nscottnichols
commented
8 months ago Was able to compile and run.
nscottnichols@x1921c1s0b0n0:~/thapi_PR_test/build/ici/bin> mpirun -n 1 ./iprof -- sycl-ls
THAPI: Trace Location /home/nscottnichols/thapi-traces/thapi_aggreg--2024-02-14--21h29m41s
BACKEND_ZE | 1 Hostnames | 1 Processes | 1 Threads |
Name | Time | Time(%) | Calls | Average | Min | Max |
zeDeviceGetSubDevices | 7.70us | 40.51% | 12 | 641.75ns | 137ns | 2.02us |
zeDeviceGet | 7.41us | 38.99% | 2 | 3.71us | 460ns | 6.95us |
zeDriverGetExtensionFunctionAddress | 2.04us | 10.75% | 2 | 1.02us | 302ns | 1.74us |
zeInit | 873ns | 4.59% | 1 | 873.00ns | 873ns | 873ns |
zeDriverGetApiVersion | 592ns | 3.11% | 1 | 592.00ns | 592ns | 592ns |
zeDriverGet | 388ns | 2.04% | 2 | 194.00ns | 156ns | 232ns |
Total | 19.01us | 100.00% | 20 |mpirun -n 16 ./iprof -- sycl-ls
THAPI: Trace Location /home/nscottnichols/thapi-traces/thapi_aggreg--2024-02-14--21h26m35s
BACKEND_ZE | 2 Hostnames | 16 Processes | 16 Threads |
Name | Time | Time(%) | Calls | Average | Min | Max |
zeDeviceGetSubDevices | 1.55ms | 85.81% | 192 | 8.06us | 134ns | 221.75us |
zeDeviceGet | 187.53us | 10.40% | 32 | 5.86us | 321ns | 15.36us |
zeDriverGetExtensionFunctionAddress | 33.87us | 1.88% | 32 | 1.06us | 256ns | 3.09us |
zeInit | 13.88us | 0.77% | 16 | 867.44ns | 507ns | 1.22us |
zeDriverGetApiVersion | 11.45us | 0.63% | 16 | 715.81ns | 462ns | 1.12us |
zeDriverGet | 9.21us | 0.51% | 32 | 287.91ns | 134ns | 576ns |
Total | 1.80ms | 100.00% | 320 |Testing on larger application now.
TApplencourt
Based on an idea on @nscottnichols
This PR add a new binary (
sync_daemon_fs) who is responsible of performing 2 operation:The communication is on
signal, so need to for some loop.The end goal is to implement another daemon that will use MPI (for faster and more robust SYNC) using the same API. Then a compile time, we can choose with daemon to use (the filesystem base one, or the MPI one).