[x265] [PATCH] Performance: Prevent small thread-pools if NUMA disabled and # CPUs > MAX_POOL_THREADS

[Steve Borho]

On 08/25, pradeep at multicorewareinc.com wrote:
> # HG changeset patch
> # User pradeep
> # Date 1440476198 -19800
> #      Tue Aug 25 09:46:38 2015 +0530
> # Node ID ba08fde2e66bb66f0410fd5bb7a28f62df0043e0
> # Parent  a28a863393994d8fb1d58c721352d9b4ec8c46ee
> Performance: Prevent small thread-pools if NUMA disabled and # CPUs > MAX_POOL_THREADS
> 
> When NUMA is disabled and if # CPUs is > MAX_POOL_THREADS (64 or 32 depending on 64-bit
> or 32-bit builds), the last pool may have few threads. This patch allows the last pool to
> exist only if it has > MAX_POOL_THREADS/2 threads; the 50% number is a heurstic.
> 
> Feature gains performance of 5% on Intel Xeon E5-2699v3 measured in slower preset with 4K video.
> 
> diff -r a28a86339399 -r ba08fde2e66b source/common/threadpool.cpp
> --- a/source/common/threadpool.cpp	Mon Aug 24 14:04:32 2015 +0530
> +++ b/source/common/threadpool.cpp	Tue Aug 25 09:46:38 2015 +0530
> @@ -289,6 +289,11 @@
>          }
>      }
>  
> +    // In the case that numa is disabled and we have more CPUs than 64,
> +    // spawn the last pool only if the # threads in that pool is > 1/2 max (heuristic)
> +    if ((numNumaNodes == 1) && (cpusPerNode[0] % MAX_POOL_THREADS < (MAX_POOL_THREADS / 2)))
> +        cpusPerNode[0] = ((int)(cpusPerNode[0] / MAX_POOL_THREADS)) * MAX_POOL_THREADS ;

the math looks fishy, how about? cpusPerNode[0] -= cpusPerNode[0] % MAX_POOL_THREADS;

it probably deserves an x265_log debug message saying the pool size is
capped to prevent asymmetry

watch the space before semicolon

>      numPools = 0;
>      for (int i = 0; i < numNumaNodes; i++)
>      {
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel

-- 
Steve Borho