Thanks to Timur for chairing this Low Latency session.

Michael Wong

Guilherme

Ronen Friedman

Guy Davidson

Timur Doumler

John McFarlane

Tristan Sizemore

Ka Ming Chan

Detlef Vollmann# CH

Piotr Grygorczuk

Vishal Oza

Sam Obeng

René Ferdinand Rivera Morell {US - C++ Alliance} (René Ferdinand Rivera
Morell)

Brett Searles

Henry Miller

Jake Fevold


14:01:07 Okay, everybody, it's. i'm actually waiting for let me see, i'm
actually waiting for teamer to call in because he that he would share this.
14:01:24 Oh, yeah, there he is. excellent. Timu is in
14:01:32 Hello, everyone! Can you hear me? Yes, we can. Yes, Hi!
14:01:38 Guy. Hi, Teama, Thank you for calling in i've made you Co-host.
14:01:42 Okay, great, you should be able to control everything so yeah yeah
just I'm, not.
14:01:49 I haven't done this before on this particular study group so i'm
not sure. I should probably keep an eye on the conversation.
14:01:55 Make sure stuff doesn't overrun make sure people don't interrupt
each other.
14:01:59 That kind of stuff Yeah, that's about it and i've sent an agenda,
but it's mostly loosely based on the idea of up of what we've done before.
14:02:09 So. I think this is mostly a discussion about he the low latency
aspects. Given that this is this month's topic.
14:02:20 And We were also going to do the games called would put trees, but
he can't call.
14:02:25 He cannot call in today, so yeah so we'll be mostly low lency, and
you can.
14:02:31 You can close it whenever you think you're done yeah yeah I I am
recording the call, and i'm also recording the transcript.
14:02:40 You can if if if somebody on you know can still take some loose
notes that would be fine as well, not, then i'll just take it from the
transcript, and that way Yeah, Do you think we should Still, have a
14:02:53 proper scribe. for the session it'd be good if somebody could do a
little bit of scribing.
14:03:00 But like I said i've kind of automated enough that I can kind of
take on the transcript.
14:03:04 It's the transcript won't be great because the machine translation
will look really strange sometimes, because as you can see right, The The
machine translation literally takes everything.
14:03:15 Okay. So then, yeah. I was under the impression that you would be
discussing Patricia's paper.
14:03:21 So now that this is not happening yeah he's not here.
14:03:26 So I I send an agenda out last night you can just take that agenda.
14:03:28 I'll take yeah, you can share screen on that. Did you want
14:03:35 I actually have to disappear briefly to go to
14:03:38 My fourth covid like my fourth covid shot I forgot that I've
scheduled it doing this call, but it it just means that I can still call
in. I will still be listening in and just in case something goes
14:03:49 wrong. But you should be okay. Let me see i'm just gonna make sure
I might have put given the the sharing to the room.
14:03:57 Okay, Okay, So you should be able to see the agenda.
14:04:04 Now, yeah, I want to. a team of the co-host.
14:04:05 Sorry, Samuel, bang I I gave it. I took. Yeah, I was about to
remind you.
14:04:12 No, no, nothing, nothing not nothing personal. i'm sure you'll
find as a goals, too.
14:04:17 But I just wanna make sure that a few people have co-hosts just in
case something goes wrong.
14:04:21 I mean
14:04:30 Can you all see my screen, or rather the agenda?
14:04:35 Yes, Okay. great. So Then I don't know Michael, before you go away.
14:04:41 One last question: Is there any formal ceremony around like the
code of conduct?
14:04:47 And all of that, because look at it at their leisure.
14:04:54 The only thing is we probably do have. We should record down the
names of people who actually are here, and then
14:05:02 The Yeah. but but but i'm not to worry about that, because zoom is
gonna record all of that, anyway.
14:05:10 And The only thing I would say is that there's a little bit of a
logistics I guess when you get you know
14:05:18 So what we we can go to Section 2.1 right away.
14:05:22 Then I can just drop off and then i'll i'll come back on after I
get to my destination.
14:05:27 Okay, 2.1. The logistics. So Cpp: Con did happen. A number, you
guys were there so you can give your impression of it.
14:05:36 The minutes minute, and but because it's a face to face it
actually goes on the wiki whereas these minutes, because the virtual I
actually send them out to the reflector I know it's a bit of a
14:05:47 weird disconnect but that's that's been the convention in the
past, where face to face goes on the wiki face to face. minutes goes on.
14:05:57 The wiki, and then the the virtual minutes gets sent out
14:06:01 So literally, So basically the whole world can see it in that way.
14:06:04 But that's really about it I think We're still on track for our
face to face meeting in Cona.
14:06:13 I don't anticipate having an idea* meeting there unless someone
wants to.
14:06:18 I will probably be leaving on Friday to go to my next meeting.
14:06:22 But you know there's there's you know you guys can, can can talk
about that logistics a little more.
14:06:30 Yeah, It would be interesting to hear who here is anticipating
being an kon in person.
14:06:37 So I I will be there I was see Michael you said You're gonna be
there Okay, I don't see any other hands.
14:06:47 But yeah, I think that's interesting. Oh, jake are there any
people who will be participating remotely
14:07:03 I I might drop into C 6. if if there's an meeting. I have There is
a meeting because we're going to be discussing linear algebra.
14:07:12 There's new a new version of the paper is Well, i'm going to make
sure it's ready by by by mailing deadline.
14:07:20 You mean staying up all night, but i'm very nearly ready on any
new version, so I hope Sc.
14:07:24 6 will come, being otherwise, i'll be quite crumpy
14:07:30 Alright, should we should we find the scribe? Is anybody willing
to take notes?
14:07:34 Or do you all think that the Zoom automated thing is enough
14:07:43 I actually think the zoom is okay? i'm not sure how many people
actually read minutes.
14:07:48 You got to Frank normally. you know it's it's discussing points on
papers.
14:07:53 Yeah, I do read minutes from something like you know core telecoms
or the the plenary, or you know Bsi meetings as well, but probably not
these, because like there's a lot of kind of casual conversation going on I
14:08:05 guess we're just going to be captured in zoom anyway.
14:08:11 Okay, so you can also see the agenda role call of participants.
14:08:16 I don't know. If do we do this here, yeah you can go ahead and do
that if you want alright.
14:08:24 So like everybody introduces themselves briefly as that it Okay,
all right.
14:08:28 I'm gonna go with the order that I have here on my screen.
14:08:37 First one is guy
14:08:38 Yeah, hello, everyone, My name's God davidson's head of
engineering practice.
14:08:41 You created assembly. i've worked for 23 years on the Total world
franchise.
14:08:43 Everyone should go and buy a copy of every game right now.
14:08:48 I work on the I I 7 c* committee I'm.
14:08:52 At the moment trying to stuff linear algebra into the standard as
soon as possible, adding this to C.
14:08:57 * cut off alright, Michael right my name's Michael wong i'm the
14:09:05 I guess the chair of a number study groups as well as part of the
Directions Group.
14:09:09 I've been in C. for about 25 years now.
14:09:13 Almost a quarter century. So thank you for joining
14:09:16 Thanks, Michael Giller. Hi! can you hear me yeah I'm.
14:09:24 A research engineer at Ait Austin Institute Technology.
14:09:27 And this is my first time here so yeah so sorry for the notes zoom
doesn't show us your last name.
14:09:34 What's your what's your full name for the notes the Glenn car.
14:09:38 Those are the Soza hutrigues it's a very long name.
14:09:40 So how could you? Maybe Ronan
14:09:50 Yes, yes. do you hear me? Yes, hey? i'm writing software for a red
hat short, and in a short while for will be writing for Ibm currently in
the last few years, involved in a large project the storage area before
that many using
14:10:16 embedded the umishal yeah so I would consider myself a C. Geek
kind of have been going into the standards, and you know, listening to
these committee meetings.
14:10:37 Haven't gone there for come here for a while but i'm interested in
low latency finance embedded systems and games.
14:10:45 Thank you. Sam. Hi! My name is Samuel Bay, and I just finished my
Mc.
14:10:55 Asian intelligence and invasion before then. I did about several
years of software before going to school. Currently, i'm done with school,
and I am yet to get my new job.
14:11:05 But I love C, and I have worked with it, and I love B here so.
14:11:11 Thank you. thank you Renee yeah let's see i'm one of the
sub-chairs or games in this group.
14:11:23 And also frequent the bunch of the other groups i'm into tooling.
14:11:28 I work for a game development company called disbelief I'm.
14:11:35 Also board member of the C alliance i'm a contributor to boost.
14:11:42 I have my own various libraries in Github. yeah.
14:11:48 I think that's all. Thank you Jack Yeah. my name is I work for
Boomer to lp.
14:12:01 Thanks. John John Mcfonin. hi there I know a joke Phone: i'm a
member of I think on the Thank you.
14:12:21 Tristan, hey? I just joined this as she put sorry we can't hear
you very well.
14:12:37 Could you possibly turn up your volume a little bit?
14:12:40 Can you hear me now? yes, so hi guys i'm kristin I just joined
this Sg* group.
14:12:55 Via email request I asked Michael Wong I could join.
14:13:02 I've just joined this observer I c.
14:13:15 Professional work at a company called Butt do crypto trading.
14:13:28 It's a mobile app i'm just observed
14:13:37 Thank you. We have 2 other people who joined the call.
14:13:41 I ha probably not going to pronounce this properly. apologies.
14:13:46 Coming, Chan: yeah, hi there, you put, Nancy is very correctly,
actually.
14:13:54 Okay, do you? wanna tell us a little bit about yourself.
14:14:00 Oh, yes, sure. I i'm actually sitting in London I worked in the
finance area.
14:14:08 Hi Co. he would get for a hedge fund.
14:14:11 Call Pauli asked me, which is a actually a us hedgehog.
14:14:17 So I use c in quant quantitative library.
14:14:25 I in general interested in the low latency and actually everything
else as well.
14:14:33 So hopefully I would be able to contribute something. But then
14:14:37 At the moment I could I just want to listen in and the I'm.
14:14:42 Sure it is about what you guys are developing yeah that's just
myself.
14:14:48 Thank you, thank you, and and finally Deadlif. Oh, my name is
Tetra Foyman, i'm from Switzerland, however, mainly in embedded. But right
now I also have a go check that has Low latency
14:15:06 in the sense that I have to do I have to boot. something from an
embedded system to a server. and that has to be done at the 6 deadline
that's for me.
14:15:22 Thank you, and we have one more person who just joined.
14:15:26 Pyoto. Yeah. waiting for them to connect to audio
14:15:37 Pyotr! Can you hear us? Can you hear me
14:15:46 Hello, Filter. Hello, So sorry just to connect it now.
14:15:55 And my microphone didn't work so
14:16:05 So my name is Peter and I was part of those calls before just to
join today.
14:16:14 And I work for intel in the embedded systems.
14:16:19 So yeah, thank you. My name is Steve Modumna.
14:16:25 I am developer advocate at japanese so we make tools.
14:16:28 So i'm into tooling I I also spent a lot of time working in audio
particular music production software, which is another low agency,
application, because it's kind of real-time number crunching.
14:16:41 So I'm very interested in low latency and real-time applications
of C.
14:16:46 And i've also been active on the c committee since I think 2,016.
14:16:51 So yeah, I there's seems to be a couple of people here on the
call, who have not been here before so welcome.
14:16:57 And thank you very much for joining i'm just gonna go through the
agenda here.
14:17:02 Which I got from Michael. action items from previous meetings I
don't know what that is.
14:17:10 If you Michael is still on the call we don't have any you don't
have any.
14:17:16 You don't have any okay great so you already discussed general
logistics.
14:17:19 We had a meeting at Tp. call and the minutes are here on the wiki
14:17:24 I I'm literally just reading out the minutes here sorry the agenda
here.
14:17:28 So we do have future meetings planned.
14:17:32 You will not have a call in november because we're gonna be
meeting face-to-face and Cono.
14:17:38 At least some of us will. however, there are calls planned for
December, January and march at the dates in the agenda, and they're going
to be continuing rotating games embedded and
14:17:51 finance national latency. Michael, do you have anything to add on
that?
14:17:56 Yeah, we could probably at this point stop saying that we have
we're gonna be doing more stricter.
14:18:01 Membership attendance and because of that I think everybody has to
choose meeting number, which is, unfortunately, but there are ways that we
can get around it, especially alliance. So maybe you can talk a little bit
about that rules change do you know what
14:18:20 i'm talking about teaching Yes. So so from my understanding
Basically, if you want to participate in a committee meeting, including a
telecom, including study group telecoms like this one.
14:18:35 You basically have to be on the Isa Global Directory, or you have
to be appointed as a as delegate or as an alternate by by your national
body.
14:18:45 So if you are not on that iso delegates list, then you can attend
one meeting as a guest, and after that you should get into touch with your
national body.
14:18:57 And become a member of Iso by a delegate of Iso
14:19:03 So. I guess that means this already applies to this telecon.
14:19:09 Am I right? No, not exactly we're gonna have some we're gonna have
a bit of a grace student in particular is kind of different.
14:19:18 It's always been an open outreach group at which people did not
have to be official members. but now that is changing. and so I have to
find a way to as well for this i'm i'm i'm
14:19:32 i'm basically just putting off but I think that there's a a way
it's not easy for obviously for people to join to become members.
14:19:43 So there's a way at which you can join and become A member would
be plus alliance. Now send some. Maybe he can tell us a little bit about
that I don't know if they can do that for us.
14:19:58 It's actually the c foundation not the supposed to no.
14:20:08 So. so actually, a guy has his end up. Yeah. Sorry.
14:20:16 Excuse me. Yeah. there are 2 organizations which are offering
membership to people such that they become members of the Iso.
14:20:25 They they can record in the iso directory there's the cpus
foundation, also boost foundation.
14:20:30 You can write to other organizations. request to be to join, and
you will be covered.
14:20:39 As members of those organizations through insights.
14:20:43 Renee the I don't know if the Boost Foundation has actually
finalized their procedures for that.
14:20:52 I know they have. they have a request for it and the board met,
and they're okay with it.
14:20:58 But they I don't think they've actually gone through and actually
become a member.
14:21:02 Okay, I will. I will say with David sample leave.
14:21:06 It was all good. But yeah let's let's wait until this action in
that in the public domain there has been an announcement in the public
domain about the sequence Foundation. there.
14:21:16 So you can contact Actually, you just write to have such a sutter
at gmail dot com shots or also nina nina.
14:21:29 Rans, Who is the the secretary, I think of the or one of another
member of the Yeah.
14:21:38 Either, either. Herborneina, contacting them will do.
14:21:42 Yes, it's it's the 9 situation.
14:21:50 But this is the thing that happens with international bureaucracy.
14:22:01 Let me just
14:22:08 Yeah. So if any of you want to keep participating.
14:22:12 And these calls and want to become a member. Please contact one of
those organizations.
14:22:21 All right. I think we can now then move on to to paper reviews,
unless there's any other kind of bookkeeping or general logistics business.
14:22:37 Can someone sum up if in a few minutes the face to face meeting
for those of us who cannot read the this link?
14:22:48 Sure. Maybe I can like do a little of pubits.
14:22:52 And then other people who were there can like fill in the blanks.
14:22:56 So we had I think 3 papers one was
14:23:02 I had a paper on kind of reading the bits of an object. it's like
reading object representations paper P.
14:23:10 1839, I think, which we discussed because it's relevant to the
latency.
14:23:17 That was the very long discussion. Then we had quite a long
discussion about at least. voice paper, which is basically not not the
published paper yet, but more like a big word document with lots of ideas
future proposals that are relevant
14:23:37 for low latency and these kinds of applications. So there was a
lot of discussion around, like all the kinds of features that were in that
document that some people thought might be a good *.
14:23:51 And We had a third paper which I can't recall now, maybe somebody
else remembers what that was
14:24:00 Or I can just cheat and look into the minutes and then probably
i'll remember
14:24:11 Oh, yes, there was a paper about graph data structures which I
wasn't in the room for so I can't really comment on that.
14:24:21 Oh, I think I was in a room but I wasn't really following this
question on that one very well
14:24:31 Yeah. Does anybody else have any more info about that?
14:24:38 Or does that answer like roughly what was going on there?
14:24:41 I think the the other thing that was going on there that was
really interesting is that that was the first of meeting of a committee
study group where we try to do hybrid attendance where we had a bunch of
people in the
14:24:53 room, and then also a bunch of people joining online at the same
time, And which is something we want to do in Coney.
14:25:00 And this was kind of the grand rehearsal for that, and we had
quite a few problems with everybody hearing, like everybody on line,
hearing what the people in the room were saying, and things like that.
14:25:11 So I didn't go very smoothly but I think it mostly worked. and it
gave us kind of good kind of good input on on what to improve for kona,
because for corner there's gonna be a lot of sessions running in this
14:25:24 kind of hybrid mode where you can attend remotely, or you can be
in the room.
14:25:29 So that was a bit of an experiment which I think was successful.
14:25:32 Maybe not in a sense that it went very smoothly, but in a sense
that we we got valuable data on on how to pull off such a hybrid meeting.
14:25:39 So hopefully in Kona. it's going to be smoother than it was there.
14:25:51 Alright. So then, on the agenda, we have a bunch of papers.
14:26:01 We have P. 2, 5, 3, 2 which is removing exception.
14:26:07 Pointer from the receiver's concept does anybody wanted to present
something about that, or
14:26:22 I don't I don't think there's authors for any of these papers
present.
14:26:26 I think today there's just going to be a discussion about low Lane
Sea.
14:26:29 That's all that's been being discussed on the reflectance in the
last couple of weeks.
14:26:35 I think you had some you had. you have started the discussion
thread and maybe I think what we need to do is just summarize that for
everybody on what's been What's been discussed so far in case Well, for
people who
14:26:49 don't follow the the reflected easily I don't know if that helps.
14:26:53 Okay, yeah, i'm not sure what discussion this this relates to in
particular.
14:26:57 So I'm not sure if anybody else here knows what this is about and
could maybe talk about that
14:27:10 Yes, but 2 messages on the on the reflector in the last 2 weeks, I
believe
14:27:19 Okay, could you potentially. that's something that we can screen
share and look at or
14:27:34 Which kind of not sure how to to conduct this discussion because
i'm just not aware of this particular i'm I don't know what what emails
you're referring to
14:27:49 Actually, I I believe. Michael is referring to your original email
with the request for financial, about their low latency, or how real time
requirements.
14:28:08 Oh, that Okay, yeah, yeah, yeah, Yeah. Okay, that's obviously i'm
aware of because I I wrote this email wasn't related to any kind of
proposal.
14:28:18 So I I sorry I just didn't realize that this is what this is what
you were talking about.
14:28:23 But sure I can talk about that. Yeah, I I can definitely talk
about that.
14:28:28 And then the other email what's the other email i'm just looking
at the reflector right now.
14:28:42 Just trying to figure out like what's the agenda basically for
today.
14:28:45 So we can talk about the the email thread that I started
14:28:50 And then, basically, is there anything else that that you should?
We should talk about
14:29:00 I have a feeling it might have been stefan's response to email as
well.
14:29:06 The whole thread was quite interesting. Okay, so i'm just so, then.
14:29:12 So so that's then just talk about that thread and if there's no
other topics.
14:29:16 Then we can join after that. So yeah, I can. let me.
14:29:20 Then I can actually share screen share and and bring that email up
here.
14:29:25 If if you like. just give me a second please sorry i'm a bit.
14:29:33 I'm a bit slow today. it's been it's been a long day, and it's
late in the evening.
14:29:38 Here. Here we go! Hey! Let me share this.
14:29:57 Can you? Can you see my screen with an email thread on it?
14:30:02 Yes, okay, great. So just a little bit of context, what this was
about.
14:30:07 So I was actually due to give a talk. Actually, that guy Davidson
has invited me to at his company about low latency, c.
14:30:16 And as I was preparing this talk, i'll sorry debt that you have
your hand up, No, okay.
14:30:25 So, as I was preparing this talk it, was kind of about kind of
typical properties of low latency low latency applications, and how I
wanted to figure out how in which ways they're similar and which way is that
14:30:38 different, so obviously in low latency whether it's games or
finance or audio processing or awesome.
14:30:48 Also quite a lot of embedded use cases. Not only do we care about
whether a piece of code does.
14:30:54 The thing it's supposed to do but also how fast it does it right?
14:30:57 So like. There is kind of an implicit kind of deadline, or we want
to get the answer faster.
14:31:01 We care about latency. So so I was kind of looking at
14:31:06 Where different use cases differ. So, for example, in finance,
versus an audio was in gaming.
14:31:15 You have kind of different different time scales there, right? so
in.
14:31:18 So I was kind of familiar with audio very much, and not so much
familiar with
14:31:23 Maybe a little bit familiar with games, but not so much familiar
with high frequency trading in particular.
14:31:28 Yeah, and I I just so sorry. I think like those 2 portions of the
high frequency training.
14:31:36 I think there's one that's low latency and then there's one that's
high throughput where what what you're doing is you're kind of the hype
group but you don't necessarily care
14:31:45 about time. You kind of want to push all of the data.
14:31:48 The high frequency training would be more towards towards like a
fpgas or solar flare cards where you what you're dealing with is you want
to when you're getting information from the exchange you want to
14:32:01 react to that as fast as possible, whether that's order
information or whether that's market data, you want to be able to to at
least put some some state on your State machines.
14:32:16 So that, and maybe have some code that might execute
14:32:22 That might be more towards the high frequency. trading as far as I
kind of had experience with.
14:32:29 Yes, yeah, I think that's accurate like I think if you're in
finance. i'm not a finance person, but I guess you have lots of
applications where you just kind of crunch.
14:32:37 Lots of numbers you you I don't know run some risk models or
whatever you do there. And then it's just a lot of kind of processing we
care about throughput, because there's a lot of data coming in Yeah, that
that that
14:32:48 can be one thing there's also the other side side where you you
basically want to react.
14:32:52 So, for example, let's say that you find like a you know a company
price changes, or like you know, a future or commodities price change.
14:33:03 You want to be able to react as soon as you get that portion
that's kind of a different portion.
14:33:08 That's more the low latency, stuff where high frequency is
basically you you're getting all of the number crunching in your you're
forwarding that into you know a a you know container that can you know Do
all
14:33:21 the processing. Yeah. So So I was specifically looking at use
cases where we care about low latency and not high throughput.
14:33:29 So it seems like that kind of orthogonal aspects of performance in
a way right?
14:33:33 No no that that's why you kind of shouldn't say high frequency
training.
14:33:38 It's low latency. low latency processing rather than a high
frequency training.
14:33:44 Usually that's and that's, usually related to more or less
algorithmic trading due to the fact that you want your algorithms to
respond as fast as possible to, or at least in a controlled way depending
on if you're dealing
14:33:56 with one exchange, or if you're dealing on with many exchanges.
14:34:00 You want to do that? you know, you know and they small if you're
dealing with like one very fast exchange, where you have all your your
servers, or your you know your order routers, etc.
14:34:17 You know, very close like, you know, no more than a couple of feet
away from the Exchange.
14:34:21 You know that's going to also help you because you don't have to
worry about the lag Now, if you're kind of placing those orders on a bunch
of different exchanges.
14:34:32 You might have to kind of you know. you know, work with the lag to
to make sure that you, you kind of have your orders delivered at multiple
exchanges.
14:34:42 So like you worst case would be that you didn't have the you know
the the slower exchange react about at the same time as the higher exchange
that's like 1 one stuff that that's like one way you can do
14:34:56 that. if you're just doing something like arbitraging right yeah,
that's interesting.
14:35:05 So a bunch of people reply to to this this thread here.
14:35:10 So Stephan wrote reply: Wesley, author And and from what I kind of
learned from that, is that kind of low latency.
14:35:23 Financial applications seem to have at least 2 interesting
properties.
14:35:26 That kind of distinguish them from all other low latency
applications.
14:35:29 I am aware of, like audio processing or gaming.
14:35:34 One is the time scales, whereas, you know, games or audio operates
usually kind of in the order of maybe one or 10 kind of milliseconds.
14:35:45 Finance seems to be operating on much, much shorter time scale.
14:35:48 So are you like, and and kind of certain high, free, concentrating
scenarios.
14:35:52 You want to. if you get information from the exchange, you want to
then send something there within an order of a microsecond or even even
faster.
14:36:02 So I thought that was quite remarkable. And the other thing that I
found interesting is kind of the deadline you have, where, for example, in
games and in audio processing, you know what you deadline is right?
14:36:13 So you have a certain frame rate for example if you're producing
video, or you have some kind of audio buffer size.
14:36:18 If you're producing audio and you kind of know that you have to
produce data within for example, 1 ms or 10 ms in order to not calculate.
So in order not to drop a frame so you can
14:36:32 know what latency is expected of you in order to you know get the
desired result, Whereas, these high frequency trailing up applications have
this curious probability that you don't actually know exactly what your
deadline is you obviously need
14:36:44 to be faster than all your competitors. But you don't know exactly
how fast that is.
14:36:50 So you in the end, like kind of just try to be as fast as
possible, which is also kind of quite unique.
14:36:57 In this kind of wider field from from kind of my perspective which
I found very interesting, Right?
14:37:18 Basically it's you know you do although I think like sometimes
there's also the idea of being kind of slow If you're dealing with multiple
exchanges simultaneously and placing those orders basically the idea is to
get you
14:37:22 know. Get your orders you don't know exactly what time and I think
also, another thing is networking high frequency training usually has some
sort of networking issue.
14:37:28 That's kind of why? you don't know the time scale. Because you
basically have to worry about how the exchange goes and processes those
orders.
14:37:37 If those the exchange is order management system is very slow,
then you know you can't really do anything about that other than wait.
14:37:47 And you don't know about that yeah they can they can always
change, you know, change their their order book or their order processing
structure over over time either to slow it down to deal with like you know,
certain hardware issues
14:38:02 or, you know, make it reliable, or maybe for some sort of
regulation purposes, or something like that.
14:38:14 Right? Yeah, that's interesting
14:38:24 Yeah, So basically, basically, the whole point I think you're
talking about is not necessarily high frequency trading.
14:38:31 But low latency or algorithmic trading, which really is about like
reacting to the orders.
14:38:39 And I think, you know, I don't think anyone does like audio
processing or embedded systems.
14:38:46 Truly too much on the network. Usually they're on the local device
and kind of in an enclosed area where I think high frequency training or
algorithmic trading is more on the more on the network.
14:39:00 It. It may be something that may be processed on the cloud, or It
may be something that's processed, you know, in in some sort of embedded
system, or sorry not a or like a network.
14:39:16 A you know just a plain old Internet connection or you can you
have stuff that's related to web software.
14:39:23 It's that might be not necessarily the same as like a true socket
connection.
14:39:30 Then you also have it connections that also also work on that type
of system.
14:39:39 So. Yeah, that's interesting, I think it's not quite right that
other low latency of the applications don't work with networking.
14:39:46 Obviously you have audio and gaming where something like this app
zoom. right is is, you know, a real time application that uses networking,
or you know, in games here you have networking as Well, it's just that I
think
14:39:58 the time scales are different, because you were talking about tens
of milliseconds, and i'm actually curious to hear about.
14:40:03 You know. Maybe there's some gaming people like die, for example,
who could talk a little bit more about the kind of networking aspect?
14:40:10 I think that the 2 differences I can see is like the time scales,
and the other one is, whether you, whether or not you control your
networking stack.
14:40:18 I think Please correct me. i'm wrong but my understanding is that
in and trading you have your own like customized kind of network cards and
customized drivers, where you kind of bypass the kernel and like some
things in between you can do
14:40:33 even like with Fpgas, and and so you have like completely
customized stack there.
14:40:37 Whereas if you're doing audio games you do use the network, and
you do want to be fast. You don't want to have a unacceptably slow latency,
you know, when you do you know I don't know playing some
14:40:48 first person shooter over the network or something but you don't
it's It's just regular consumer hardware, which you don't have any control
over unless you want to console maybe but still then you don't
14:40:55 know what networks. set up you have so well well I think There's 2
portions you have your own network where you're trying to minimize.
14:41:04 You're doing, you're having cards like solar flare, or you know
it's custom fpgas.
14:41:10 Some people might do like you know if you're trying to get in.
14:41:12 You might do stuff with, you know. regular conventional hardware.
you know.
14:41:16 Maybe use a Gq. or something like that I mean this may be if
you're a smaller shop that you know, are trying to start in into the low
latency algorithmic trading or high frequency, trading.
14:41:28 They might do that. but I I just i'm saying like Yeah, you know
you have 2 portions of the network.
14:41:35 You have the exchange, and then then you have actually, you have
may have 3 you may have.
14:41:42 You might have a broker like a, you know, interactive broker, or
something else.
14:41:45 That kind of you know, takes your orders, and sensitive to the
exchange.
14:41:49 Then you have the exchange like the cme group ice your X etc. New
York Stock Exchange, where he basically you're sending the orders directly
to the Exchange.
14:42:00 And then they are trying to, You know, process those orders as
fast, you know, doing an order matching so structure.
14:42:09 So you basically, you have control over your side up to the
Exchange.
14:42:12 But then, you know, the exchange has to deal with when you're
sending the orders to that that might be, you know, another stuff where you
know the processing.
14:42:24 The issue is processing. You want to keep your side as fast as
possible, but you might want you might not actually have any sort of
control over.
14:42:34 Once you send the orders you could send, It could mean, maybe
take, like, you know, a day, or it may take a you know, a couple seconds
for those orders to be processed, or the order that you had sent to be
processed
14:42:50 Where I think, like with a game, you basically, you know, once
once everything you know, the system updates.
14:42:58 You know your networking system kind of is there?
14:43:02 Everyone tries to be as fast as possible, and has the same code.
14:43:06 You know. you know they they're trying to play the same game on
the same, maybe slightly different hardware.
14:43:13 But, like, you know, the same game structure Everyone is a
different, you know you have the buyers and sellers on on the Exchange, and
that's probably one of the difference between the you know, embedded
systems, and high frequency
14:43:29 trading or actually low latency trading. So sorry about that.
14:43:32 But yeah, so hi! I appreciate high frequency. can kind of be
something where you're taking you're taking the market data.
14:43:40 You're processing that to maybe I think the this was like about a
time.
14:43:46 I I think Stefan and I were working at the same company where we
had.
14:43:52 We were using like a getting some market data, and just kind of
calculating the implied.
14:43:59 This was this: wasn't something that I was doing per se. But this
was something that a colleague was was doing where he was taking the the
data and processing that on a nvidia gpu card to
14:44:14 calculate implied. implies and this was like a thing that was
happening like once every month or so.
14:44:24 4 eyes introducing Continental exchange.
14:44:27 He was just doing that to get, like all of the all of you know,
calculating the implied pricing from the from the actually price press data
14:44:48 Right, so that that was more towards the high frequency training.
14:44:53 But you know you will latency, is us and you basically the idea is
here you want in the high frequency trading year, trying to spread out a
lot of borders, and you don't really care about you know if they're filled
14:45:04 or not. You just want to do you you do a lot of badge processing
where low agency is more like a you know, you're sending out that are
little bit more smart and you're trying to do it more towards
14:45:19 algorithmic structures
14:45:23 Yeah, it's interesting because I heard like from some as somebody
who comes from the outside like I heard the terms high frequency trading
low Latin situating and algorithmic trading, or basically being used
interchangeably so
14:45:37 where there's any difference between any of them right basically
the idea is hybrid frequency is like you're sending a bunch of orders in
went batch where low latency you're trying to I think low
14:45:49 latency is usually more closer to the out algorithmics.
14:45:53 Trading algorithmic. you're you're having more of A. You know what
what you're supposed to do where low latency is.
14:46:00 Basically, you want to react as much as possible. So low latency
and algorithmic training might be kind of very much closely tied together.
14:46:11 Where high frequency is more like you know the idea of sending a
bunch of batches. not necessarily knowing what you're doing, but you're
trying to take advantage of of something you might have some sort of
processing, but they may maybe
14:46:22 not be more of a smaller processing. Yeah, and you know a lot of
stuff might be a Os kernel.
14:46:29 One thing that i'm kind of trying to do on my my side is a
basically experiment with like a training system where where you're using
14:46:40 You're using a and micro kernel operating system and keeping all
of your training structure.
14:46:47 In the user space very much connected, you know, once you once you
initialize your your system.
14:46:57 Basically, i'm trying to look at using linux, 3, which is kind of
like a micro kernel and just trying to prevent like, you know, the the
kernel system from you know disrupting disrupting
14:47:11 the the software happening, although Chris is more like a
experimental idea without an exchange, it's kind of worthless.
14:47:22 So pick that as you will
14:47:32 That if you're doing you know probably not in the high frequency,
you might not be.
14:47:42 Yeah, you might you basically, with algorithmic training you're
trying to, you know, set like, you know, what what the conditions are for.
for you know, making the order, You're not necessarily sending a bunch of
orders you might send like a
14:47:56 pure pure orders. But you want to react when you have market data
as best as possible.
14:48:01 At least That's from my point of view there might be other people
who, who, you know, are also in the same domain, but not necessarily not
necessarily there.
14:48:13 I think, one of the structures that i'm kind of trying to, you
know, get myself a little bit of associated with this called Mediterranean,
which uses c.
14:48:24 In order to create the algorithmic structures. But this is more
like for smaller, you know, smaller smaller time of traders.
14:48:34 I I did think like there May be other people who are like you know
if you're a hedge fund, or you're if you're something like Bloomberg or
something you might not actually kind of use the
14:48:43 Mediterranean platform. you might actually use like you know raw
data that you're getting from the Exchange, like the Cme group, ice, New
York Stock Exchange, etc.
14:48:57 Nasdaq, etc. so you might be trying to do something towards
towards that
14:49:11 I could have raised my hand. Okay.
14:49:18 Any more on this tedler related to this i'd be interested in.
14:49:31 Well, if you have a deadline of microseconds do we actually use
parallelism to meet that deadline, or try?
14:49:47 Or do you try to do everything as fast as possible on a single
core?
14:49:52 I would say it might depend it might be if you're dealing with
like a high frequency training where your number crunching.
14:50:00 Then parallelism can help if you're dealing with like you know,
just sending out single orders.
14:50:05 I would say that maybe it might be better to to do it everything
as fast as possible.
14:50:12 Rather than you know, doing doing stuff normally.
14:50:19 It's basically kind of it depends you have 2 different kind of
strategies.
14:50:23 You have the high frequency trading strategy where you're
basically sending a bunch of orders.
14:50:29 You know, a little bit mindlessly. and like hoping some of those
get bills, or you know, hoping that you know you get paid, or something
like that.
14:50:39 Then you have the low B Latency version work where you are trying
to, you know.
14:50:45 Make sure that I mean it may be just running once on a single
core, or you might have maybe a core that's dedicated to 2 different
environments like you might have one that's right doing something like
cheese futures or one that's dealing with oil
14:51:03 or something something else so that that's what you know that's
what the idea of the hypothesis trading?
14:51:13 Yes, it's basically you have or that's what's related to the
trading.
14:51:19 The the side you So yeah Thank you.
14:51:29 Interesting. Okay, do you mind if i'm remove myself sure.
14:51:35 So one thing I can add like from the audio processing perspective.
14:51:41 Is that also like if you have a like no latency, application
there, where you're doing like real time processing, you don't want to be
doing any multi-shotting, there.
14:51:49 So typically you have one thread which is kind of the real time
thread where you do your processing, where you have to generate like a new
audio buffer.
14:51:56 Every millisecond and you don't wanna do that on multiple threads,
because, you know, you have to synchronize those threads. and then you have
to interact with a threat scheduler in order to do that
14:52:07 and and that's not going to have kind of a deterministic execution
time.
14:52:11 And so you you can't rely anymore on kind of being below your
deadline.
14:52:17 So it tends to be that you have, like in particular in this
particular domain, that you have a single higher priority thread which is
doing the real-time processing, and is not doing any parallel stuff at all
and all the other
14:52:31 threads that deal with like the Gui, or the networking, or dis
disk access, or whatever it is.
14:52:37 They do that kind of independently and if they need to exchange
data. you'll use like a single producer single consumer lock free fifo, in
order to do that. because that's wait free not just lock free So
14:52:49 you can again reason about execution. time there and this way you
can kind of like, get in that data in and out of the real time fit.
14:52:56 So that's kind of how it works in this particular domain of audio
processing.
14:52:59 But bunch of hands. Guy was first I believe yeah i'm I you said,
that's We don't want to using threats because of the normalistic nature of
switching.
14:53:13 And so on. I might be talking as a rubber shirt.
14:53:17 But do you think that asynchronous co routines my offer?
14:53:21 Determinism or sufficient determinism. if I may just directly
reply to that. My understanding of core teams is that, And they don't know
anything about threads, right?
14:53:35 They don't. they're not like concurrent or parallel in any way,
just by themselves.
14:53:39 They're just a way of passing control to a different context.
14:53:43 Like. If you want to do parallel stuff with core teams, you have
to add that on top, right.
14:53:48 By the way, you write your promise type or your whatever it is
you're available, and all of that that's where the concurrent stuff or the
parallel stuff goes and and that's going to use kind of the same
synchronization
14:54:01 mechanisms, as is also in the language that's my understanding.
14:54:07 So you can obviously use core teams in these contexts and I think
quotings are great because they're very kind of low overhead.
14:54:12 They're very efficient. but at that point you're not doing
anything in parallel.
14:54:17 If you want to do stuff in parallel you're gonna then have to add
some to that synchronization mechanisms on top.
14:54:24 And then again, you're gonna run into the same problems as with
all the other language mechanisms.
14:54:28 But i'm curious what other people said I think because in terms of
hands we had then that left, and get out
14:54:40 Cooling is not a concurrency it doesn't know about the other fresh?
14:54:43 It just seems back and forth. Leave it with trust because there's
no time
14:55:00 Up to low latency folks. What do you guys do? If you do use the
Stl do you solve the determinism problem?
14:55:10 Do you just not use stology just use a version of sql that doesn't
have any
14:55:19 Does anyone want to direct your reply to that yeah I can We don't
throw exceptions and we minimize allocation.
14:55:29 Those the 2 greatest contributors to non deterministic behavior, c.
14:55:35 Plus plus We minute survive, not throwing exceptions that does
rule that certain parts of the Stl.
14:55:44 But by by locking out allocations that if it actually eliminates
all the containers, so we do have to accept some aspects of owners in which
we mitigate by right well it's a creative assembly to
14:55:58 mitigate that by writing our own allocators. for example, using
it, Reiner allocators using cool allocations.
14:56:05 We are allocating. all the same, you know objects the same size.
14:56:08 You can. being much more certain about you, you can place a much
stronger, shorter upper band on the amount of time allocation will take.
14:56:17 Yeah, I mean audio people do the same thing. I have actually a
whole talk about that called using like, I think, real-time programming
with the standard library, or something like that, where like I talk about
like the subset of the Stl: that you
14:56:30 can use that actually has the deterministic runtime.
14:56:33 And yeah, that eliminates everything that allocates memory.
14:56:38 It means all dynamic containers, everything that has type or Asia.
14:56:40 You can't do any of that stuff you can't do anything that kind of
might have a lock inside.
14:56:47 You can't yeah So so it's it's a peculiar subset of the Stl.
14:56:52 And a lot of people they they, instead right kind of their own
replacements for this.
14:56:58 So you can't use the vector because it's allocating. So you might
want to write a static vector right? and actually a lot of the kind of Sd
14 proposals that we have been looking at target exactly those use case
14:57:10 and proposed facilities kind of to accomplish that, to avoid
things like allocations or locks.
14:57:17 So I think debt left was was next yeah about my Oh, which no
question was already onset. But also about the Sdl: Well, about exceptions.
14:57:32 Well, if you get an exceptions, then you have something that you
can't meet your deadline, anyway.
14:57:42 So that case is not really a problem, and having no exceptions, is
typically pretty deterministic.
14:57:53 Time wise. So that is not a big problem. And yeah, about
allocators.
14:57:59 We already heard a lot, and that's a very important thing you have
to look at.
14:58:05 If you need determinism. Okay, be shy. Yeah.
14:58:14 So I was just kind of, I think, like this was about like
exceptions, like I, and like, like, this exceptions and kind of like the
low latency portion.
14:58:27 So basically I was just saying in the financial industry, maybe
you might be able to use some exceptions or some of the exception handling,
although, it, it might be kind of the case.
14:58:39 By case but basis you know. So some people, some companies might
might allow them.
14:58:45 They might like. I think the the idea is like, you know you you
use most of the Stl.
14:58:50 Maybe a little bit of boost and like, you know, there is some
cases where you might do something you might use like the standard
containers.
14:59:01 Oh, although maybe you might have to kind of you know like, you
know, modify or kind of use, a different like, like, instead of using a
vector you might want to use a deck or a queue in order to process, this
14:59:17 you know, push in data, and it basically because we're not
necessarily looking at deterministic data.
14:59:23 We don't know when we're gonna get the information from the
exchanges, or anything else. we don't.
14:59:29 I I think that in this case it might be better to have like a
queue, and I think also possibly looking at stuff from the Exchange, or
sorry from the from Concepts and co-.
14:59:44 Routines. That probably is also like might be useful for for
people, at least on the financial low latency level compared to compared to
somebody who's doing something on a embedded system, or gaming gaming thing
where you can
15:00:01 write a co routine that actually handles handles like a message
type that you're getting from the exchange.
15:00:11 Yeah. So can I just say one more thing about the whole exceptions
thing?
15:00:18 Because somebody said, Well, if you're throwing an exception you
don't care anymore about the determinism, because at that point you're like
in failure.
15:00:23 Mode that's not quite true or it's not true for every use case.
15:00:28 So we know that on most platforms as long as you don't throw an
exception like having exceptions in your code doesn't have runtime overhead.
15:00:37 It. It does have runtime overhead I think on windows 32 bit, but
it doesn't have runtime overhead on any other desktop or mobile platform
that i'm familiar with it has
15:00:47 overhead in size, binary size, so that that's important for
embedded systems.
15:00:53 But more importantly, there are scenarios where the like even the
arrow path needs to be deterministic like.
15:00:58 For example, if you're doing audio processing you have to call
back right every millisecond you have a callback, and then you get a
pointer, you have to feel like new audio frames into like the array
15:01:10 that's this pointer pointing at and that's going to be set out to
the speakers.
15:01:15 So. you cannot. There, just give up on your determinism and say,
oh!
15:01:20 Some exception was thrown somewhere i'm just gonna not do
anything, because then you're gonna not write any data into the buffer.
15:01:25 You're gonna get an audible glitch or click which you know in the
worst case, might actually destroy your speakers, because it's like a very
shop kind of like discontinuity in your waveform.
15:01:35 So if you encounter an error, you you have to do something else.
15:01:38 You have to, you know, fade out, or you know output.
15:01:44 Maybe some noise or output silence, or or do something else.
15:01:48 But you do have to deterministically produce some data.
15:01:51 Right. So you can't just give up and say Oh, an exception has been
thrown.
15:01:56 I don't care anymore about this function you know like returning a
result within a millisecond.
15:02:02 You just can't do that, and I Imagine that There are quite a few
embedded like use cases again.
15:02:08 I'm not an embedded guy but there are quite a and a imagine that
quite a lot of embedded.
15:02:13 Use cases where you also have this kind of callback or deadline,
where you have to get a result within x milliseconds, no matter what like
I'm thinking about automotive or robotics, or you know medical kind of
devices
15:02:25 and maybe there's some people here who can comment on this stuff.
15:02:29 But in those use cases you cannot just say Oh, an exception has
been thrown.
15:02:33 I don't care anymore about like the the runtime of this function
being being non-deterministic.
15:02:37 You just can't do that, and therefore you just end up not using
exceptions at all right.
15:02:43 And I think, Patrice actually at one of the earlier earlier
meetings actually mentioned that you know there is like a per potential
performance overhead, with just exception handling kind of even if you
don't even call that the
15:02:59 exception. he was doing some sort of benchmarking, I believe.
15:03:03 This was, I think, like, you know, very, very long time ago.
15:03:07 But it it is kind of it should be kind of a note noted, I think,
that you know it's there, or it is known.
15:03:16 I I think that Sutter was actually mentioning like making it, you
know, kind of turning the exception handling P path to be more
deterministic in order to or kind of in the return type to basically not
necessarily
15:03:33 dynamically allocate when you're doing an exception when you're
generating, and string that you're kind of doing the the exception that was
probably one of the things that actually hurts with the exception
15:03:45 handling. Yes, so So there are 2 different things here. one is
what happens when you throw an exception, and what happens when you don't
throw an exception.
15:03:52 If you throw an exception currently that's a dynamic allocation
which is not a deterministic and that's just the way like the language
works.
15:04:04 You cannot really do anything about that. So her was addressing
that problem with his proposal Right?
15:04:09 He was trying to find a deterministic way. of like throwing
catching exceptions in a different way that doesn't require rtti and memory
allocations.
15:04:18 So that's one problem by the way. i'm curious you know what
happened to this proposal.
15:04:22 I think they have to know developments there in the last 3 years
since this was released.
15:04:27 But if anybody knows anything else i'm i'm curious but then the
other thing is once you finish all initial discussion. Yes, and so the
other thing is what happens.
15:04:37 If you have a try, catch block there in your code.
15:04:41 And you have exceptions enabled but you don't actually throw an
exception.
15:04:45 Does that have any overhead, and the heck and benchmarks there?
15:04:48 I think the number of people did that. I think Ben Craig also had
a paper where he was looking at this
15:04:55 It's kind of subtle so but at the end of the day it boils down to
15:04:58 You have, like 2 possible strategies how to implement exceptions,
right?
15:05:02 Because you need to store all this information about how to unwind
the stack that that needs to go somewhere.
15:05:08 Right. So either you generate that information at one time and
then you get runtime overhead, which is what the windows 30 two-bit does,
or you generate that information at compile time, and you store it
somewhere in your binary and that's
15:05:20 what windows? 60 four-bit does That's what Linux does.
15:05:25 That's what like android does so you don't notionally have runtime
overhead.
15:05:31 However, you have you have more stuff in your binary so that's
gonna affect code layout.
15:05:36 So you can still indirectly effect performance anyway.
15:05:40 So, but it's kind of very hard to measure and it depends on how
exactly you set up your benchmark, and so that kind of stuff.
15:05:47 So that's how far my knowledge on this topic goes if if anybody
knows more about this.
15:05:52 I would be very happy like very, very curious, but let's first
hear a guy his hands Do you want to go first?
15:06:03 Oh, I thought I thought you wanted to reply to this but well, I
wouldn't just definitely find something else.
15:06:10 Okay, that's just regarding habits exceptions it's it's a great
paper.
15:06:17 But you know. he's one man who is short on time and it's been
devoting all of his energy to secretly, the next the new syntax called
about developing a new syntax as far as
15:06:29 i'm understand he's expecting to wind in some of these other stuff
that you know is Meticlasses proposals, and is exception proposals into the
new syntax rather than channeling them into the
15:06:44 into the into c What are the problems with the exceptions since
there's an awful lot of time being invested.
15:06:54 Buy some companies into making care the exception safe where it's
not necessarily the case.
15:06:57 The form you know they're they're operating on a different
strains, So that has been pushed back against her static exceptions.
15:07:04 Paper. So so basically it's it's not going anywhere is that
basically the executive summary Oh, rather pessimistic one.
15:07:15 Yeah, okay, Well, that's good to know at least get left
15:07:22 I've also personally said, Well, if you saw an exception you don't
care about your real time deadline, and with that I mean well, I do not in
safety.
15:07:37 So the point is you have to go to your safe fall back before you
scroll your exception, and that is now the reason why you don't care about
your deadline anymore.
15:07:54 Because you do everything that you need to do before your deadline
before you scroll your exception.
15:08:04 Yeah, So I think it really depends on the use case. I think I can
see, use cases where that's the case.
15:08:10 I can also see use cases. where you're like on some kind of
regular callback that you can't just stop right.
15:08:21 And so you just need to, simply, which is the dummy.
15:08:22 Call that more or less, that not on the dummy a safe call back
that works like you fading out, or whatever you have to do that first, and
then you go for the exception.
15:08:39 Yeah, that means, Michael. Yeah, So quite a few quite a few
streams or thoughts there. i'm i'm highly interested in this, because i'm I
care both about the high performance side and the safety side they're
almost like opposite ends
15:08:58 of of the spectrum, sometimes not all the time. Sometimes they do
coincide.
15:09:02 So yeah, too. dead love. Yeah, I get it that's a pretty good
technique standard technique where you do all the real time stuff, right?
15:09:10 And then you, and then you throw the exception. So anything that
requires a real-time response. You get that done, and then you you
15:09:17 You put you then you go to the part that you don't care about with
the deadline is no longer important.
15:09:24 The the hoop paper. I have talked to her to see what his
intentions are, because I care about it in some ways not exactly in this
current form.
15:09:35 Because there's an th the it has an abi incompatibility, because
it causes a different another parameter like the Vv. table parameter that's
hidden to be added to your to your function
15:09:48 call and so in that way it's not it's not yet ideal.
15:09:54 But as far as I understand it, hopes held it back.
15:09:58 In favor of trying to let's get c 23 through first, and then he
might come back to it.
15:10:04 Because this is obviously a big discussion about about exceptions
and how to handle it.
15:10:09 And of course, in this group we've been shepherding through a
paper from a gentleman at
15:10:15 Is he at all now, or something like that? What he did his PHD.
15:10:20 Thesis on deterministic exceptions for embedded systems. and
15:10:24 His name is Jeremy Renwick. i'm sure you can easily Google his
paper.
15:10:28 And anyway, we've reviewed this in this group 2 or 3 times now
trying to find a way building his system of compile time exceptions.
15:10:39 That is deterministic for embedded systems that works with
embedded systems.
15:10:45 Into the C Standard and we've not figured out a way at which, so
far, it's a great experience paper.
15:10:53 The fact of it is that we don't know what it takes to change the
standard, and to to to make that possible.
15:11:04 Even now, I mean, we know that with the c standard exceptions
doesn't have to be built exception.
15:11:11 Information does not have to be built on the on the heap.
15:11:13 It could actually be built on the stack the standard Doesn't
Doesn't say it has to be built on the Hebrew.
15:11:19 It just says it has to be built somewhere it's just that all
compiler vendors have used a heap to build that exception.
15:11:24 Information that's what causes the non-determinism, the memory
allocations right?
15:11:30 The dynamic memories. and this is why but it doesn't have to be,
you know i'm I'm.
15:11:36 I actually wrote an exception system for Ibm compiler is, an I put
it on the heap just because the directions from that time you know, if you
have big iron machines, space is not really a matter of contention so you
could pretty
15:11:53 much use as much space as you want once exception is thrown and
the's what they do that's what they do Once the exception is on you just
start gobbling up space to store all that information and that's what
15:12:03 causes, the the slowdown, the unwinding, the personality routines
all takes time.
15:12:08 And this is what this is. This is why all the exception systems
that we have caters do that kind of that kind of big iron, big mainframe
systems.
15:12:17 Because they have lots of memory they don't cater to embedded
systems where memory is limited because that's not what our bosses told us
to do.
15:12:24 But now it's different. it's getting to the point where you do
have limited resource limited memory, and you do want to do a an exception
system that conforms to that that means putting it on the stack and the
15:12:37 problem, and that's okay, and both herbs paper and Jeremy Reynolds
paper essentially do that they try to put the exception stack frames and
the the exception frame information on the Stack and this is
15:12:52 why they can be much more deterministic they're proven to be
deterministic by data, because they've done benchmarks and all that on on
these kinds of things.
15:13:00 The problem is, no one has built it, even though these things,
this paid these papers at that PHD.
15:13:06 Thesis has been out for 3, or 4 years. now no one has built it,
and I don't anticipate people building it for another 3 to 5 years. so I
just don't see it as being an active solution, even though in
15:13:18 theory, it should all work and multiple people have proven it.
15:13:23 So are you basically saying that this, the problem is not in the
specification.
15:13:28 But the problem is in the kind of compile implementation.
15:13:31 The problem is exactly what i'm saying I don't believe that
interesting Yeah, the specification specification the C standard does not
prohibit you by by by to implement exception on the stack.
15:13:42 I can point to the exact paragraph I gave and talks about this in
details. Say you know there's nothing here that says you have to put it on
the heap.
15:13:49 It's just that by convention everyone has put it on the heap.
15:13:53 So, Guy, I i'll call you in a second but I just wanna reply to
this doesn't that or maybe yeah guy go first.
15:13:59 Maybe it's gonna say actually in discussion with her about the
exceptions.
15:14:05 Paper. the issue of running out of memory has been quite
important. Because at the moment, one problem with throwing exceptions is,
if you're throwing an exception, because you run as memory why?
15:14:16 Do you put the exception but it was observed that the whole out of
memory. Business is pretty much meaningless now, because memory allocation,
It's not simply a case.
15:14:30 You know it's it's often it's simply a case of you know I like you
know marking a page is available for being written to, or something like
that.
15:14:36 I mean you actually run out of memory, or you observeably run out
of memory long after you've made the allocation.
15:14:43 So the actual classifications of situations in which you can throw
exceptions, have been diminishing.
15:14:50 Putting exceptions on, you know the the opening exceptions on the
heat thing.
15:14:55 I I think it's becoming a it's it's sort of a non issue, because
memory just doesn't work in the same way that it works in the fine to you
in the 1980 s when all this was when all this was
15:15:07 pairing. Wait, you mean running out of memory is not an issue,
because obviously obviously putting an exception on the heap, can still,
you know, actually result in a in a memory location that you can where you
can observe the the issue is that what
15:15:23 what when a memory allocation is made then for example if if you
like, you can request medication from of anybody's direct experience
windows.
15:15:35 But I believe the same is true with in the Linux.
15:15:37 You then the allocation doesn't fail until you actually try and
write the memory. but by by which point it's too, it's it's too late you
don't have to the failure happens doesn't happen at
15:15:49 the point of allocation, it happens to the point of use.
15:15:52 So. yeah, what you want to do is show the exception at the point
of allocation. Not at the point.
15:15:56 And That's that's not necessarily something that that can be done
anymore.
15:15:59 But like, from what I understand, like herb saying that if you run
out of memory, that shouldn't be treated as an exception, like basically
like bad Alloc, or something, it should be treated as you run out of
resources behavior is
15:16:15 undefined, and like the whole thing. So I think I think that
actually that is actually very reasonable approach.
15:16:24 But I I'm very curious about like I wanna make a comment about one
other thing that Michael said earlier, because I think that's really
fascinating.
15:16:31 It's not a perspective, I have really heard before that the kind
of issue of non deterministic, exceptions, and I'm not talking about
running out of memory, which is exceptions in general.
15:16:45 The the issue of non-deterministic exceptions is not an issue of
language specification.
15:16:50 It is an issue of tooling and compiler technology.
15:16:53 And so that makes a lot of sense to me I just haven't heard this
before, And so I wonder then what the point is off things like hubs.
15:17:02 Okay, if if the problem is in the implementation of an exception
mechanism rather than in its specification, if we could theoretically do
this today, the only point is that the only point of his paper that we have
a different syntax which
15:17:16 is distinct from the existing mechanism, basically for backwards.
15:17:21 Compatibility purposes. is that like the only interesting thing?
15:17:24 Yeah, Well, there, there's also. the abi break right but but
there's not the point of the paper is to switch people's mindset from
things that.
15:17:33 It has to be only be implemented on this on the heap, and now it
could also be implemented on the stack.
15:17:39 But this putting on a stack causes another, another another kind
of problem, which is that the Abi is going to match your previous calling.
15:17:47 Ap Api. Yeah. So if you do care about backpacks compatibility and
still supporting the old exception mechanism at the same time, Yeah, and
like the Abi, then you need like 2 parallel
15:18:01 mechanisms, which is what Herb has done there, and his proposal.
15:18:04 But you. don't strictly need that if you only care about the
non-determinism.
15:18:09 You don't care about backwards compatibility you don't need a new
syntax for any of this.
15:18:13 Is that my correct like t of understanding but you don't actually
need a new syntax there, and you could just implement that says, Oh, i'm
gonna now compile it using a stack mechanism fine go ahead
15:18:31 fascinating, i'm not going to agree with any other library. Okay,
fine. just go nuts. Well, it feels that to me that maybe the who need it,
and who do things anyway like compiling with f no exceptions, or having
their own
15:18:50 if St. L. or whatever. right, then you might as well give them
another compiler flag, and then it least like the compiler is gonna take
care of this stuff for you.
15:18:59 You don't have to do it yourself. so it feels like that would be a
great solution, actually.
15:19:03 So I hope maybe somebody's gonna somebody's gonna do that Anyway,
debt and I've had signs up
15:19:12 Actually, I I believe, Gcc. has some kind of this compile of like
could be, because Spec.
15:19:23 In 2,001. I think it was when we heard from Ibm the first time
about this table based approach.
15:19:34 At that time everybody was using the tech approach and only
afterwards.
15:19:42 It got into the Itadium Abi And and that is the real problem.
15:19:53 Because, I can remember, I think it was 2,005 or 2,006 at the
beginning of Wt.
15:20:02 21 meeting how a tenant from apple at that time.
15:20:12 Speech with an official statement from Apple and Photoshop.
15:20:21 If the exception, Abi we'll change they will do strongly against
any standard version that contains that one, because that would mean that
the plugins for Photoshop don't work anymore.
15:20:44 For example, and all the the the exception api if you don't have a
pure C interface between 2 components.
15:20:58 The exception Api is a very important part of that interface.
15:21:07 And since this is so, you really need to keep that from the
compiler. point of view, maybe because that is what your customers want,
and that is only 2 for the desktop systems.
15:21:26 As soon as you are on embedded systems where you compile
everything you yourself, anyway, in most times you don't care.
15:21:37 Yeah. for example, if you have queues Steinba queue is all the
plugins very go through the exception.
15:21:46 Interface.
15:21:50 Guy that's absolutely fascinating delt I shall remember that.
15:22:00 I was going to suggest that we might have great ideas about well,
that's just right.
15:22:06 A stack-based exception, Adelaide.
15:22:08 But there's a you know there's a month. you know there's a cost to
doing all these things.
15:22:15 This is possibly going to be too great for our compiler vendors to
15:22:17 That I think with the possible exception of Gcc.
15:22:22 And clang unless people actually start saying that we want a stack
based exception outlet.
15:22:29 Then we're we're not going to get one in the General case, unless
somebody is prepared to write off the clang in Gcc.
15:22:37 And you know, make it obvious, and advertisers extensively one of
the problems we've always had with talking about exceptions simply
measuring the wretched things it's been.
15:22:48 It's always been very hard to compare exception safe and exception
unsafe code, because one makes different assumptions.
15:22:54 Fundamentally all through your code base, based on the decision of
whether or not you're going to use exceptions
15:23:05 Yeah, but I think it's a good point that you know we don't know
whether that's realistic expectation from kind of compiler vendors that
they provide basic exception. so whether it's it's too costly or whether
there is
15:23:19 a market for it because it's it's weird like i've been in the low
agency business now for over a decade, and I have never heard about this
before.
15:23:27 So I wonder if it's just me being completely ignorant or whether
it's just not as widely known as it should be that you know that is
something that's technically possible.
15:23:38 And that's the reason why there is no demand for it currently
15:23:42 Yeah, if I may. you you're probably not that different than most
of us. I myself really only learned of this in the last 4 or 5 years after
I looked at the paper from her I looked at this Guy's.
15:23:55 Phd. pieces. It became clear to me that this was the truth.
15:23:57 I imagine not a lot of people know about this and because it's one
of those paragraph in the standard that is pretty well hidden, and but
there's no there's lots of paragraph in the standard
15:24:08 as well hidden, unless unless we actually, unless we actually
wrote it.
15:24:12 Most of us gonna struggle to find it. But yeah, you know the the
thing is that I I mean, I I suppose.
15:24:23 Yeah, I I suppose that the the thing that I guess I want to
transfer across is I've been looking at this for a while, trying to figure
out what to do with with low latency and I pretty much feel like I I know
15:24:35 we all a lot of the problems are we guys we've already talked a
lot about dynamic memory.
15:24:39 Well, people are using tool allocators static a whole static chunk
to make sure that it's deterministic.
15:24:47 Okay, so that's kind of solvable kind of solvable with exceptions,
because it's so infused in the Standard Library.
15:24:53 It's hard to solve, because it's not just about not using the
exception.
15:24:58 I do use. use a whole different exception. system that stack base,
or somehow you just bracket out the exception, which is why i'm very
curious about the solutions you guys talk about i'm mostly coming from this
from a safety
15:25:09 self-driving call point of view. and yeah, you have to have your
own Stl: Okay, that's not great, but doable.
15:25:15 I guess I can somehow make my own sdl that doesn't use exceptions.
15:25:23 That's not that's happened quite that's happened a lot before, you
know. ea electronic odd had their own Stl.
15:25:29 And I imagine that most of it doesn't use any exceptions and just
use error code or something like that.
15:25:34 The electronic the electronic Sdl: you devoted the exception
throwing problem by simply not implementing things that could error which
was which was awkward.
15:25:47 The main reason why the Estl existed was to deal with memory,
allocation, and fragmentation.
15:25:53 Oh, okay. So they didn't deal with the exception at all. Okay,
it's good.
15:25:57 So i've seen kind of 2 approaches one is you just artificially
restrict your you compile with exceptions, but you just artificially
restrict yourself to the subset of Stl.
15:26:10 That that can never throw any right and Then you're right kind of
your own facilities on top, and you don't use exceptions, or you just throw
the whole scl out the window you write your own and then you
15:26:21 compile with no exceptions, and as an error mechanism, you use
either return codes, or you use something like stood expected, which, you
know is now coming in C 3.
15:26:33 But I think pretty much every like application framework under the
sun has something similar already.
15:26:39 It's just that now. we're gonna get it as a vocabulary type which
makes it useful across api boundaries, which is another very interesting
topic.
15:26:45 I'm not going to go into now. but yeah this is kind of what I've
seen people do in this space.
15:26:53 And I don't know anything about automotive for finance.
15:26:54 I'm talking about games audio. you know these kinds of kind of
consumer, like thing like low latency things that kind of run consumer
software on consumer hardware.
15:27:06 Roughly speaking, using a off the shelf kind of operating system
like windows.
15:27:13 Mac Linux, rather than, you know, something like their metal or
the real time operating system, or something.
15:27:19 Guy. So one final point, I guess, on the whole, rewriting the sdl
business.
15:27:26 I think we're writing sdl is a really bad idea, because the
problem that you're solving is that the containers that the str provides
are not up or not containers that you want to use
15:27:37 rewriting vectors that it doesn't throw exceptions means that you
don't have a vector anymore.
15:27:41 You've got something else that looks like a vector that actually
It's still a vector as we understand it's in the standard.
15:27:48 And you know we have an object called a dynamic array, which but
he's like back to. but it doesn't throw exceptions, and that's what we use
in the general case instead.
15:28:03 I think you know, rewriting the St. I was throwing the baby out
with the bathwater.
15:28:07 The Stl does carefully describe what the expected behavior is in
containers, and that behavior can be met without throwing exceptions.
15:28:15 You know, the all all the you know the the type tests that or the
the typewriter is that the container should have, don't imply exception
handling, So it's just it's it's quite used to right around
15:28:27 containers that that operate with the rest of the you know, with
the algorithm, for example.
15:28:33 So that is true. But for example, in Audio, we have another
constraint, which is, we can never ever make a dynamic memory allocation in
the real-time path, and Right? you're in consent.
15:28:44 So then, yeah, you write our own container which Isn't. A. vector
which is something like static, vector which has a static capacity where
everything is like inside the object.
15:28:52 And then the Ab. the Api changes then you can't have a pushback
anymore.
15:28:56 You have to have a try push back which can fail right and stuff
like that.
15:28:59 So so. yeah, I guess you're right like either you you can meet
your needs with like the existing api or you can't, and that's kind of a
decision you have to make at that point
15:29:15 That was good information. I guess I guess my summary Is that okay?
15:29:18 So you can't rely on this static mechanism coming anytime soon.
15:29:22 I mean it works but it's nobody nobody's implemented It'll
probably take somebody 3 or 5 years to do it. You can You can marginally
rely on the idea that you can bracket yourself away from the stuff that
15:29:32 throws exception, which is kind of what I think Guys trying to
point me to.
15:29:36 And I agree. Rewriting The Sel is just growing the whole baby out
with the bathwater. not good.
15:29:40 So what what else? What choices do people have left Now at this
point You still have to.
15:29:45 If you're gonna use the standard library, and assuming that's
that's a fundamental requirement, because you can always write your own
application and don't use exceptions.
15:29:51 But that's not that's Not the world We live in especially in the
commercial gaming.
15:29:55 I put hft world people will have to find a way to Marshall,
sometimes using error codes in the time.
15:30:04 Critical path, and sometimes using exceptions in the one in a
1,000 case, time, non critical path, exceptions, as I understand, it, was
never supposed to be used in in any kind of hard, or even soft, real-time
requirements.
15:30:21 That was never be honest design. It was designed for the one in a
1,000 case.
15:30:24 That's that an error happens so most errors are not one in a 1,000.
15:30:28 They're one in a 100 or one in 10 so i'm just thinking that. is
there a way that we can you know.
15:30:33 Do we? you know well like, like, exceptions like sorry like like
variants and optional where somehow you can take an exception, and somehow
it's maybe you can still pass out an error code if you need it so they can
keep
15:30:49 the software running. I don't know like something like that I
think we have to have an an immediate like an like a current best practice
solution for people, because we know that the other great solutions are not
gonna land anytime soon So I
15:31:02 think I have one possible answer to that. But Guy had his end
first.
15:31:06 Sure I wanted to point out that exceptions solve the problem of
unexpected things happening.
15:31:16 Now in the game, domain the it's the kinds of inputs that the game
has tremendously limited, you know, tremendously limited.
15:31:29 Basically those that the number deterministic inputs you'll get
into a game is from a controller.
15:31:34 Maybe from a file system and that's it so actually throwing
exceptions.
15:31:41 It's kind of outside of some of the expected scope of gain
developments, anyway.
15:31:47 This doesn't answer your question, Michael, but I just wanted but
I just wanted to make clear that you know you're using exceptions.
15:31:54 Need to be a a for exceptional cases from outside of your system.
15:31:57 Okay, that that's good data. Yeah, yeah I mean for a call. We have
to kind of figure out all the possible it exception cases.
15:32:05 But but yeah, no, I I I understand. Yeah, so another approach
that, I think is really cool, is, if you use something like stood expected
which we now have in understand that you can actually build a lot of
things, like especially since it's a
15:32:22 vocabulary type that you can pass across like api boundaries, and,
like different libraries, can agree on on that being kind of the expected
unexpected type that we now use.
15:32:31 There's like really cool patterns, you can do and i've shown some
of them in my cbp cone keynote a few weeks ago, where you know, you can
kind of pass pass expected across interfaces you can use them
15:32:45 in algorithms where, like, you can make them generic on the type
of error that the expected contains.
15:32:52 You can have these like things. We have an expected of a variant
of different errors.
15:32:59 And then you! you kind of have different errors coming in from
different layers, and, like you kind of collate them where you do your
error handling, and you can do is to visit on the variant, and then the if
that looks very much like a
15:33:08 try catch block, except it actually forces you to like catch all
the different error types.
15:33:12 Otherwise you get a compiler error so that's like really really
interesting.
15:33:17 Kind of design patterns there, or like coding patterns there, that
give you a lot of the functionality of expect exceptions, accept like the
deterministic and efficient with the downside that there was just a lot of
15:33:31 syntactic overhead, right? because it's a library type.
15:33:35 Yeah, you have to like explicitly construct these error types and
stuff like that.
15:33:38 So it's not syntactically as neat as exceptions are, and it's and
there's no like separate kind of mechanism like you can throw out of a
function. you can still only leave a function through a return statement
right
15:33:49 so there's a lot more syntactic noise.
15:33:53 But you get effectively like a very, very similar behavior, and
sometimes even better stuff, like compile time checked like you.
15:34:00 You caught all the cases and things like that, and you can.
15:34:02 You can do that with it. expected quite, quite nicely. Thank you.
15:34:09 So i'm kind of a bit conscious of time you have 25 min left, and
wondering if there was anything else on the agenda that people want us to
discuss today, or if not, we can just like let this discussion continue
until you run
15:34:23 out of time. But yeah, that on the agenda they were like a few
papers here which I don't think we have the authors in for
15:34:34 There. there was meant to be a discussion about games, topics. No,
those those are for different week, different months.
15:34:41 Every month I alternate between different topics like games and and
15:34:48 Okay, so none of these are relevant for today.
15:34:54 Okay, that's great. Then we can just keep this discussion going
for another 20 min, and then we can stop. If people If people have had
enough of us.
15:35:02 Yeah, let's just see if there are any more hands I think it's a
very fascinating discussion.
15:35:07 By the way, I i'm learning a lot here and I will definitely look
at that transcript later.
15:35:13 As well. so thanks for everybody who's but Yeah, no thanks, thank
you.
15:35:19 Thank you timer for volunteering you've done you've done a great
job.
15:35:21 This is fantastic. i'm actually back but I have to leave again
soon.
15:35:27 So i'm happy If we stop here. so and then, that way. I can just
save the transcript right doesn't. Oh, can I have a ask you a question?
15:35:35 Are you gonna publish the transcript somewhere? Is that going to
be accessible?
15:35:38 Oh, oh, oh, because this is not a face to face. I I emailed the
entire transcript to the reflector.
15:35:45 : Yeah. So yeah, okay, Cool: alright. So is there any more
discussion on this?
15:35:55 Otherwise we can can wrap up alright. Thank you. Thanks.
15:36:02 Everybody. yeah, normally next month, so the next. meeting is
going to be on December the seventh.
15:36:14 And it will be game I believe it's also a Wednesday.
15:36:17 It's always wednesday isn't it it's also a Wednesday.
15:36:26 It's always wednesdays, isn't it but because time change
differently. i'm always just using my own time, and just let everyone else
convert to it, whatever their time is.
15:36:36 Yeah. So So maybe we like another approach is to always give the
time in Utc, because that's like really unambiguous.
15:36:41 And everybody knows what that means. But yeah I don't know if I
had to look at that.
15:36:46 I think Utc moves around, too, because of stay like savings.
15:36:50 Time. yeah time to time just move around so one thing you see,
alright.
15:37:00 So then i'll see some of you in kona and then i'll see all of you
hopefully on December the seventh.
15:37:09 And Thank you very much for this discussion. cheers guys, Goodbye.

On Wed, Oct 12, 2022 at 12:13 PM Patrice Roy <patricer_at_[hidden]> wrote:

> I'll be in class today during the meeting so I cannot make it, sadly :(
>
> Le mar. 11 oct. 2022 à 23:31, Michael Wong via SG14 <sg14_at_[hidden]>
> a écrit :
>
>> Topic: SG14 Low Latency Monthly This meeting is focused on Low Latency.
>> There were several Low latency discussions on the reflector this month and
>> this would be a good time to review and summarize to see if a paper can be
>> jointly published. Alternatively, we can continue with the Games paper that
>> was started at CPPCON.
>>
>>
>> Hi,
>>
>> Michael Wong is inviting you to a scheduled Zoom meeting.
>>
>> Topic: SG14 monthly
>> Time: 2nd Wednesdays 02:00 PM Eastern Time (US and Canada)
>> Every month on the Second Wed,
>>
>> Join from PC, Mac, Linux, iOS or Android:
>> https://iso.zoom.us/j/93151864365?pwd=aDhOcDNWd2NWdTJuT1loeXpKbTcydz09
>> Password: 789626
>>
>> Or iPhone one-tap :
>> US: +12532158782,,93151864365# or +13017158592,,93151864365#
>> Or Telephone:
>> Dial(for higher quality, dial a number based on your current
>> location):
>> US: +1 253 215 8782 or +1 301 715 8592 or +1 312 626 6799 or +1
>> 346 248 7799 or +1 408 638 0968 or +1 646 876 9923 or +1 669 900 6833
>> or 877 853 5247 (Toll Free)
>> Meeting ID: 931 5186 4365
>> Password: 789626
>> International numbers available: https://iso.zoom.us/u/abRrVivZoD
>>
>> Or Skype for Business (Lync):
>> https://iso.zoom.us/skype/93151864365
>>
>> Agenda:
>>
>> 1. Opening and introduction
>>
>> ISO Code of Conduct
>> <
>>
>> https://isotc.iso.org/livelink/livelink?func=ll&objId=20882226&objAction=Open&nexturl=%2Flivelink%2Flivelink%3Ffunc%3Dll%26objId%3D20158641%26objAction%3Dbrowse%26viewType%3D1
>> *>*
>>
>> ISO patent policy.
>>
>> https://isotc.iso.org/livelink/livelink/fetch/2000/2122/3770791/Common_Policy.htm?nodeid=6344764&vernum=-2
>>
>> IEC Code of Conduct:
>>
>> https://www.iec.ch/basecamp/iec-code-conduct-technical-work
>>
>> WG21 Code of Conduct:
>>
>>
>> https://isocpp.org/std/standing-documents/sd-4-wg21-practices-and-procedures
>>
>> 1.1 Roll call of participants
>>
>> 1.2 Adopt agenda
>>
>> 1.3 Approve minutes from previous meeting, and approve publishing
>> previously approved minutes to ISOCPP.org
>>
>> 1.4 Action items from previous meetings
>>
>> 2. Main issues (125 min)
>>
>> 2.1 General logistics
>>
>> CPPCON minutes:
>> https://wiki.edg.com/bin/view/Wg21virtual2022-07/SG14
>>
>> Future meeting plans
>>
>> *No call Nov due to Kona F2F:
>> *Dec 7, 2022 02:00 PM ET Games
>> *Jan 11, 2022 02:00 PM ET: Embedded
>> *Feb 8, 2022 02:00 PM ET: Finance/low Latency
>> *Mar 8, 2022 02:00 PM ET: Games
>>
>> 2.2 Paper reviews
>> Discussion on Embedded:
>> Review latest mailings:
>> P2532 Removing exception_ptr from the receivers concept
>> Based on the last meeting and the discussions here.
>> P2544 C++ Exceptions are becoming more and more problematic
>> We might want to chime in here.
>> /Paul
>> P. S. P2327 de-deprecating volatile received a "consensus" straw poll.
>>
>>
>> Discussion on Low Latency/Finance topics
>>
>> http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p1839r4.pdf
>>
>> Patrice's paper on games.
>>
>> P2300
>> Swift
>>
>>
>>
>> Discussion about Games topics:
>>
>> P2388R1 - Minimum Contract Support: either Ignore or Check_and_abort
>> <http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p2388r1.html>
>>
>> Patrice's WIP on Games issues.
>>
>> Finance topics from July 14, 2021.
>>
>> https://lists.isocpp.org/sg14/2021/06/0636.php
>>
>> https://lists.isocpp.org/sg14/2021/07/0642.php
>>
>> 2.2.1 any other proposal for reviews?
>>
>> Deterministic Exception for Embedded by James Renwick
>>
>> https://www.pure.ed.ac.uk/ws/portalfiles/portal/78829292/low_cost_deterministic_C_exceptions_for_embedded_systems.pdf
>>
>> Freestanding?
>>
>> SG14/SG19 features/issues/defects:
>>
>>
>> https://docs.google.com/spreadsheets/d/1JnUJBO72QVURttkKr7gn0_WjP--P0vAne8JBfzbRiy0/edit#gid=0
>>
>> 2.3 Domain-specific discussions
>>
>> 2.3.1 SIG chairs
>>
>> - Embedded Programming chairs: Ben Craig, Wouter van Ooijen and Odin
>> Holmes, John McFarlane
>>
>> - Financial/Trading chairs: Staffan TjernstrÃm
>> Carl Cooke, Neal Horlock,
>> - Games chairs: Rene Riviera, Guy Davidson and Paul Hampson, Patrice
>> Roy
>>
>> - Linear Algebra chairs: Bob Steagall, Mark Hoemmen, Guy Davidson
>>
>> 2.4 Other Papers and proposals
>>
>> 2.5 Future F2F meetings:
>>
>> 2.6 future C++ Standard meetings:
>> https://isocpp.org/std/meetings-and-participation/upcoming-meetings
>>
>> -
>>
>> 3. Any other business
>> Reflector
>> https://lists.isocpp.org/mailman/listinfo.cgi/sg14
>> As well as look through papers marked "SG14" in recent standards committee
>> paper mailings:
>> http://open-std.org/jtc1/sc22/wg21/docs/papers/2015/
>> http://open-std.org/jtc1/sc22/wg21/docs/papers/2016/
>>
>> Code and proposal Staging area
>> https://github.com/WG21-SG14/SG14
>> 4. Review
>>
>> 4.1 Review and approve resolutions and issues [e.g., changes to SG's
>> working draft]
>>
>> 4.2 Review action items (5 min)
>>
>> 5. Closing process
>>
>> 5.1 Establish next agenda
>>
>> 5.2 Future meeting
>>
>>
>> *No call Nov due to Kona F2F:
>> *Dec 7, 2022 02:00 PM ET Games
>> *Jan 11, 2022 02:00 PM ET: Embedded
>> *Feb 8, 2022 02:00 PM ET: Finance/low Latency
>> *Mar 8, 2022 02:00 PM ET: Games
>> _______________________________________________
>> SG14 mailing list
>> SG14_at_[hidden]
>> https://lists.isocpp.org/mailman/listinfo.cgi/sg14
>>
>