ucx

ucx copied to clipboard

## What connect lanes in reverse ordering ## Why ? - to optimize multiple re-scheduling of requests with multi-fragment protocols - to avoid re-initializtion of started protocol according to #8386

evgeny-leksikov

UCT/API: add dmabuf to md_mem_query attributes

13

## What Add dmabuf fd field in md_mem attributes ## Why ? Needed by UCT/IB to register device memory exposed as a dmabuf

Akshay-Venkatesh

## What Implementation of `ucp_proto_t::clean` for eager protocols Depends on https://github.com/openucx/ucx/pull/8386 ## Why ? The method was intoduced in https://github.com/openucx/ucx/pull/8386 and should be implemented for all protocols

evgeny-leksikov

Build failure

2

Using clang 13, ucx 1.13.0-rc1 and CUDA 11.7 ``` $ grep 'implicit conversion' build.txt cuda_copy/cuda_copy_ep.c:120:31: error: implicit conversion from enumeration type 'CUresult' (aka 'enum cudaError_enum') to different enumeration type 'cudaError_t'...

haampie

Core dumps from ucx_info -d

3

### Describe the bug Segfaults/core dumps when running `ucx_info -d` ### Steps to Reproduce - ucx 1.13.0 release installation with `./configure --prefix=$HOME/bin/ucx-1.13.0` ``` ucx_info -v # Version 1.13.0 # Git...

cniethammer

AMD x86 cluster: git HEAD hanging now but 1.12.1 working

7

### Describe the bug Was using 1.12.1 on an AMD x86 cluster, then switched to github head, programs working with IB transports (rc, dc) now hang. Single node runs with...

tonycurtis

UCX ERROR ibv_reg_mr failed: Cannot allocate memory

25

### Describe the bug I am running Openfoam-v2012 simulation (turbulent channel with particles) on cluster with 128 cores . The simulation works fine till 1 hour and then it shows...

abhishekiitkgp9191

UCC mpitest needs -x UCX_IB_PREFER_NEAREST_DEVICE=n flag to run on helios

3

### Describe the bug UCC mpitest fails to run on helios machine on hpcadvisorycouncil, unless flag is added -x UCX_IB_PREFER_NEAREST_DEVICE=n. Note that for example on Rome machine on hpcadvisorycouncil, mpitest...

shimmybalsam

UCP: Ignore path idx for non-bw lanes

2

## What Ignore path indexes when adding new lane and there is existing non-bw lane with the same tl resources ## Why ? The problem shown on the picture below...

brminich

Failing unit tests with ucx 1.14.0 and ROCm 5.1.0

4

### Describe the bug ROCm related unit test failed see [rocm.log](https://github.com/openucx/ucx/files/9076538/rocm.log) ### Setup and versions - CentOs stream 8 - rdma-core-55mlnx-37-1.55103.x86_64 ``` ibv_devinfo hca_id: mlx5_0 transport: InfiniBand (0) fw_ver: 20.32.1010...

greole