Automatic CPU-GPU affinity

[eile]

On multi-socket systems, performance may vary widely depending on which core a thread executes:

readback "channel" RGBA/401/101/0 1920x1200: 249MPix/sec (8.80081ms, 113FPS)
readback "channel" RGBA/401/101/1 1920x1200: 243MPix/sec (9.02805ms, 110FPS)
readback "channel" RGBA/401/101/2 1920x1200: 257MPix/sec (8.54776ms, 116FPS)
readback "channel" RGBA/401/101/3 1920x1200: 241MPix/sec (9.10284ms, 109FPS)
readback "channel" RGBA/401/101/4 1920x1200: 263MPix/sec (8.34019ms, 119FPS)
readback "channel" RGBA/401/101/5 1920x1200: 262MPix/sec (8.37638ms, 119FPS)
readback "channel" RGBA/401/101/6 1920x1200: 173MPix/sec (12.6451ms, 79FPS)
readback "channel" RGBA/401/101/7 1920x1200: 173MPix/sec (12.6364ms, 79FPS)
readback "channel" RGBA/401/101/8 1920x1200: 173MPix/sec (12.6432ms, 79FPS)
readback "channel" RGBA/401/101/9 1920x1200: 173MPix/sec (12.6438ms, 79FPS)
readback "channel" RGBA/401/101/a 1920x1200: 175MPix/sec (12.5237ms, 79FPS)
readback "channel" RGBA/401/101/b 1920x1200: 174MPix/sec (12.5634ms, 79FPS)

Provide a, preferably automatic, way to configure CPU affinity. hwloc seems to be the most promising package for this.

[eile]

It also matters for network IO:

[[eilemann@node01 Equalizer]$ numactl --cpunodebind=0 ./release/bin/netperf -s node01i:4242:RDMA
Recv perf: 1758.73MB/s (1758.73pps) from RDMA#5000000#node01i##4242#default#
Recv perf: 1787.89MB/s (1787.89pps) from RDMA#5000000#node01i##4242#default#
Recv perf: 1771.19MB/s (1771.19pps) from RDMA#5000000#node01i##4242#default#

[eilemann@node01 Equalizer]$ numactl --cpunodebind=1 ./release/bin/netperf -s node01i:4242:RDMA
Recv perf: 2028.07MB/s (2028.07pps) from RDMA#5000000#node01i##4242#default#
Recv perf: 2022.14MB/s (2022.14pps) from RDMA#5000000#node01i##4242#default#
Recv perf: 1921.8MB/s (1921.8pps) from RDMA#5000000#node01i##4242#default#

[eile]

See also https://svn.open-mpi.org/trac/hwloc/ticket/54

[eile]

Design doc in https://github.com/Eyescale/Equalizer/wiki/Affinity

[eile]

FindHWLOC needs version checking. The Ubuntu version is too old, it doesn't have hwloc_bitmap_t which is deprecated in newer versions. Later we'll need also your new code. Please do this in https://github.com/Eyescale/CMake and then merge to Eq (see doc).

[marwan-abdellah]

This issue was resolved in the commit 08f4ae5

[eile]

Implemented except node thread affinity, as the small gain does not warrant the implementation overhead right now.