[openblas] Build for high numbers of threads (4096) by default instead of 32 #3667

ViralBShah · 2021-09-30T22:42:35Z

Core counts are going up a lot nowadays - I suggest we use 128 max threads by default.

Note that this is only updating the common.jl file, which will not trigger the building of the libraries. That will get picked up in the next run. I wanted to open this PR to have discussion on the defaults and prepare for bumping things up whenever we are ready.

Core counts are going up a lot nowadays - I suggest we use 128 max threads by default.

vchuravy · 2021-10-01T02:11:32Z

What is the consequence on startup time?

ViralBShah · 2021-10-01T02:19:42Z

Presumably also will use more memory. what is a reliable way to test both? Startup time probably with

time ./julia -e exit

Is there a better way to look at real memory usage than just eyeballing htop?

inkydragon · 2021-10-01T04:21:00Z

Maybe /usr/bin/time -v julia -e "a=rand(5_000_000); BigFloat.(a)"

Output

    Command being timed: "julia -e a=rand(5_000_000); BigFloat.(a)"
    User time (seconds): 0.99
    System time (seconds): 0.56
    Percent of CPU this job got: 142%
    Elapsed (wall clock) time (h:mm:ss or m:ss): 0:01.09
    Average shared text size (kbytes): 0
    Average unshared data size (kbytes): 0
    Average stack size (kbytes): 0
    Average total size (kbytes): 0
    Maximum resident set size (kbytes): 828132
    Average resident set size (kbytes): 0
    Major (requiring I/O) page faults: 0
    Minor (reclaiming a frame) page faults: 32552
    Voluntary context switches: 7
    Involuntary context switches: 16
    Swaps: 0
    File system inputs: 0
    File system outputs: 0
    Socket messages sent: 0
    Socket messages received: 0
    Signals delivered: 0
    Page size (bytes): 4096
    Exit status: 0

or

# sudo apt install valgrind -y
valgrind --trace-children=yes --tool=massif  julia -e "a=rand(5_000_000); BigFloat.(a)"
ms_print massif.out.* | less

Output

--------------------------------------------------------------------------------
Command:            julia -e a=rand(5_000_000); BigFloat.(a)
Massif arguments:   (none)
ms_print arguments: massif.out.2674
--------------------------------------------------------------------------------


    MB
175.7^                                                                       @
     |        @:::::::@::::::::::::@:::::::::@:::::::::::##:::::::::::::@::::@
     |        @:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |        @:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |        @:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |     @::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |     @::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |     @::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |     @::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |     @::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
     |   @:@::@:: : ::@:  ::::::   @:::::: : @:::        # :: :: :::::::@::::@
   0 +----------------------------------------------------------------------->Gi
     0                                                                   6.243

Number of snapshots: 64
 Detailed snapshots: [3, 5, 8, 16, 24, 33, 37 (peak), 53, 63]

--------------------------------------------------------------------------------
  n        time(i)         total(B)   useful-heap(B) extra-heap(B)    stacks(B)
--------------------------------------------------------------------------------
  0              0                0                0             0            0
  1    113,336,545        3,633,576        3,607,086        26,490            0
  2    219,235,835        3,649,968        3,623,470        26,498            0
  3    324,720,416       97,585,728       97,427,879       157,849            0
99.84% (97,427,879B) (heap allocation functions) malloc/new/new[], --alloc-fns, etc.
->81.98% (80,000,000B) 0x55F76B7: jl_profile_init (signal-handling.c:294)
| ->81.98% (80,000,000B) 0xC0ADEFA: julia_init_55051.clone_1 (Profile.jl:60)
|   ->81.98% (80,000,000B) 0xC0ADCE3: julia___init___55048.clone_1 (in /home/woclass/packages/julias/julia-1.6/lib/julia/sys.so)
|     ->81.98% (80,000,000B) 0xC0ADCF5: jfptr___init___55049.clone_1 (in /home/woclass/packages/julias/julia-1.6/lib/julia/sys.so)
|       ->81.98% (80,000,000B) 0x5599459: _jl_invoke (gf.c:2237)
|         ->81.98% (80,000,000B) 0x5599459: jl_apply_generic (gf.c:2419)
|           ->81.98% (80,000,000B) 0x55D21C5: jl_apply (julia.h:1703)
|             ->81.98% (80,000,000B) 0x55D21C5: jl_module_run_initializer (toplevel.c:72)
|               ->81.98% (80,000,000B) 0x55B9ADB: _julia_init (init.c:794)
|                 ->81.98% (80,000,000B) 0x55F5A70: repl_entrypoint (jlapi.c:696)
|                   ->81.98% (80,000,000B) 0x4007A8: main (loader_exe.c:51)
|
....

ViralBShah · 2021-10-01T22:52:38Z

I am doing time ./julia -e "using LinearAlgebra; @show BLAS.get_num_threads(); exit" on Julia master (which is built for 32 max openblas threads).

With 8 threads, it is 0.24s
With 32 threads, it is 0.29s

When I replace openblas with the version from OpenBLASHighCoreCounts_jll, I see no meaningful change in startup time or memory usage, but the machine I am testing on only goes to 40 cores.

ViralBShah · 2021-10-01T23:08:17Z

@staticfloat @chriselrod Any thoughts here?

staticfloat · 2021-10-02T00:20:41Z

Losing 50ms is rather significant, IMO. I'd like to better understand why we're losing all that time.

ViralBShah · 2021-10-02T00:28:08Z

I think 50ms is not too bad given the performance boost (only on systems with large numbers of cores). As a user, I want Julia to start with as many threads as cores. I thought we lose time because of the openblas per-thread buffer allocation that happens when you dlopen it. @vtjnash probably knows this quite well.

We could try to estimate the true number of cores and use half as many threads possibly (avoiding the allocation one openblas thread per hyperthread), but I don't know how reliably one can do such a thing.

ViralBShah · 2021-10-02T13:52:23Z

The impact on pure Julia startup time is on the order of 5ms on arctic1 (which is now several years old).

ViralBShah · 2021-10-05T17:25:58Z

I propose merging this, and using 512 for the HighCoreCounts version. Thoughts?

staticfloat · 2021-10-05T21:36:50Z

After talking it over with Viral, I think we can come to a happy compromise:

Investigate whether building OpenBLAS with high max core counts is feasible on all platforms, and if it is, get rid of OpenBLASHighCoreCount and just have a single OpenBLAS_jll that has its max CPU count set to something very high like 4096.
Re-instate a Julia __init__() limit that does something like:

ENV["OPENBLAS_NUM_THREADS"] = get(ENV, "OPENBLAS_NUM_THREADS", string(min(Sys.CPU_THREADS, 32)))

We should probably put that in LinearAlgebra instead of Base.

This way, we get:

Reasonable performance by default (32 cores is enough for most machines).
Reduced latency increase on manycore machines.
The ability to override Julia's default limit and run with 128 cores if you really want to.

I've looked at Sys.maxrss() with a variety of different OPENBLAS_NUM_THREADS settings, and it doesn't appear to change that much (~1MB difference between 1 and 128 threads) so I don't think memory usage is an issue when setting extremely high core counts.

ViralBShah · 2021-10-05T21:55:29Z

We may want to merge JuliaLang/julia#42473 first - but I think these environment variables need to get set before we dlopen openblas. Is there a chance it gets dlopened earlier?

chriselrod · 2021-10-05T22:01:13Z

I think that sounds reasonable.
Eventually we should use physical core counts instead of threads
JuliaLang/LinearAlgebra.jl#671

How did you time the difference in startup times?

time OPENBLAS_NUM_THREADS=1 julia --startup=no -e '0'

seems too noisy for me to notice a difference.

ViralBShah · 2021-10-05T22:11:30Z

We should absolutely do physical core counts if we can get them reliably. Note you may often want to do /usr/bin/time.

staticfloat · 2021-10-05T22:13:28Z

I actually use @btime from BenchmarkTools, e.g.

@btime run(`./julia -e0`)

chriselrod · 2021-10-05T22:27:09Z

I still don't see a marked difference. Minimum time seems slightly worse, while median and mean improve:

> OPENBLAS_NUM_THREADS=1 julia --startup=no -e 'using BenchmarkTools
, LinearAlgebra; @show(BLAS.get_num_threads()); display(@benchmark run(`$(Base.julia_cmd()) --startup=no -e0`))'
BLAS.get_num_threads() = 1
BenchmarkTools.Trial: 42 samples with 1 evaluation.
 Range (min … max):  108.952 ms … 124.790 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     119.121 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   119.068 ms ±   4.407 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

           ▁              ▁         ▁▄ ▁ ▁ ▁         ▁▁  ▁ █
  ▆▁▁▁▁▆▁▆▁█▁▁▁▆▁▁▆▁▁▁▁▁▁▁█▆▁▁▁▁▁▁▁▆██▆█▆█▁█▁▁▆▁▁▆▆▁▁██▆▁█▆█▆▆▆ ▁
  109 ms           Histogram: frequency by time          125 ms <

 Memory estimate: 4.00 KiB, allocs estimate: 87.⏎ 

> OPENBLAS_NUM_THREADS=1 julia  --startup=no -e 'using BenchmarkTools
, LinearAlgebra; @show(BLAS.get_num_threads()); display(@benchmark run(`$(Base.julia_cmd()) --startup=no -e0`))'
BLAS.get_num_threads() = 1
BenchmarkTools.Trial: 43 samples with 1 evaluation.
 Range (min … max):  108.923 ms … 129.074 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     118.254 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   117.794 ms ±   4.956 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

    ▃█   ▃          █▃  ▃   █ █  ▃ ▃       █   █
  ▇▁██▁▁▁█▁▁▁▇▁▁▁▁▇▁██▁▁█▁▇▁█▁█▇▁█▇█▁▇▁▁▇▇▁█▁▇▁█▁▇▇▁▁▁▁▁▁▁▁▁▁▁▇ ▁
  109 ms           Histogram: frequency by time          129 ms <

 Memory estimate: 4.00 KiB, allocs estimate: 87.⏎

> OPENBLAS_NUM_THREADS=28 julia --startup=no -e 'using BenchmarkTool
s, LinearAlgebra; @show(BLAS.get_num_threads()); display(@benchmark run(`$(Base.julia_cmd()) --startup=no -e0`))'
BLAS.get_num_threads() = 28
BenchmarkTools.Trial: 44 samples with 1 evaluation.
 Range (min … max):  111.792 ms … 124.431 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     113.549 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   114.528 ms ±   2.479 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

   ▃   █ ▆▃ ▁ ▃
  ▇█▁▁▄█▄██▄█▁█▄▁▁▁▁▄▁▁▄▁▄▄▁▁▇▇▇▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄ ▁
  112 ms           Histogram: frequency by time          124 ms <

 Memory estimate: 4.00 KiB, allocs estimate: 87.⏎ 

> OPENBLAS_NUM_THREADS=28 julia --startup=no -e 'using BenchmarkTool
s, LinearAlgebra; @show(BLAS.get_num_threads()); display(@benchmark run(`$(Base.julia_cmd()) --startup=no -e0`))'
BLAS.get_num_threads() = 28
BenchmarkTools.Trial: 44 samples with 1 evaluation.
 Range (min … max):  111.519 ms … 120.237 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     113.767 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   114.196 ms ±   2.124 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

            ▅█      ▅
  █▅▁█▅█▅█▅▁██▁▁▁▅█▅█▅▁▁█▅▅▁▁▅▁▁▁▁▅▅▁▁▁▁█▅▁▅▁▁▁▁▁▁▁▅▁▁▁▁▁▁▅▁▁▁▅ ▁
  112 ms           Histogram: frequency by time          120 ms <

 Memory estimate: 4.00 KiB, allocs estimate: 87.⏎

This is on a 14 core / 28 thread CPU.

staticfloat · 2021-10-05T22:33:59Z

This looks very similar to what I was seeing; you're seeing a 3ms increase in minimum time (which, IMO, is the only thing we should be paying attention to, as all other variation is most likely due to other processes adding noise). If you had a 128 core machine, you'd probably see an increase closer to 10ms.

ViralBShah · 2021-10-05T22:50:08Z

So, how do we feel about merging this with 4096 max threads? :-)

staticfloat · 2021-10-05T22:54:27Z

I would like someone to do a quick sanity check on Windows and macOS to see what happens if you build OpenBLAS with 4096 cores. It's not too hard; just build Julia via make USE_BINARYBUILDER_OPENBLAS=0 after having changed these lines to set the limit to 4096. Then compare the memory usage there with the memory usage of a normal Julia, (startup time would be good to know as well)

ViralBShah · 2021-10-05T22:57:03Z

We'll certainly get mac testers, but who might be able to do this for Windows?

staticfloat · 2021-10-06T05:44:55Z

I checked on windows; it's about the same as on Linux.

ViralBShah · 2021-10-06T12:01:42Z

@staticfloat can you check startup time impact when you use as many threads as cores? OpenBLAS by default will launch as many threads as hyperthreads which may be much worse for startup time impact due to oversubscribing.

ViralBShah · 2021-10-06T13:53:52Z

On mac with the current master, there's some makefile breakage:

➜  julia git:(master) ✗ make USE_BINARYBUILDER_OPENBLAS=0 VERBOSE=1
make[1]: *** No rule to make target `scratch/objconv/build-compiled', needed by `scratch/openblas-d909f9f3d4fc4ccff36d69f178558df154ba1002/build-compiled'.  Stop.
make: *** [julia-deps] Error 2

Found a way forward, but also filed:
JuliaLang/julia#42519

staticfloat · 2021-10-07T22:22:49Z

I'm satisfied that this is reasonable now. Let's merge this, and then work on the Julia-side thread choice logic.

ViralBShah · 2021-10-08T01:06:24Z

Note that this only changes common.jl. We need a separate PR to actually build a new OpenBLAS with this, after which we do the version bump on julia master. Should we keep the openblas high core counts around - or get rid of it?

…d of 32 (JuliaPackaging#3667) * [openblas] Build for 4096 threads by default We're building with an extremely high thread limit, then limiting in Julia according to our hardware introspection

[openblas] Build for 128 threads by default

e519043

Core counts are going up a lot nowadays - I suggest we use 128 max threads by default.

ViralBShah changed the title ~~[openblas] Build for 128 threads by default~~ [openblas] Build for 128 threads by default instead of 32 Sep 30, 2021

ViralBShah mentioned this pull request Oct 1, 2021

Remove openblas set_num_threads in julia __init__ JuliaLang/julia#42442

Merged

ViralBShah changed the title ~~[openblas] Build for 128 threads by default instead of 32~~ [openblas] Build for high numbers of threads (4096) by default instead of 32 Oct 5, 2021

Update common.jl

86fcf2a

staticfloat merged commit 433dc26 into master Oct 7, 2021

staticfloat deleted the vs/openblas-threads branch October 7, 2021 22:23

ViralBShah mentioned this pull request Oct 8, 2021

[OpenBLAS/lapack] Apply the patch from xianyi/OpenBLAS#3392 #3708

Merged

vtjnash mentioned this pull request Aug 9, 2022

addprocs fails on Julia v1.8 triggering internal error JuliaLang/julia#45434

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[openblas] Build for high numbers of threads (4096) by default instead of 32 #3667

[openblas] Build for high numbers of threads (4096) by default instead of 32 #3667

ViralBShah commented Sep 30, 2021 •

edited

Loading

vchuravy commented Oct 1, 2021

ViralBShah commented Oct 1, 2021 •

edited

Loading

inkydragon commented Oct 1, 2021 •

edited

Loading

ViralBShah commented Oct 1, 2021 •

edited

Loading

ViralBShah commented Oct 1, 2021 •

edited

Loading

staticfloat commented Oct 2, 2021

ViralBShah commented Oct 2, 2021 •

edited

Loading

ViralBShah commented Oct 2, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

chriselrod commented Oct 5, 2021

ViralBShah commented Oct 5, 2021 •

edited

Loading

staticfloat commented Oct 5, 2021

chriselrod commented Oct 5, 2021 •

edited

Loading

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 6, 2021

ViralBShah commented Oct 6, 2021 •

edited

Loading

ViralBShah commented Oct 6, 2021 •

edited

Loading

staticfloat commented Oct 7, 2021

ViralBShah commented Oct 8, 2021

[openblas] Build for high numbers of threads (4096) by default instead of 32 #3667

[openblas] Build for high numbers of threads (4096) by default instead of 32 #3667

Conversation

ViralBShah commented Sep 30, 2021 • edited Loading

vchuravy commented Oct 1, 2021

ViralBShah commented Oct 1, 2021 • edited Loading

inkydragon commented Oct 1, 2021 • edited Loading

ViralBShah commented Oct 1, 2021 • edited Loading

ViralBShah commented Oct 1, 2021 • edited Loading

staticfloat commented Oct 2, 2021

ViralBShah commented Oct 2, 2021 • edited Loading

ViralBShah commented Oct 2, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

chriselrod commented Oct 5, 2021

ViralBShah commented Oct 5, 2021 • edited Loading

staticfloat commented Oct 5, 2021

chriselrod commented Oct 5, 2021 • edited Loading

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 5, 2021

ViralBShah commented Oct 5, 2021

staticfloat commented Oct 6, 2021

ViralBShah commented Oct 6, 2021 • edited Loading

ViralBShah commented Oct 6, 2021 • edited Loading

staticfloat commented Oct 7, 2021

ViralBShah commented Oct 8, 2021

ViralBShah commented Sep 30, 2021 •

edited

Loading

ViralBShah commented Oct 1, 2021 •

edited

Loading

inkydragon commented Oct 1, 2021 •

edited

Loading

ViralBShah commented Oct 1, 2021 •

edited

Loading

ViralBShah commented Oct 1, 2021 •

edited

Loading

ViralBShah commented Oct 2, 2021 •

edited

Loading

ViralBShah commented Oct 5, 2021 •

edited

Loading

chriselrod commented Oct 5, 2021 •

edited

Loading

ViralBShah commented Oct 6, 2021 •

edited

Loading

ViralBShah commented Oct 6, 2021 •

edited

Loading