Raul Akhmetshin comments

Results 47 comments of


                                            Raul Akhmetshin

UCT/API: Removed count and stride fields from uct_iov_t structure.

Please find performance test results showing the impact of removing the fields from the structure. iovcnt | Avg Latency v0*, ns | Avg Latency v1**, ns | Diff, ns*** |...

UCT/API: Removed count and stride fields from uct_iov_t structure.

> @yosefe Just saw this. Can you share the v0 code ? Seems like this is Thunderx-2. You have to pin processes on this platform otherwise it is very noisy....

UCT/API: Removed count and stride fields from uct_iov_t structure.

> FYI, my command line does not accept this size format. @shamisp Yes, I forgot about this. Sorry for the inconvenience. For quick workaround please replace `max_iov = attr.cap.am.max_iov;` by...

UCT/API: Removed count and stride fields from uct_iov_t structure.

> @rakhmets can you please tree/branches for v0/v1 ? thanks. @shamisp v0 was mid-december master branch. And v1 was v0 + changes from this pull request. I think it is...

UCT/API: Removed count and stride fields from uct_iov_t structure.

bot:pipe:retest

openmpi pml ucx cannot be selected when linking to cuda object file

Ok, I will take a look.

UCP/CORE: Fixed wiface activation counter.

/azp run

CMAKE: Add targets once in case of multiple find_package() - v1.16.x

> we'd also need `NEWS` file update under rc3? I've just created a pre-release... But I can remove tag and RC to avoid having two RCs for two single PRs.

CMAKE: Add targets once in case of multiple find_package() - v1.16.x

> > we'd also need `NEWS` file update under rc3? > > I've just created a pre-release... But I can remove tag and RC to avoid having two RCs for...

GTEST/UCT: Do not query BAR1 memory using NVML under valgrind.

> can you please commit it to #9531 to verify that it fixes the issue? @brminich pushed the changes to https://github.com/openucx/ucx/pull/9491