Benchmark tool for uniform neighbor sampling C API #3071
Conversation
ChuckHastings
added
improvement
| @@ -53,6 +53,7 @@ target_link_libraries(cugraphtestutil | |||
| PRIVATE | |||
| cuco::cuco | |||
| GTest::gtest | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| @@ -68,6 +69,7 @@ function(ConfigureTest CMAKE_TEST_NAME) | |||
| GTest::gtest | |||
| GTest::gtest_main | |||
| NCCL::NCCL | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| @@ -95,6 +97,7 @@ function(ConfigureTestMG CMAKE_TEST_NAME) | |||
| GTest::gtest_main | |||
| NCCL::NCCL | |||
| MPI::MPI_CXX | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| @@ -126,6 +129,7 @@ function(ConfigureCTest CMAKE_TEST_NAME) | |||
| cugraph_c_testutil | |||
| GTest::gtest | |||
| GTest::gtest_main | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| @@ -152,6 +156,7 @@ function(ConfigureCTestMG CMAKE_TEST_NAME) | |||
| GTest::gtest_main | |||
| NCCL::NCCL | |||
| MPI::MPI_CXX | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| @@ -412,6 +417,7 @@ if(BUILD_CUGRAPH_MG_TESTS) | |||
| MPI::MPI_CXX | |||
| PRIVATE | |||
| GTest::gtest | |||
| rt | |||
There was a problem hiding this comment.
Now we should be able to remove this.
| raft::handle_t* raft_handle = static_cast<raft::handle_t*>(create_raft_handle(prows)); | ||
| handle = cugraph_create_resource_handle(raft_handle); | ||
|
|
||
| #ifdef AFTER_3601_MERGES |
There was a problem hiding this comment.
I think 3601 here should be 3061.
|
|
||
| rmm::device_scalar<vertex_t> d_num_vertices(number_of_vertices, handle.get_stream()); | ||
| raft::random::RngState rng_state(seed); | ||
|
|
There was a problem hiding this comment.
We may need to consider when #ifdef NO_CUGRAPH_OPS is true.
And should this low level testing utility function be dependent on cugraph-ops? (and under the current cugraph-ops implementation, this will be served by 1 CUDA warp).
We may consider an alternative.
Say when count << num_vertices,
Generate count + alpha random indices and remove duplicates and take first count (generate additional random numbers if the resulting unique indices are fewer than count).
If count is in the same order of num_vertices, use thrust::shuffle to shuffle indices from 0 to num_vertices - 1 and take the first count indices.
| number_of_vertices = vertex_ends.back(); | ||
| int rank = handle.get_comms().get_rank(); | ||
| if (rank > 0) base_index = vertex_ends[rank - 1]; | ||
| } |
There was a problem hiding this comment.
What will be the behavior of this function if count is larger than number_of_vertices?
BradReesWork
dismissed stale reviews from jnke2016 and naimnv
The base branch was changed.
|
Not going to merge this. We don't have a real call for this capability - it was a temporary branch to do some performance testing. If we do have a need to do performance testing at the C API we should create a more comprehensive mechanism for this. |
This PR contains a new benchmarking tool for timing the uniform neighbor sampling algorithm from the C API.
The code is a mix of C/C++ - to take advantage of the C++ test infrastructure that we have built.
This branch contains a hack (adding
rtto the libraries to work around the problem corrected by #3049). Once PR 3049 is merged we can undo that hack and this will be ready to review/merge.