Unverified Commit 796750f6 authored by Jin Hai's avatar Jin Hai Committed by GitHub
Browse files

Merge pull request #414 from fishpenguin/0.6.0-yk-update-config 

Support build index with multiple gpu and update config
parents 6d9ed166 c451f8cb

CHANGELOG.md

+2 −1
Original line number Diff line number Diff line
@@ -17,8 +17,9 @@ Please mark all change in change log and use the ticket from JIRA. 
## Feature

- \#12 - Pure CPU version for Milvus

- \#77 - Support table partition

- \#226 - Experimental shards middleware for Milvus

- \#127 - Support new Index type IVFPQ

- \#226 - Experimental shards middleware for Milvus

- \#346 - Support build index with multiple gpu



## Improvement

- \#275 - Rename C++ SDK IndexType

core/conf/server_cpu_config.template

+8 −7
Original line number Diff line number Diff line
@@ -27,9 +27,7 @@ metric_config: 
    port: 8080                      # port prometheus uses to fetch metrics, must in range [1025, 65534]



cache_config:



  cpu_cache_capacity: 16            # GB, CPU memory used for cache, must be a positive integer

  cpu_cache_threshold: 0.85         # percentage of data that will be kept when cache cleanup is triggered, must be in range (0.0, 1.0]

  cpu_cache_capacity: 16            # GB, size of CPU memory used for cache, must be a positive integer

  cache_insert_data: false          # whether to load inserted data into cache, must be a boolean



engine_config:
@@ -37,7 +35,10 @@ engine_config: 
                                    # if nq >= use_blas_threshold, use OpenBlas, slower with stable response times

  gpu_search_threshold: 1000        # threshold beyond which the search computation is executed on GPUs only



resource_config:

  search_resources:                 # define the device used for search computation

    - cpu

  index_build_device: cpu           # CPU used for building index
 
gpu_resource_config:

  enable_gpu: false                 # whether to enable GPU resources

  cache_capacity: 4                 # GB, size of GPU memory per card used for cache, must be a positive integer

  search_resources:                 # define the GPU devices used for search computation, must be in format gpux

    - gpu0

  build_index_resources:            # define the GPU devices used for index building, must be in format gpux

    - gpu0

core/conf/server_gpu_config.template

+7 −8
Original line number Diff line number Diff line
@@ -27,10 +27,7 @@ metric_config: 
    port: 8080                      # port prometheus uses to fetch metrics, must in range [1025, 65534]



cache_config:

  cpu_cache_capacity: 16            # GB, CPU memory used for cache, must be a positive integer

  cpu_cache_threshold: 0.85         # percentage of data that will be kept when cache cleanup is triggered, must be in range (0.0, 1.0]

  gpu_cache_capacity: 4             # GB, GPU memory used for cache, must be a positive integer

  gpu_cache_threshold: 0.85         # percentage of data that will be kept when cache cleanup is triggered, must be in range (0.0, 1.0]

  cpu_cache_capacity: 16            # GB, size of CPU memory used for cache, must be a positive integer

  cache_insert_data: false          # whether to load inserted data into cache, must be a boolean



engine_config:
@@ -38,8 +35,10 @@ engine_config: 
                                    # if nq >= use_blas_threshold, use OpenBlas, slower with stable response times

  gpu_search_threshold: 1000        # threshold beyond which the search computation is executed on GPUs only



resource_config:

  search_resources:                 # define the devices used for search computation, must be in format: cpu or gpux

    - cpu

gpu_resource_config:

  enable_gpu: true                  # whether to enable GPU resources

  cache_capacity: 4                 # GB, size of GPU memory per card used for cache, must be a positive integer

  search_resources:                 # define the GPU devices used for search computation, must be in format gpux

    - gpu0

  build_index_resources:            # define the GPU devices used for index building, must be in format gpux

    - gpu0
 
  index_build_device: gpu0          # CPU / GPU used for building index, must be in format: cpu or gpux

core/src/cache/GpuCacheMgr.cpp

+2 −2
Original line number Diff line number Diff line
@@ -37,7 +37,7 @@ GpuCacheMgr::GpuCacheMgr() { 
    Status s;



    int64_t gpu_cache_cap;

    s = config.GetCacheConfigGpuCacheCapacity(gpu_cache_cap);

    s = config.GetGpuResourceConfigCacheCapacity(gpu_cache_cap);

    if (!s.ok()) {

        SERVER_LOG_ERROR << s.message();

    }
@@ -45,7 +45,7 @@ GpuCacheMgr::GpuCacheMgr() { 
    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32);



    float gpu_mem_threshold;

    s = config.GetCacheConfigGpuCacheThreshold(gpu_mem_threshold);

    s = config.GetGpuResourceConfigCacheThreshold(gpu_mem_threshold);

    if (!s.ok()) {

        SERVER_LOG_ERROR << s.message();

    }

core/src/db/engine/ExecutionEngineImpl.cpp

+17 −5
Original line number Diff line number Diff line
@@ -144,7 +144,14 @@ ExecutionEngineImpl::HybridLoad() const { 
    }



    const std::string key = location_ + ".quantizer";

    std::vector<uint64_t> gpus = scheduler::get_gpu_pool();



    server::Config& config = server::Config::GetInstance();

    std::vector<int32_t> gpus;

    Status s = config.GetGpuResourceConfigSearchResources(gpus);

    if (!s.ok()) {

        ENGINE_LOG_ERROR << s.message();

        return;

    }



    // cache hit

    {
@@ -355,6 +362,7 @@ ExecutionEngineImpl::CopyToGpu(uint64_t device_id, bool hybrid) { 


Status

ExecutionEngineImpl::CopyToIndexFileToGpu(uint64_t device_id) {

    gpu_num_ = device_id;

    auto to_index_data = std::make_shared<ToIndexData>(PhysicalSize());

    cache::DataObjPtr obj = std::static_pointer_cast<cache::DataObj>(to_index_data);

    milvus::cache::GpuCacheMgr::GetInstance(device_id)->InsertItem(location_, obj);
@@ -578,12 +586,16 @@ ExecutionEngineImpl::GpuCache(uint64_t gpu_id) { 
Status

ExecutionEngineImpl::Init() {

    server::Config& config = server::Config::GetInstance();

    Status s = config.GetResourceConfigIndexBuildDevice(gpu_num_);

    if (!s.ok()) {

        return s;

    std::vector<int32_t> gpu_ids;

    Status s = config.GetGpuResourceConfigBuildIndexResources(gpu_ids);

    for (auto id : gpu_ids) {

        if (gpu_num_ == id) {

            return Status::OK();

        }

    }



    return Status::OK();

    std::string msg = "Invalid gpu_num";

    return Status(SERVER_INVALID_ARGUMENT, msg);

}



}  // namespace engine