Ilia Yastrebov
Ilia Yastrebov
## What Second batch commit for DPU xgvmi daemon It contains handlers for host-to-peer and peer-to-peer messages no the daemon side. For reference: First batch commit https://github.com/openucx/ucx/pull/9536 is merged. Second...
### Describe the bug "gpu on worker 0" test `dcx/test_ucp_tag_mem_type.reuse_buffers_mrail/15 ` failed with assertion, which seems to be unrelated to my change (https://github.com/openucx/ucx/pull/9525): ``` [ RUN ] dcx/test_ucp_tag_mem_type.reuse_buffers_mrail/15 [ INFO...
## What First CI test for perftest_daemon. ## Why ? XGVMI functionality must be covered by CI tests to prevent regressions. ## How ? Some refactoring was needed to allow...
## What Few unit tests for XGVMI UMR approach to check that: - Lazy initialization works as expected - All UMR invariants are respected - UMR mkey pool is properly...
There is a bug introduced in recent commit https://github.com/openucx/ucx/pull/9650, which is exposed in 2 ways: - assertion failure in debug build in function `ucp_worker_iface_deactivate : wiface->activate_count > 0` - in...
## What Fix for [RM#3886801](https://redmine.mellanox.com/issues/3886801) I managed to reproduce this issue on rock machines in 100% of the cases, but only when running this test under high CPU load. This...
## What `UCP_MEMH_FLAG_IMPORTED` was added in https://github.com/openucx/ucx/pull/8584, as the only memory handle flag. But this flag is only set in the memh parent, and never propagated to user memh. Despite...
## What This is fix for [RM#3919044](https://redmine.mellanox.com/issues/3919044) There are multiple test failures reported as: ``` ==54391== Conditional jump or move depends on uninitialised value(s) ==54391== at 0x50AE1A2: ucs_rcache_check_overlap (rcache.c:860) ==54391==...
This is double commit of https://github.com/openucx/ucx/pull/9880, into v1.17.x branch Fix for [RM#3918537](https://redmine.mellanox.com/issues/3918537) I managed to reproduce this issue on rock machines in 100% of the cases, but only when running...
## What Redesigned peer accessible cache in uct_cuda_ipc component. There were several design flaws in existing implementation: - uct_cuda_ipc_component stores a reference to the last created MD. However if last...