Replace rmm::device_vector & thrust::host_vector with rmm::device_uvector & std::vector, respectively.

[seunghwak]

• Replace rmm::device_vector with rmm::device_uvector for better concurrency in multi-stream executions
• Replace thrust::host_vector with std::vector

This PR partially addresses #1390

[seunghwak]

@afender you may want to review this.

[afender]

Looks good but :

• consider using the latest streams features from RAFT
• could we think of a test to verify it is now executing on the stream?

[afender]

Do we really need this sync?

If so, consider handle.get_stream_view.synchronize() instead.
Or handle.wait_on_user_stream but this triggers worker stream sync as well

[seunghwak]

This is necessary as without this, bucket_sizes (which is std::vector and is not aware of stream) is not guaranteed to hold up-to-date data and is used right after this sync.

CUDA_TRY(cudaStreamSynchronize(handle.get_stream())); this is also used all around the codebase, so I will create an issue to replace all at once.

[afender]

Consider replacing
thrust::fill(rmm::exec_policy(handle_ptr_->get_stream())->on(handle_ptr_->get_stream()),
by
thrust::fill(rmm::exec_policy(handle.get_stream_view()),

This applies to all thrust calls

[seunghwak]

Yes, same story here. This should be replaced all at once.

[seunghwak]

@afender #1422 is created to transition cuGraph to rmm's new cuda_stream_view.

[seunghwak]

• could we think of a test to verify it is now executing on the stream?

This is tricky I guess; the first thing is this is not a black-or-white thing and this will affect only performance not correctness.

I guess we need cuGraph-wise benchmark scheme to track multi-stream execution of (every) graph analytics?

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (branch-0.19@3f13ffc). Click here to learn what that means.
The diff coverage is n/a.

@@              Coverage Diff               @@
##             branch-0.19    #1421   +/-   ##
==============================================
  Coverage               ?   60.75%           
==============================================
  Files                  ?       70           
  Lines                  ?     3134           
  Branches               ?        0           
==============================================
  Hits                   ?     1904           
  Misses                 ?     1230           
  Partials               ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3f13ffc...9350f96. Read the comment docs.

[afender]

• could we think of a test to verify it is now executing on the stream?

This is tricky I guess; the first thing is this is not a black-or-white thing and this will affect only performance not correctness.

I guess we need cuGraph-wise benchmark scheme to track multi-stream execution of (every) graph analytics?

Ok, for now I can rebase my concurrency branch and generate a profile after this gets merged.

[BradReesWork]

@gpucibot merge