Faiss IVFPQ (GPU) Enable raft-on cause /dev/nvidiactl ioctl race

[neezeeyee]

Severe Performance Degradation Due to NVIDIA Driver (nvidiactl) ioctl Contention When RAFT is Enabled in Faiss GPU Build

Environment

• Faiss Version: 1.9.0

• RAFT Version: 24.06.00

• GPU

• Faiss Build Configuration:

  # Docker build snippet
RUN set -ex \
  && mkdir -p /root/logs \
  && rm -rf /lib64/libstdc++.so.6.0.25-gdb.py \
  && ldconfig \
  && cd cmake-3.29.4 \
  && ./configure --prefix=/usr/local/cmake \
  && make && make install \
  && ln -s /usr/local/cmake/bin/cmake /usr/bin/cmake \
  && cd .. && rm -rf cmake-3.29.4 \
  && cd raft-24.06.00 \
  && ./build.sh libraft \
  && cd ..

ENV raft_DIR=/home/raft-24.06.00/cpp/build/

RUN cd faiss-1.9.0 \
  && cmake -B build . -DFAISS_ENABLE_GPU=ON -DFAISS_ENABLE_PYTHON=OFF -DBUILD_TESTING=OFF -DCUDAToolkit_ROOT=/usr/local/cuda/targets/x86_64-linux/include -DBUILD_SHARED_LIBS=ON -DFAISS_ENABLE_RAFT=ON  -DCMAKE_BUILD_TYPE=Release  -DCMAKE_CUDA_ARCHITECTURES="70;75;80;86;90" \
  && make -C build install -j6 && cd .. \
  && rm -rf faiss-1.9.0

Problem Description

We run GPU searches using the following workflow

std::unique_ptr<faiss::Index> index_cpu_ptr(faiss::read_index(faiss_local_file_path.c_str(), 0));
faiss::Index* cpu_ptr = cpu_unique_ptr.release();
 faiss::gpu::StandardGpuResources res;
faiss::Index* index_gpu = faiss::gpu::index_cpu_to_gpu(&res, 0, cpu_ptr, &options);
index_gpu->search();

Within a single k8s Pod, multiple GPU indices (gpu_index) may coexist, where each index is bound to an isolated StandardGpuResources, Under normal circumstances, search operations across these indices should not interfere with each other due to resource isolation.

1. Cross-Pod Driver-Level Contention

When monitoring multiple k8s Pods sharing the same GPU host, we observed heavy contention through:

@ioctl_file[b, /dev/nvidiactl, NV_ESC_RM_FREE]:  // 
[0]                 2248 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|

@ioctl_file[b /dev/nvidiactl, NV_ESC_RM_VID_HEAP_CONTROL]: 
[0]                 2248 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|

2. Within-Pod Contention

raft race

Solve

remove DFAISS_ENABLE_RAFT=ON and recomplie faiss

[cjnolet]

Hi @neezeeyee. There have been significant updates to this since June of last year (version 24.06 denotes June of 2024). Are you able to update FAISS to use the new cuVS library? It's also possible updating RAFT could work @tarang-jain do you recall what version of RAFT was released w/ FAISS 1.9.0?

[tarang-jain]

Faiss 1.9.0 had RAFT=24.06 (https://github.com/facebookresearch/faiss/blob/v1.9.0/cmake/thirdparty/fetch_rapids.cmake)

[cjnolet]

Thanks @tarang-jain. @neezeeyee in that case, it might be easier to try the latest version of FAISS (1.10), which upgrades the RAFT dependency to cuVS (and provides a wealth of new capabilities and bug fixes).

[neezeeyee]

@cjnolet Are there any known issues with FAISS 1.9.0 and RAFT 24.06 when performing GPU-based searches? Will upgrading versions definitely resolve the problem or only potentially resolve it? What is the root cause of the current observed issue? thx

[mdouze]

Normally there should be one StandardGpuResources per device not per index.

[neezeeyee]

@mdouze But the concurrency level is determined by the GpuResources configuration. Each GpuResources instance can handle only one request at a time. which means that If all indexes share the same GpuResources, search operations across different indexes will be queued sequentially.