llvm dev - Jul 2020 - [cfe-dev] RFC: Replacing the default CRT allocator on Windows

Hello,

I was wondering how folks were feeling about replacing the default Windows CRT
allocator in Clang, LLD and other LLVM tools possibly.

The CRT heap allocator on Windows doesn't scale well on large core count
machines. Any multi-threaded workload in LLVM that allocates often is impacted
by this. As a result, link times with ThinLTO are extremely slow on Windows.
We're observing performance inversely proportional to the number of cores.
The more cores the machines has, the slower ThinLTO linking gets.

We've replaced the CRT heap allocator by modern lock-free thread-cache
allocators such as rpmalloc (unlicence), mimalloc (MIT licence) or snmalloc (MIT
licence). The runtime performance is an order of magnitude faster.

Time to link clang.exe with LLD and -flto on 36-core:
  Windows CRT heap allocator: 38 min 47 sec
  mimalloc: 2 min 22 sec
  rpmalloc: 2 min 15 sec
  snmalloc: 2 min 19 sec

We're running in production with a downstream fork of LLVM + rpmalloc for
more than a year. However when cross-compiling some specific game platforms
we're using other downstream forks of LLVM that we can't change.

Two questions arise:

  1.  The licencing. Should we embed one of these allocators into the LLVM tree,
or keep them separate out-of-the-tree?
  2.  If the answer for above question is "yes", given the tremendous
performance speedup, should we embed one of these allocators into Clang/LLD
builds by default? (on Windows only) Considering that Windows doesn't have a
LD_PRELOAD mechanism.

Please see demo patch here: https://reviews.llvm.org/D71786

Thank you in advance for the feedback!
Alex.

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200702/40adf06e/attachment.html>

I'd appreciate the speed-up due to the inclusion of an alternative
malloc and the ease of using it if its source was included in the LLVM
repository. We already have multiple sources from external projects
included in the repository (among them gtest, gmock, ConvertUTF,
google benchmark, ISL, ...) and don't see a reason why it should be
different for a malloc implementation.

AFAIK replacing malloc is quite common for Windows projects. The
Windows default implementation (called "low fragmentation heap") has
different optimization goals.

I'd start with including it in the repository and providing the option
to enable it, with the possibility to change it to the default after
some experience has been collected, maybe even with multiple malloc
implementations.

Michael

Not against this for the executables, but I just wanted to 100% check that
it is possible to override malloc/free for clang and not for the
libclang.dll?

I think it's fine to make the built executables use a different allocator,
but it'd be a bigger pain if we force an allocator on users that link
against the LLVM libraries or shared libraries by default.

Cheers,
-Neil.

On Thu, Jul 2, 2020 at 5:54 AM Michael Kruse via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> I'd appreciate the speed-up due to the inclusion of an alternative
> malloc and the ease of using it if its source was included in the LLVM
> repository. We already have multiple sources from external projects
> included in the repository (among them gtest, gmock, ConvertUTF,
> google benchmark, ISL, ...) and don't see a reason why it should be
> different for a malloc implementation.
>
> AFAIK replacing malloc is quite common for Windows projects. The
> Windows default implementation (called "low fragmentation heap")
has
> different optimization goals.
>
> I'd start with including it in the repository and providing the option
> to enable it, with the possibility to change it to the default after
> some experience has been collected, maybe even with multiple malloc
> implementations.
>
> Michael
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>

-- 
Neil Henning
Senior Software Engineer Compiler
unity.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200702/99a160f9/attachment.html>

On 02/07/2020 05:20, Alexandre Ganea via llvm-dev wrote:
> Time to link clang.exe with LLD and -flto on 36-core:
> 
>    Windows CRT heap allocator: 38 min 47 sec
> 
>    mimalloc: 2 min 22 sec
> 
>    rpmalloc: 2 min 15 sec
> 
>    snmalloc: 2 min 19 sec

These numbers all seem very close (apart from the baseline).  How many 
runs did you do and what was the jitter?

FWIW, I'm using snmalloc on FreeBSD instead of jemalloc and clang is 
around 2% faster, so it might be worth considering this as an option for 
all platforms.  It's likely to be a big win on anything where dlmalloc 
is the default allocator.

Snmalloc currently supports macOS, Windows, Linux, FreeBSD, NetBSD, 
OpenBSD, Haiku, and OpenEnclave (adding other POSIXy systems is fairly 
trivial, can be completely trivial if you don't want any 
non-standard-POSIX behaviour) on x86, ARM, and PowerPC (RISC-V and MIPS 
under review).

I am obviously biased towards snmalloc, since I'm one of the authors, 
and happy to help out anyone wanting to integrate it with LLVM.  Note 
that snmalloc requires C++17, so would need to be conditional on LLVM 
being built with a vaguely modern compiler.

David

Hello David,

Please see below.

-----Message d'origine-----
De : llvm-dev <llvm-dev-bounces at lists.llvm.org> De la part de David
Chisnall via llvm-dev
Envoyé : July 2, 2020 8:04 AM
À : llvm-dev at lists.llvm.org
Objet : Re: [llvm-dev] RFC: Replacing the default CRT allocator on Windows
> These numbers all seem very close (apart from the baseline).  How many runs
did you do and what was the jitter?
Three runs, and I took the last value. Once the Windows cache is hot, the
numbers are very stable. The ThinLTO cache is not enabled, and I used
/opt:lldltojobs=all to extend the ThreadPool to all hardware threads.

> FWIW, I'm using snmalloc on FreeBSD instead of jemalloc and clang is
around 2% faster, so it might be worth considering this as an option for all
platforms.  It's likely to be a big win on anything where dlmalloc is the
default allocator.
I didn't mention, but the compile-time experience has was improved, in the
range of 5-10% depending on the file to compile. When using integrated
compilation, ie. all TUs compile in the same process, the gain is in the range
of 60%. But in that case there are other effects that come into play.

> I am obviously biased towards snmalloc, since I'm one of the authors,
and happy to help out anyone wanting to integrate it with LLVM.  Note that
snmalloc requires C++17, so would need to be conditional on LLVM being built
with a vaguely modern compiler.
snmalloc currently compiles as part of the LLVM codebase with a few
C++17-related constexpr warnings. However the contentious issue is the commit
size, which //could be// a showstopper for certain users. A runtime flag
-fno-integrated-crt-alloc could somehow mitigate this issue. However this only
exacerbates with high core count machines.

Peak commit when linking clang.exe with LLD and -flto on 36-core:

  Windows CRT heap allocator: 14.9 GB
  mimalloc: 19.8 GB
  rpmalloc: 31.9 GB
  snmalloc: 42 GB


_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Have you tried Microsoft's new "segment heap" implementation? Only
apps
that opt-in get it at the moment. Reportedly edge and chromium are getting
large memory savings from switching, but I haven't seen performance
comparisons.

If the performance is good, seems like that might be the simplest choice

https://docs.microsoft.com/en-us/windows/win32/sbscs/application-manifests#heaptype

https://www.blackhat.com/docs/us-16/materials/us-16-Yason-Windows-10-Segment-Heap-Internals.pdf

On Thu, Jul 2, 2020, 12:20 AM Alexandre Ganea via cfe-dev <
cfe-dev at lists.llvm.org> wrote:
> Hello,
>
>
>
> I was wondering how folks were feeling about replacing the default Windows
> CRT allocator in Clang, LLD and other LLVM tools possibly.
>
>
>
> The CRT heap allocator on Windows doesn’t scale well on large core count
> machines. Any multi-threaded workload in LLVM that allocates often is
> impacted by this. As a result, link times with ThinLTO are extremely slow
> on Windows. We’re observing performance inversely proportional to the
> number of cores. The more cores the machines has, the slower ThinLTO
> linking gets.
>
>
>
> We’ve replaced the CRT heap allocator by modern lock-free thread-cache
> allocators such as rpmalloc (unlicence), mimalloc (MIT licence) or snmalloc
> (MIT licence). The runtime performance is an order of magnitude faster.
>
>
>
> Time to link clang.exe with LLD and -flto on 36-core:
>
>   Windows CRT heap allocator: 38 min 47 sec
>
>   mimalloc: 2 min 22 sec
>
>   rpmalloc: 2 min 15 sec
>
>   snmalloc: 2 min 19 sec
>
>
>
> We’re running in production with a downstream fork of LLVM + rpmalloc for
> more than a year. However when cross-compiling some specific game platforms
> we’re using other downstream forks of LLVM that we can’t change.
>
>
>
> Two questions arise:
>
>    1. The licencing. Should we embed one of these allocators into the
>    LLVM tree, or keep them separate out-of-the-tree?
>    2. If the answer for above question is “yes”, given the tremendous
>    performance speedup, should we embed one of these allocators into
Clang/LLD
>    builds by default? (on Windows only) Considering that Windows doesn’t
have
>    a LD_PRELOAD mechanism.
>
>
>
> Please see demo patch here: https://reviews.llvm.org/D71786
>
>
>
> Thank you in advance for the feedback!
>
> Alex.
>
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200702/7c79c7d9/attachment.html>

Thanks for the suggestion James, it reduces the commit by about ~900 MB (14,9 GB
-> 14 GB).

Unfortunately it does not solve the performance problem. The heap is global to
the application and thread-safe, so every malloc/free locks it, which evidently
doesn’t scale. We could manually create thread-local heaps, but I didn’t want to
go there. Ultimately allocated blocks need to share ownership between threads,
and at that point it’s like re-writing a new allocator. I suppose most
non-Windows platforms already have lock-free thread-local arenas, which probably
explains why this issue has gone (mostly) unnoticed.


De : James Y Knight <jyknight at google.com>
Envoyé : July 2, 2020 6:08 PM
À : Alexandre Ganea <alexandre.ganea at ubisoft.com>
Cc : Clang Dev <cfe-dev at lists.llvm.org>; LLVM Dev <llvm-dev at
lists.llvm.org>
Objet : Re: [cfe-dev] RFC: Replacing the default CRT allocator on Windows

Have you tried Microsoft's new "segment heap" implementation? Only
apps that opt-in get it at the moment. Reportedly edge and chromium are getting
large memory savings from switching, but I haven't seen performance
comparisons.

If the performance is good, seems like that might be the simplest choice

https://docs.microsoft.com/en-us/windows/win32/sbscs/application-manifests#heaptype

https://www.blackhat.com/docs/us-16/materials/us-16-Yason-Windows-10-Segment-Heap-Internals.pdf

On Thu, Jul 2, 2020, 12:20 AM Alexandre Ganea via cfe-dev <cfe-dev at
lists.llvm.org<mailto:cfe-dev at lists.llvm.org>> wrote:
Hello,

I was wondering how folks were feeling about replacing the default Windows CRT
allocator in Clang, LLD and other LLVM tools possibly.

The CRT heap allocator on Windows doesn’t scale well on large core count
machines. Any multi-threaded workload in LLVM that allocates often is impacted
by this. As a result, link times with ThinLTO are extremely slow on Windows.
We’re observing performance inversely proportional to the number of cores. The
more cores the machines has, the slower ThinLTO linking gets.

We’ve replaced the CRT heap allocator by modern lock-free thread-cache
allocators such as rpmalloc (unlicence), mimalloc (MIT licence) or snmalloc (MIT
licence). The runtime performance is an order of magnitude faster.

Time to link clang.exe with LLD and -flto on 36-core:
  Windows CRT heap allocator: 38 min 47 sec
  mimalloc: 2 min 22 sec
  rpmalloc: 2 min 15 sec
  snmalloc: 2 min 19 sec

We’re running in production with a downstream fork of LLVM + rpmalloc for more
than a year. However when cross-compiling some specific game platforms we’re
using other downstream forks of LLVM that we can’t change.

Two questions arise:

  1.  The licencing. Should we embed one of these allocators into the LLVM tree,
or keep them separate out-of-the-tree?
  2.  If the answer for above question is “yes”, given the tremendous
performance speedup, should we embed one of these allocators into Clang/LLD
builds by default? (on Windows only) Considering that Windows doesn’t have a
LD_PRELOAD mechanism.

Please see demo patch here: https://reviews.llvm.org/D71786

Thank you in advance for the feedback!
Alex.

_______________________________________________
cfe-dev mailing list
cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200703/cbe57497/attachment.html>