Adding Pinned host buffer benchmarks #576
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
nirandaperera
added
improvement
ResultsSystem: nvcc: NVIDIA (R) Cuda compiler driver 
|
Signed-off-by: niranda perera <[email protected]>
Signed-off-by: niranda perera <[email protected]>
Signed-off-by: niranda perera <[email protected]>
nirandaperera
force-pushed
the
pinned_host_buffer_bench
branch
from
f6a2faf to
ddc5452
Compare
There was a problem hiding this comment.
Overall looks good but we need to run a smoke test in CI: https://github.com/rapidsai/rapidsmpf/blob/f67d8ecc3a5ff83d3d0bf92dc10f2be2a499ea24/ci/run_cpp_benchmark_smoketests.sh
Signed-off-by: niranda perera <[email protected]>
| auto latency_to_first = std::chrono::duration_cast<std::chrono::nanoseconds>( | ||
| first_allocation_time - start_time | ||
| ) | ||
| .count(); | ||
| auto first_round_duration_ns = | ||
| std::chrono::duration_cast<std::chrono::nanoseconds>( | ||
| first_round_end - start_time | ||
| ) | ||
| .count(); | ||
| auto second_round_duration_ns = | ||
| std::chrono::duration_cast<std::chrono::nanoseconds>( | ||
| second_round_end - first_round_end | ||
| ) | ||
| .count(); |
There was a problem hiding this comment.
These timings are at least partly nonsense. The latency to first makes sense. The first_round_duration kind of doesn't because we've synced after the very first allocation, but ok. I guess it makes kind of sense.
The second_round_duration includes the time to deallocate all of the allocations from the first round. This makes no sense.
There was a problem hiding this comment.
Oh yeah. Good point. I was simply following this
https://github.com/rapidsai/rmm/blob/branch-25.12/cpp/benchmarks/async_priming/async_priming_bench.cpp
I guess the same issue is here as well.
Co-authored-by: Lawrence Mitchell <[email protected]>
Co-authored-by: Lawrence Mitchell <[email protected]>
Signed-off-by: niranda perera <[email protected]>
3 participants
This PR adds benchmarks for pinned host buffer
Depends on #549