Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Tuesday, 25 August 2015 at 05:12:55 UTC, Laeeth Isharc wrote: What's the best reference to learn more about PGAS? I've seen a few presentations, https://www.osc.edu/sites/osc.edu/files/staff_files/dhudak/pgas-tutorial.pdf http://www.inf.unideb.hu/~fazekasg/english/New_Programming_Paradims/PGAS.pdf papers, http://www.cs.rochester.edu/u/cding/amp/papers/full/The%20Asynchronous%20Partitioned%20Global%20Address%20Space%20Model.pdf and a stackoverflow question http://stackoverflow.com/questions/31992761/what-are-the-differences-between-tls-and-pgas Not sure what is the best one.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
More info on the Go 1.5 concurrent GC, a classic one: https://blog.golang.org/go15gc
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Tuesday, 25 August 2015 at 07:18:24 UTC, Ola Fosheim Grøstad wrote: On Tuesday, 25 August 2015 at 05:09:56 UTC, Laeeth Isharc wrote: On Monday, 24 August 2015 at 21:57:41 UTC, rsw0x wrote: [...] Horses for courses ? Eg for Andy Smith's problem of processing trade information of tens of gigs where Python was choking, I guess nobody in their right mind would use Rust. I don't think there is much difference between C, D or Rust in terms of computing. The core semantics are similar. With Rust you have the additional option of linear type checking. But Rust programmers of course want to use idiomatic linear typing as much as possible and that makes designing graph-like structures a challenge. An option implies you can turn it off, has this changed since the last time I used Rust?(admittedly, a while back) Memory safety doesn't seem like it's the top priority for scientific computing as much as fast turnarouds and performance... in my opinion, anyways.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Tuesday, 25 August 2015 at 05:09:56 UTC, Laeeth Isharc wrote: On Monday, 24 August 2015 at 21:57:41 UTC, rsw0x wrote: On Monday, 24 August 2015 at 21:20:39 UTC, Russel Winder wrote: For Python and native code, D is a great fit, perhaps more so that Rust, except that Rust is getting more mind share, probably because it is new. I'm of the opinion that Rust's popularity will quickly die when people realize it's a pain to use. Horses for courses ? Eg for Andy Smith's problem of processing trade information of tens of gigs where Python was choking, I guess nobody in their right mind would use Rust. I don't think there is much difference between C, D or Rust in terms of computing. The core semantics are similar. With Rust you have the additional option of linear type checking. But Rust programmers of course want to use idiomatic linear typing as much as possible and that makes designing graph-like structures a challenge.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Tuesday, 25 August 2015 at 07:21:13 UTC, rsw0x wrote: An option implies you can turn it off, has this changed since the last time I used Rust?(admittedly, a while back) Rust supports other reference types, so you decide by design whether you want to use linear typing or not?
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Monday, 24 August 2015 at 21:20:39 UTC, Russel Winder wrote: The issue here for me is that Chapel provides something that C, C++, D, Rust, Numba, NumPy, cannot – Partitioned Global Address Space (PGAS) programming. This directly attacks the multicore/multiprocessor/cluster side of computing, but not the GPGPU side, at least not per se. Yes, I agree that highly parallel batch programming requires a specialized approach. Though for most applications I think something like Pony will fit better. But they have to figure out how to do migration and load balancing I suppose. Javascript is moving towards programming using futures/promises, it is already built into Chrome. While that isn't actors per se, I think the computing objects approach will be more familiar for programmers over time and that will open up for actor-based languages too. Maybe.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Monday, 24 August 2015 at 21:20:39 UTC, Russel Winder wrote: For Python and native code, D is a great fit, perhaps more so that Rust, except that Rust is getting more mind share, probably because it is new. I'm of the opinion that Rust's popularity will quickly die when people realize it's a pain to use.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Sun, 2015-08-23 at 19:42 +, via Digitalmars-d-learn wrote: […] Yes, of course it is, but given it's typical use context I find it odd that they didn't go more towards higher level constructs. For me Go displaces Python where more speed is required, though I wish it was more pythonic… (neither C++, Rust or D are really eligible) I wondered about the Go/Python issue last year, and even did a London Python Group session on the topic. Clearly organizations like Canonical are displacing Python with Go for some of their projects, almost certainly rightly so: as you say Go has duck typing yet is strongly statically types, where Python (via PEP 484 and MyPy) is only now getting type hinting. However I think we will see that Go and Python are actually getting traction in different spaces and so are not competing as much as many would have us believe. For Python and native code, D is a great fit, perhaps more so that Rust, except that Rust is getting more mind share, probably because it is new. It would be good to get rid of C and C++ as the languages of Python extensions. Of course systems like Numba change the Python performance game, which undermines D's potential in the Python-verse, as it does C and C++. Currently I am investigating Python/Numba/Chapel as the way of doing performance computing. Anyone who just uses Python/NumPy/SciPy is probably not doing performance computing, NumPy is so slow (*). The issue here for me is that Chapel provides something that C, C++, D, Rust, Numba, NumPy, cannot – Partitioned Global Address Space (PGAS) programming. This directly attacks the multicore/multiprocessor/cluster side of computing, but not the GPGPU side, at least not per se. (*) Which comment allows this piece to be attacked as I have been attacking similar comments elsewhere ;-) -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
. Of course systems like Numba change the Python performance game, which undermines D's potential in the Python-verse, as it does C and C++. Currently I am investigating Python/Numba/Chapel as the way of doing performance computing. Anyone who just uses Python/NumPy/SciPy is probably not doing performance computing, NumPy is so slow (*). Can you elaborate ? The issue here for me is that Chapel provides something that C, C++, D, Rust, Numba, NumPy, cannot – Partitioned Global Address Space (PGAS) programming. This directly attacks the multicore/multiprocessor/cluster side of computing, but not the GPGPU side, at least not per se. What's the best reference to learn more about PGAS? Thanks. Laeeth.
Re: Role of D in Python and performance computing [was post on using go 1.5 and GC latency]
On Monday, 24 August 2015 at 21:57:41 UTC, rsw0x wrote: On Monday, 24 August 2015 at 21:20:39 UTC, Russel Winder wrote: For Python and native code, D is a great fit, perhaps more so that Rust, except that Rust is getting more mind share, probably because it is new. I'm of the opinion that Rust's popularity will quickly die when people realize it's a pain to use. Horses for courses ? Eg for Andy Smith's problem of processing trade information of tens of gigs where Python was choking, I guess nobody in their right mind would use Rust. But maybe D isn't quite yet what you would choose for a highly complex mass consumer product like a browser? Btw had a nice little test the other day. I haven't yet done much with intraday data - I only just lately took the FX data I had sitting around and pulled it into my data server.Someone asked me a question about something in that domain - would have taken them a few weeks at least to begin to get an answer. I have never been the fastest programmer, but took a couple hundred lines and three hours to write the analysis from scratch, and half the time was figuring out how to get it to display nicely in Mathgl. Took less than 10 minutes to run on my home machine using debug Dmd and pulling data from my server over the Internet with a spinning drive (so probably could get it down to a minute or two). I know it takes a big Wall Street firm an hour to run the same task. That gets in the way of using rapid iteration to explore the data.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 06:54:43 UTC, Ola Fosheim Grøstad wrote: On Saturday, 22 August 2015 at 06:48:48 UTC, Russel Winder wrote: But one that Google are entirely happy to fully fund. Yes, they have made Go fully supported on Google Cloud now, so I think it is safe to say that Google management is backing Go fully. I'm kinda hoping for Go++... The other day I thought it'd be hilarious if I did a Bjarne and wrote a preprocessor to generate Go code that would accept a superset of Go syntax but added generics, function overloading, etc. And, of course, called it Go++. Alas, 'tis too much work for just the lulz. I'd rather spend the time making D better. Atila
Re: post on using go 1.5 and GC latency
On Sunday, 23 August 2015 at 12:49:35 UTC, Russel Winder wrote: You are mixing too many factors here. General purpose has nothing to do with performance, it is to do with can the language describe most if not all forms of computation. Go is a general purpose programming language just like C, C++, D, Rust, Haskell, OCaml. Yes, of course it is, but given it's typical use context I find it odd that they didn't go more towards higher level constructs. For me Go displaces Python where more speed is required, though I wish it was more pythonic… (neither C++, Rust or D are really eligible)
Re: post on using go 1.5 and GC latency
On Sun, 2015-08-23 at 11:26 +, rsw0x via Digitalmars-d-learn wrote: […] https://groups.google.com/forum/#!msg/golang -dev/pIuOcqAlvKU/C0wooVzXLZwJ 25-50% performance decrease across the board in 1.4 with the addition of write barriers, to an already slow language. Garbage collection is a hard problem for performance oriented code. I am sure someone will come up with a way of improving the Go one. Across the board though does worry me, my codes never get a sweep. D's garbage collector could also do with some work. random benchmarks of Go performing 3x(+) slower than C/C++/D, some of these predate Go 1.4. https://github.com/kostya/benchmarks https://benchmarksgame.alioth.debian.org/u64/benchmark.php?test=alll ang=golang2=gccdata=u64 https://togototo.wordpress.com/2013/07/23/benchmarking-level -generation-go-rust-haskell-and-d/ (gcc-go performed the _worst_) https://togototo.wordpress.com/2013/08/23/benchmarks-round-two -parallel-go-rust-d-scala-and-nimrod/ (and again) https://github.com/logicchains/LPATHBench/blob/master/writeup.md (once again, Go is nowhere near C/C++/D/Rust. Where is it? Hanging out with C#/Mono.) Thanks for those pointers, I shall have a look at them. Sadly though not for a couple of weeks due to various commitments. Go is slow. These aren't cherrypicked, just random samples from a quick Googling. But why use absolutes. Go may be slow for you in your context, but that doesn't mean the observation applies everywhere (note the interesting turn of phrase). For my (admittedly small) codes that do not cause a garbage collect and are basically just a loop, I find things are fine. So for a π approximation sequential code using 64-bit: C, gcc -O3: 8.847241 C++, gcc -O3: 8.916043 Fortran, gfortran -O3: 8.893000 D, ldc -O -release: 8.722329 D, dmd -O -release: 8.787744 Rust, cargo --release: 8.715818 Go, gccgo: 8.823525 Go, 6g: 8.824643 Go is definitely not slow there then. Thus Go is not slow. Where is Go performing C-level speeds at? D claims this, and D shows it does. Go falls into the fast enough category, because it is _not_ a general purpose programming language. So unless multiple randomly sampled benchmarks are all wrong, I'm going to stick with 'Go is slow.' You are mixing too many factors here. General purpose has nothing to do with performance, it is to do with can the language describe most if not all forms of computation. Go is a general purpose programming language just like C, C++, D, Rust, Haskell, OCaml. If for you Go is not performant enough, that is fine. But for many people in various contexts, Go is more the comparable with other languages. This is not just Go is fast enough, but Go is as fast as any other option. -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Sat, 2015-08-22 at 11:06 +, Laeeth Isharc via Digitalmars-d -learn wrote: […] Builds in Go 1.5 will be slower by a factor of about two. The automatic translation of the compiler and linker from C to Go resulted in unidiomatic Go code that performs poorly compared to well-written Go. Analysis tools and refactoring helped to improve the code, but much remains to be done. Further profiling and optimization will continue in Go 1.6 and future releases. For more details, see these slides and associated video. This is about compiler performance, not about generated code performance. Anyone interested in performance with Go currently uses gccgo: the standard Go compiler does not generate particularly well optimized code. This has been a resourcing choice to date, it is not a failing. gccgo on the other hand makes use of the whole GCC optimization chain. On the other hand gc is blindingly fast at compilation compared to gccgo. This seems reminiscent of dmd vs. ldc and gdc! -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Sat, 2015-08-22 at 09:27 +, rsw0x via Digitalmars-d-learn wrote: […] The performance decrease has been there since 1.4 and there is no way to remove it - write barriers are the cost you pay for concurrent collection. Go was already much slower than other compiled languages, now it probably struggles to keep up with mono. I know Walter hates it when people mention the word but: benchmarks. As soon as someone say things like it probably struggles to keep up with mono further discussion of the topic is probably not worth entertaining without getting some agreed codes and running them all on the same machine. I agree the standard Go compiler generates not well optimized code, but gccgo generally does, and generally performs at C-level speeds. Of course Java often performs far better than that, and often fails to. You have to be careful with benchmarking and performance things generally. -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 12:48:31 UTC, rsw0x wrote: The problem with D's GC is that there's no scaffolding there for it, so you can't really improve it. At best you could make the collector parallel. If I had the runtime hooks and language guarantees I needed I'd begin work on a per-thread GC immediately. If you had a fiber local reference type and some guarantees related to that, you probably could do a per-fiber GC and collect when fibers are waiting.
Re: post on using go 1.5 and GC latency
On Sunday, 23 August 2015 at 11:06:20 UTC, Russel Winder wrote: On Sat, 2015-08-22 at 09:27 +, rsw0x via Digitalmars-d-learn wrote: […] The performance decrease has been there since 1.4 and there is no way to remove it - write barriers are the cost you pay for concurrent collection. Go was already much slower than other compiled languages, now it probably struggles to keep up with mono. I know Walter hates it when people mention the word but: benchmarks. As soon as someone say things like it probably struggles to keep up with mono further discussion of the topic is probably not worth entertaining without getting some agreed codes and running them all on the same machine. I agree the standard Go compiler generates not well optimized code, but gccgo generally does, and generally performs at C-level speeds. Of course Java often performs far better than that, and often fails to. You have to be careful with benchmarking and performance things generally. https://groups.google.com/forum/#!msg/golang-dev/pIuOcqAlvKU/C0wooVzXLZwJ 25-50% performance decrease across the board in 1.4 with the addition of write barriers, to an already slow language. random benchmarks of Go performing 3x(+) slower than C/C++/D, some of these predate Go 1.4. https://github.com/kostya/benchmarks https://benchmarksgame.alioth.debian.org/u64/benchmark.php?test=alllang=golang2=gccdata=u64 https://togototo.wordpress.com/2013/07/23/benchmarking-level-generation-go-rust-haskell-and-d/ (gcc-go performed the _worst_) https://togototo.wordpress.com/2013/08/23/benchmarks-round-two-parallel-go-rust-d-scala-and-nimrod/ (and again) https://github.com/logicchains/LPATHBench/blob/master/writeup.md (once again, Go is nowhere near C/C++/D/Rust. Where is it? Hanging out with C#/Mono.) Go is slow. These aren't cherrypicked, just random samples from a quick Googling. Where is Go performing C-level speeds at? D claims this, and D shows it does. Go falls into the fast enough category, because it is _not_ a general purpose programming language. So unless multiple randomly sampled benchmarks are all wrong, I'm going to stick with 'Go is slow.'
Re: post on using go 1.5 and GC latency
On Sat, 2015-08-22 at 07:30 +, rsw0x via Digitalmars-d-learn wrote: […] because Go is not a general purpose language. Not entirely true. Go is a general purpose language, it is a successor to C as envisioned by Rob Pike, Russ Cox, and others (I am not sure how much input Brian Kernighan has had). However, because of current traction in Web servers and general networking, it is clear that that is where the bulk of the libraries are. Canonical also use it for Qt UI applications. I am not sure of Google real intent for Go on Android, but there is one. A concurrent GC for D would kill D. Go programs saw a 25-50% performance decrease across the board for the lower latencies. They also saw a 100% increase in performance when it was rewritten, and a 20% fall with this latest rewrite. I anticipate great improvement for the 1.6 rewrite. I am surprised they are retaining having only a single garbage collector: different usages generally require different garbage collection strategies. Having said that Java is moving from having four collectors, to having one, it is going to be interesting to see if G1 meets the needs of all JVM usages. D could make some very minor changes and be capable of a per-thread GC with none of these performance drawbacks, but nobody seems very interested in it. Until some organization properly funds a suite of garbage collectors for different performance targets, you have what there is. -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 09:16:32 UTC, Russel Winder wrote: On Sat, 2015-08-22 at 07:30 +, rsw0x via Digitalmars-d-learn wrote: [...] Not entirely true. Go is a general purpose language, it is a successor to C as envisioned by Rob Pike, Russ Cox, and others (I am not sure how much input Brian Kernighan has had). However, because of current traction in Web servers and general networking, it is clear that that is where the bulk of the libraries are. Canonical also use it for Qt UI applications. I am not sure of Google real intent for Go on Android, but there is one. [...] They also saw a 100% increase in performance when it was rewritten, and a 20% fall with this latest rewrite. I anticipate great improvement for the 1.6 rewrite. I am surprised they are retaining having only a single garbage collector: different usages generally require different garbage collection strategies. Having said that Java is moving from having four collectors, to having one, it is going to be interesting to see if G1 meets the needs of all JVM usages. [...] Until some organization properly funds a suite of garbage collectors for different performance targets, you have what there is. The performance decrease has been there since 1.4 and there is no way to remove it - write barriers are the cost you pay for concurrent collection. Go was already much slower than other compiled languages, now it probably struggles to keep up with mono.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 09:16:32 UTC, Russel Winder wrote: On Sat, 2015-08-22 at 07:30 +, rsw0x via Digitalmars-d-learn wrote: […] because Go is not a general purpose language. Not entirely true. Go is a general purpose language, it is a successor to C as envisioned by Rob Pike, Russ Cox, and others (I am not sure how much input Brian Kernighan has had). However, because of current traction in Web servers and general networking, it is clear that that is where the bulk of the libraries are. Canonical also use it for Qt UI applications. I am not sure of Google real intent for Go on Android, but there is one. A concurrent GC for D would kill D. Go programs saw a 25-50% performance decrease across the board for the lower latencies. They also saw a 100% increase in performance when it was rewritten, and a 20% fall with this latest rewrite. I anticipate great improvement for the 1.6 rewrite. I am surprised they are retaining having only a single garbage collector: different usages generally require different garbage collection strategies. Having said that Java is moving from having four collectors, to having one, it is going to be interesting to see if G1 meets the needs of all JVM usages. D could make some very minor changes and be capable of a per-thread GC with none of these performance drawbacks, but nobody seems very interested in it. Until some organization properly funds a suite of garbage collectors for different performance targets, you have what there is. I didn't mean to start again the whole GC and Go vs D thing. Just that one ought to know the lay of the land as it develops. Out of curiosity, how much funding is required to develop the more straightforward kind of GCs ? Or to take what's been done and make it possible for others to use? It needn't be a single organisation I would think if there are many that would benefit and one doesn't get bogged down in a mentality of people worrying about possibly spurious free rider problems. Since the D Foundation seems under way, it seems worth asking the question first and thinking about goals without worrying for now about what seems realistic.
Re: post on using go 1.5 and GC latency
On 8/22/2015 10:47 PM, Laeeth Isharc wrote: On Saturday, 22 August 2015 at 09:16:32 UTC, Russel Winder wrote: On Sat, 2015-08-22 at 07:30 +, rsw0x via Digitalmars-d-learn wrote: […] because Go is not a general purpose language. Not entirely true. Go is a general purpose language, it is a successor to C as envisioned by Rob Pike, Russ Cox, and others (I am not sure how much input Brian Kernighan has had). However, because of current traction in Web servers and general networking, it is clear that that is where the bulk of the libraries are. Canonical also use it for Qt UI applications. I am not sure of Google real intent for Go on Android, but there is one. A concurrent GC for D would kill D. Go programs saw a 25-50% performance decrease across the board for the lower latencies. They also saw a 100% increase in performance when it was rewritten, and a 20% fall with this latest rewrite. I anticipate great improvement for the 1.6 rewrite. I am surprised they are retaining having only a single garbage collector: different usages generally require different garbage collection strategies. Having said that Java is moving from having four collectors, to having one, it is going to be interesting to see if G1 meets the needs of all JVM usages. D could make some very minor changes and be capable of a per-thread GC with none of these performance drawbacks, but nobody seems very interested in it. Until some organization properly funds a suite of garbage collectors for different performance targets, you have what there is. I didn't mean to start again the whole GC and Go vs D thing. Just that one ought to know the lay of the land as it develops. Out of curiosity, how much funding is required to develop the more straightforward kind of GCs ? Or to take what's been done and make it possible for others to use? It needn't be a single organisation I would think if there are many that would benefit and one doesn't get bogged down in a mentality of people worrying about possibly spurious free rider problems. Since the D Foundation seems under way, it seems worth asking the question first and thinking about goals without worrying for now about what seems realistic. I believe the hardest part is finding somebody can and willing to work on it. For example I'm willing but I don't know how and there are people willing with a job and can do it. But cannot dedicated time because of money. Really it comes down to having a budget and if somebody says hey I'll do x, y and z features to pay them for their time as they do it. Even if they only do one small feature which takes a week.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 07:30:23 UTC, rsw0x wrote: On Saturday, 22 August 2015 at 06:48:48 UTC, Russel Winder wrote: On Fri, 2015-08-21 at 10:47 +, via Digitalmars-d-learn wrote: Yes, Go has sacrificed some compute performance in favour of latency and convenience. They have also released GC improvement plans for 1.6: https://docs.google.com/document/d/1kBx98ulj5V5M9Zdeamy7v6ofZXX3yPziA f0V27A64Mo/edit It is rather obvious that a building a good concurrent GC is a time consuming effort. But one that Google are entirely happy to fully fund. because Go is not a general purpose language. A concurrent GC for D would kill D. Go programs saw a 25-50% performance decrease across the board for the lower latencies. D could make some very minor changes and be capable of a per-thread GC with none of these performance drawbacks, but nobody seems very interested in it. This puts ddmd into context, bearing in mind an automated translation without won't I guess be much slower in LDC or GDC, and it's already a small difference: Release notes on go 1.5 via stack overflow. Builds in Go 1.5 will be slower by a factor of about two. The automatic translation of the compiler and linker from C to Go resulted in unidiomatic Go code that performs poorly compared to well-written Go. Analysis tools and refactoring helped to improve the code, but much remains to be done. Further profiling and optimization will continue in Go 1.6 and future releases. For more details, see these slides and associated video.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 10:47:55 UTC, Laeeth Isharc wrote: Out of curiosity, how much funding is required to develop the more straightforward kind of GCs ? A classical GC like D has is very straightforward. It is been used since the 60s, I even have a paper from 1974 or so describing the implementation used for Simula which is a precise stop-the world GC. Trivial to do. Or to take what's been done and make it possible for others to use? Therein is the trouble, a more advanced GC is intrinsically linked to the language semantics and has to be tuned to the hardware. Expect at least 2 years of work for anything approaching state-of-the-art. In the web server space you wait a lot for I/O so raw performance is not key for Go's success. Stability, memory usage and low latency is more important.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 07:02:40 UTC, Russel Winder wrote: I think Go 2 is a long way off, and even then generics will not be part of the plan. I agree that Go from Google will stay close to the ideals of the creators. I think it would be difficult get beyond that for social reasons. But I think the mechanics Go provides are generic enough that someone could build a transpiler providing more high level convenience. I am thinking along the lines of a convenient language that can compile to both Go and Javascript... I'm tempted to have a go at it. ;) Go UK 2015 was held yesterday. It was less a conference and more a Google rah rah event. It was though very clear that Google are looking for new idioms and practices to come from users other than Google, rather than what has happened to date, which is the Go central team dictating everything to everyone else. Go UK sounds interesting. I wonder if they will have one in Oslo? Probably not :-/.
Re: post on using go 1.5 and GC latency
On Fri, 2015-08-21 at 10:47 +, via Digitalmars-d-learn wrote: Yes, Go has sacrificed some compute performance in favour of latency and convenience. They have also released GC improvement plans for 1.6: https://docs.google.com/document/d/1kBx98ulj5V5M9Zdeamy7v6ofZXX3yPziA f0V27A64Mo/edit It is rather obvious that a building a good concurrent GC is a time consuming effort. But one that Google are entirely happy to fully fund. -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 06:48:48 UTC, Russel Winder wrote: But one that Google are entirely happy to fully fund. Yes, they have made Go fully supported on Google Cloud now, so I think it is safe to say that Google management is backing Go fully. I'm kinda hoping for Go++...
Re: post on using go 1.5 and GC latency
On Sat, 2015-08-22 at 06:54 +, via Digitalmars-d-learn wrote: On Saturday, 22 August 2015 at 06:48:48 UTC, Russel Winder wrote: But one that Google are entirely happy to fully fund. Yes, they have made Go fully supported on Google Cloud now, so I think it is safe to say that Google management is backing Go fully. I'm kinda hoping for Go++... I think Go 2 is a long way off, and even then generics will not be part of the plan. Go UK 2015 was held yesterday. It was less a conference and more a Google rah rah event. It was though very clear that Google are looking for new idioms and practices to come from users other than Google, rather than what has happened to date, which is the Go central team dictating everything to everyone else. -- Russel. = Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Roadm: +44 7770 465 077 xmpp: rus...@winder.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder signature.asc Description: This is a digitally signed message part
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 06:48:48 UTC, Russel Winder wrote: On Fri, 2015-08-21 at 10:47 +, via Digitalmars-d-learn wrote: Yes, Go has sacrificed some compute performance in favour of latency and convenience. They have also released GC improvement plans for 1.6: https://docs.google.com/document/d/1kBx98ulj5V5M9Zdeamy7v6ofZXX3yPziA f0V27A64Mo/edit It is rather obvious that a building a good concurrent GC is a time consuming effort. But one that Google are entirely happy to fully fund. because Go is not a general purpose language. A concurrent GC for D would kill D. Go programs saw a 25-50% performance decrease across the board for the lower latencies. D could make some very minor changes and be capable of a per-thread GC with none of these performance drawbacks, but nobody seems very interested in it.
Re: post on using go 1.5 and GC latency
On Saturday, 22 August 2015 at 10:47:55 UTC, Laeeth Isharc wrote: On Saturday, 22 August 2015 at 09:16:32 UTC, Russel Winder wrote: [...] I didn't mean to start again the whole GC and Go vs D thing. Just that one ought to know the lay of the land as it develops. Out of curiosity, how much funding is required to develop the more straightforward kind of GCs ? Or to take what's been done and make it possible for others to use? It needn't be a single organisation I would think if there are many that would benefit and one doesn't get bogged down in a mentality of people worrying about possibly spurious free rider problems. Since the D Foundation seems under way, it seems worth asking the question first and thinking about goals without worrying for now about what seems realistic. The problem with D's GC is that there's no scaffolding there for it, so you can't really improve it. At best you could make the collector parallel. If I had the runtime hooks and language guarantees I needed I'd begin work on a per-thread GC immediately.
Re: post on using go 1.5 and GC latency
Yes, Go has sacrificed some compute performance in favour of latency and convenience. They have also released GC improvement plans for 1.6: https://docs.google.com/document/d/1kBx98ulj5V5M9Zdeamy7v6ofZXX3yPziAf0V27A64Mo/edit It is rather obvious that a building a good concurrent GC is a time consuming effort.
post on using go 1.5 and GC latency
https://medium.com/@robin.verlangen/billions-of-request-per-day-meet-go-1-5-362bfefa0911 We then started analyzing the behavior of our Go application. On average the application spent ~ 2ms per request, which was great! It gave us 98 milliseconds to spare for network overhead, SSL handshake, DNS lookups and everything else that makes the internet work. Unfortunately the standard deviation of the latency was high, about 100 milliseconds. Meeting our SLA became a major gamble. With the “runtime” package of Go we started profiling the entire application and found out that garbage collection was the cause, resulting in 95-percentile latencies of 279 milliseconds… We decided to rewrite big chunks of the application to generate minimal or no garbage at all. This effectively helped reduce the interval at which garbage collection froze the rest of application to do its cleanup magic. But we were still having issues, so we decided to add more nodes to stay within our SLA. With over 80K requests per second at peak times, even minimal garbage can become a serious issue. ... Yesterday evening (19 August), the moment had finally arrived. A stable version 1.5 of Go was released, claiming: The “stop the world” phase of the collector will almost always be under 10 milliseconds and usually much less. Just a few hours after the release we rebuilt the application with the new version of Go 1.5 and ran our unit and functional tests; they all passed. It seemed too good to be true, so we put some effort in manually verifying the functionality. After a few hours we decided it was safe to release it to a single production node.
