FAISS with cuVS enabled in cuvs-bench #561
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. |
tarang-jain
added
non-breaking
| @@ -483,7 +483,14 @@ void register_search(std::shared_ptr<const dataset<T>> dataset, | |||
| ->MeasureProcessCPUTime() | |||
| ->UseRealTime(); | |||
|
|
|||
| if (metric_objective == Mode::kThroughput) { b->ThreadRange(threads[0], threads[1]); } | |||
| if (metric_objective == Mode::kThroughput) { | |||
There was a problem hiding this comment.
Let's not put algorithm-specific stuff in the common files. I agree with @achirkin about this one. If needed, you could propoagate this down to the specific algorithms, but the common stuff should be agnostic of he algoithmmic specific.
There was a problem hiding this comment.
Thanks, @cjnolet for spotting this. @tarang-jain , could you please move the warning into the faiss-related header (perhaps into the set_search_params)? There, you should be able to query the cuvs::bench::benchmark_n_threads > 1 from common/util.hpp.
| int main(int argc, char** argv) | ||
| { | ||
| rmm::mr::cuda_memory_resource cuda_mr; | ||
| // Construct a resource that uses a coalescing best-fit pool allocator | ||
| // and is initially sized to half of free device memory. | ||
| rmm::mr::pool_memory_resource<rmm::mr::cuda_memory_resource> pool_mr{ | ||
| &cuda_mr, rmm::percent_of_free_device_memory(50)}; | ||
| // Updates the current device resource pointer to `pool_mr` | ||
| auto old_mr = rmm::mr::set_current_device_resource(&pool_mr); | ||
| auto ret = cuvs::bench::run_main(argc, argv); | ||
| // Restores the current device resource pointer to its previous value | ||
| rmm::mr::set_current_device_resource(old_mr); | ||
| return ret; | ||
| } |
There was a problem hiding this comment.
We've run in many issues in ann-bench in the past doing the program-wide setting of memory resource. This also changes the behavior depending on whether ANN_BENCH_BUILD_MAIN flag is enabled. Therefore, it's best if these lines stay unmodified and same for all algorithms.
Please move this resource setting to the place where the index wrapper is constructed/destructed in the algo header file.
Mostly adapted from rapidsai/raft#2026