-
Notifications
You must be signed in to change notification settings - Fork 423
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
TOOLS/PERF: Changes to enable the rocm perf modules #4349
Conversation
Can one of the admins verify this patch? |
ok to test |
Test FAILed. |
Test FAILed. |
Mellanox CI: FAILED on 4 of 25 workers (click for details)Note: the logs will be deleted after 05-Nov-2019
|
Mellanox CI: FAILED on 5 of 25 workers (click for details)Note: the logs will be deleted after 18-Nov-2019
|
Mellanox CI: FAILED on 5 of 25 workers (click for details)Note: the logs will be deleted after 18-Nov-2019
|
# This is the 1st commit message: Changes to enable the rocm perf modules # This is the commit message #2: UCT: rdmacm on iface: null check for the UNREACHABLE event's priv_data # This is the commit message #3: UCT: rdmacm on iface - remove redundant static assert. # This is the commit message #4: GTEST/UCP: Fix clang 9 warnings for OpenMP loop variable initialization # This is the commit message #5: UCT/IB/MLX5: Modify QP via DEVX # This is the commit message #6: MEM/TYPE-DETECT: use const pointer - use const void* buffers for memory type detection # This is the commit message #7: UCT.H: fixed typo in description # This is the commit message #8: UCM/UTIL: Add error prints to mmap() failures # This is the commit message #9: TOOLS/PROFILE: Support custom pager # This is the commit message #10: UCT/RCX: MP XRQ support # This is the commit message #11: TOOLS/TESTS/BUILD-PGI: fixed compilation flags - fixed compilation flags detection for PGI compiler - optimized ICC flags settings # This is the commit message #12: UCP/CORE, GTEST/UCP: Add tests for zero-thresholds + fixes for RNDV thresholds # This is the commit message #13: GTEST/UCP: Remove duplicate test # This is the commit message #14: GTEST/UCP: Initialize m_env inside ctor # This is the commit message #15: UCS/MEMTYPE_CACHE, GTEST/UCS: Align regions by Page Table alignment # This is the commit message #16: TOOLS/PERF: Print an error if memory type is unsupported # This is the commit message #17: UCT/IB: Add multi-threaded MR handling performance on Intel E5-2697A 32 cores without patch 100G shmem_init: 25.32 secs shmem_finalize: 3.26 secs 200G shmem_init: 106.29 secs shmem_finalize 9.55 secs with patch 100G shmem_init: 1.63 secs shmem_finalize: 0.44 secs 200G shmem_init: 3.01 secs shmem_finalize 0.85 secs # This is the commit message #18: UCS/ASYNC: Support sync-removing a handler from its callback Allow removing the handler from its callback in a synchronous way. This means the function will return when the only remaining reference to the handler is of the current callback. # This is the commit message #19: EXAMPLES/JAVA: Sleep for 3 seconds after endpoint destroy Give grace period for other side to disconnect before worker is destroyed. Need to remove this when UCP close protocol is implemented. Fixes test failures with "Retry count exceeded". # This is the commit message #20: UCS/ASYNC: Code review fixes and reword comment # This is the commit message #21: UCS/STRING: Add ucs_snprintf_safe() # This is the commit message #22: UCT/TCP, UCS/SYS: Repeat connection establishment whn detected dropped conenctions # This is the commit message #23: UCT/TCP: Fix review comments # This is the commit message #24: TEST/GTEST: Print parameter name for every running test # This is the commit message #25: TEST/GTETS: reduce execution time of stress sockaddr testing under valgrind # This is the commit message #26: UCP/WIREUP: remove ucp_ep_params from ucp_wireup_ep_t And store all required info in ucp_ep_init_flags # This is the commit message #27: TEST/GTEST: Small enhancements # This is the commit message #28: TEST/UCT: Extend MM tests # This is the commit message #29: UCT/MM: Fix md_attr::cap.max_reg for xpmem # This is the commit message #30: GTEST/UCS/UCT: Enable sockaddr testing with an IPv6 address. Don't use functions that support only IPv4. ioctl is legacy and won't return IPv6. https://stackoverflow.com/questions/20743709/get-ipv6-addresses-in-linux-using-ioctl # This is the commit message #31: GTEST: sockaddr: fix testing for ipv6 # This is the commit message #32: UCS/ASYNC/TEST: Fix sync handler remove In order to support removing async handler while its callback is called, we must make sure the call happens from the same thread. Otherwise, the sync-remove operation does not really guarantee sync, and subsequent file operations (such as accept() in sockcm) will fail because fd is already closed. Refactor polling on missed events to avoid deadlock - we can assume the polled handlers come from same async context we are polling on (because of the way the are added to miss queue). Add a test to make sure the async handler is not called after it's sync-removed. # This is the commit message #33: UCS/DEBUG: Do not use undefined signal names, and terminate the array. On FreeBSD SIGSYS is 12, and SIGPIPE is 13. On the other hand, there is no SIGSTKFLT and SIGPWR. Signed-off-by: Konstantin Belousov <konstantinb@mellanox.com> # This is the commit message #34: UCP/WIREUP/TEST: Fix asymmetric endpoints connection with p2p transport # This is the commit message #35: UCP/ADDRESS: Code review fixes # This is the commit message #36: TEST/MPI: Remove accidentaly added file # This is the commit message #37: UCS/BACKTRACE: re-implemented backtrace output - added backtrace output to invalid AM ID handler - optimization for ucs_mmap_free: now ucs_mmap_free gets original object size used to allocate # This is the commit message #38: UCM/CUDA: Skip ucm_cuda_set_ptr_attr when pointer is null # This is the commit message #39: AZP: edit existing release if exist # This is the commit message #40: TEST/ASYNC: Add retries to test_async.ctx_event_block # This is the commit message #41: GTEST/UCT/ROCM: Enable ROCm unit tests in the gtest framework # This is the commit message #42: UCM/CUDA: fix ucm_cudafree_dispatch_events assertion Only call assert in ucm_cudafree_dispatch_events if cuMemGetAddressRange() call succeeds and downgrade warning to debug message when it fails. # This is the commit message #43: ARCH/ARM: Barrier code update. The change is driven by recent clarifications in linux kernel code: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=22ec71615d824f4f11d38d0e55a88d8956b7e45f Performance testing shows 8% uplift for short messages using InfiniBand. Signed-off-by: Pavel Shamis (Pasha) <pasharesearch@gmail.com> # This is the commit message #44: UCT/CUDA_IPC: Fix peer_accessible check # This is the commit message #45: RC/IFACE: fixed potential leak on failed rc iface init - added jump label for more accurate error handling
…me system and explicitly define the include files, and changes to enable the rocm perf modules
Mellanox CI: FAILED on 5 of 25 workers (click for details)Note: the logs will be deleted after 19-Nov-2019
|
Mellanox CI: FAILED on 5 of 25 workers (click for details)Note: the logs will be deleted after 19-Nov-2019
|
Mellanox CI: FAILED on 5 of 25 workers (click for details)Note: the logs will be deleted after 19-Nov-2019
|
closing this PR to fix the commit title |
What
I am including the changes to enable the perf modules for rocm to get built, so we can run ucx_perftest on rocm device.
Why ?
The rocm support on ucx_perftest is not enabled, so just to enable the ucx_perftest. This change does not change the existing functionality of rocm in ucx.