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

Wed Jul 8 11:20:54 PDT 2020

Yes sorry, the callstack was only showing “free”. However the time is equally spent between calls to HeapFree and HeapAlloc, so I don’t think this is a pathological case. We can clearly see in the profile traces, threads stalling on the heap’s critical section and then woken up later when the critical section is released into another thread.

As for the adaptative behavior for “affinity slots”, I got word that it doesn’t scale above 4 threads. From what I hear, the behavior of the segment heap is similar to the older low-fragmentation heap, in terms of multi-threaded performance/contention. Although I’d like to hear other opinions if anyone has deeper/practical knowledge with Windows’ segment heap.

De : James Y Knight <jyknight at google.com>
Envoyé : July 6, 2020 3:19 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

https://www.blackhat.com/docs/us-16/materials/us-16-Yason-Windows-10-Segment-Heap-Internals-wp.pdf seems to be the paper that goes with the sides I linked before. It says that there's some sort of adaptive mechanism that allocates per-CPU "affinity slot" if it detects lots of lock contention. Which seems like it ought to have good multithreaded behavior.

I see in your initial email that the sample backtrace is in "free", not allocate. Is that just an example, or is "free" where effectively all the contention is? If the latter, I wonder if we're hitting some pathological edge-case...e.g. allocating on one thread, and then freeing on different threads, or something along those lines.

On Thu, Jul 2, 2020, 11:56 PM Alexandre Ganea <alexandre.ganea at ubisoft.com<mailto:alexandre.ganea at ubisoft.com>> wrote:
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<mailto:jyknight at google.com>>
Envoyé : July 2, 2020 6:08 PM
À : Alexandre Ganea <alexandre.ganea at ubisoft.com<mailto:alexandre.ganea at ubisoft.com>>
Cc : Clang Dev <cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>>; LLVM Dev <llvm-dev at lists.llvm.org<mailto: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/20200708/83f54711/attachment.html>