Refactor optimizer (a13677b6) · Commits · SUMMER2020 / students / proj-2018085

core/src/scheduler/SchedInst.h

+10 −12

Original line number	Diff line number	Diff line
		@@ -24,9 +24,10 @@
		#include "Utils.h"
		#include "optimizer/BuildIndexPass.h"
		#include "optimizer/FallbackPass.h"
		#include "optimizer/HybridPass.h"
		#include "optimizer/LargeSQ8HPass.h"
		#include "optimizer/OnlyCPUPass.h"
		#include "optimizer/FaissFlatPass.h"
		#include "optimizer/FaissIVFFlatPass.h"
		#include "optimizer/FaissIVFSQ8Pass.h"
		#include "optimizer/FaissIVFSQ8HPass.h"
		#include "optimizer/Optimizer.h"
		#include "server/Config.h"

		@@ -100,15 +101,12 @@ class OptimizerInst {
		std::lock_guard<std::mutex> lock(mutex_);
		if (instance == nullptr) {
		std::vector<PassPtr> pass_list;
		pass_list.push_back(std::make_shared<LargeSQ8HPass>());
		pass_list.push_back(std::make_shared<HybridPass>());
		#ifdef MILVUS_CPU_VERSION
		pass_list.push_back(std::make_shared<OnlyCPUPass>());
		#else
		server::Config& config = server::Config::GetInstance();
		std::vector<int32_t> build_resources;
		config.GetGpuResourceConfigBuildIndexResources(build_resources);
		pass_list.push_back(std::make_shared<BuildIndexPass>(build_resources));
		#ifdef MILVUS_GPU_VERSION
		pass_list.push_back(std::make_shared<BuildIndexPass>());
		pass_list.push_back(std::make_shared<FaissFlatPass>());
		pass_list.push_back(std::make_shared<FaissIVFFlatPass>());
		pass_list.push_back(std::make_shared<FaissIVFSQ8Pass>());
		pass_list.push_back(std::make_shared<FaissIVFSQ8HPass>());
		#endif
		pass_list.push_back(std::make_shared<FallbackPass>());
		instance = std::make_shared<Optimizer>(pass_list);

core/src/scheduler/optimizer/BuildIndexPass.cpp

+6 −3

Original line number	Diff line number	Diff line
		@@ -23,11 +23,14 @@
		namespace milvus {
		namespace scheduler {

		BuildIndexPass::BuildIndexPass(std::vector<int32_t>& build_gpu_ids) : build_gpu_ids_(build_gpu_ids) {
		}

		void
		BuildIndexPass::Init() {
		server::Config& config = server::Config::GetInstance();
		std::vector<int32_t> build_resources;
		Status s = config.GetGpuResourceConfigBuildIndexResources(build_resources);
		if (!s.ok()) {
		throw;
		}
		}

		bool

core/src/scheduler/optimizer/BuildIndexPass.h

+1 −1

Original line number	Diff line number	Diff line
		@@ -34,7 +34,7 @@ namespace scheduler {

		class BuildIndexPass : public Pass {
		public:
		explicit BuildIndexPass(std::vector<int32_t>& build_gpu_id);
		BuildIndexPass() = default;

		public:
		void

core/src/scheduler/optimizer/HybridPass.cpp→core/src/scheduler/optimizer/FaissFlatPass.cpp

+69 −0

Original line number	Diff line number	Diff line
		@@ -15,33 +15,55 @@
		// specific language governing permissions and limitations
		// under the License.

		#include "scheduler/optimizer/HybridPass.h"
		#include "scheduler/optimizer/FaissFlatPass.h"
		#include "cache/GpuCacheMgr.h"
		#include "scheduler/SchedInst.h"
		#include "scheduler/Utils.h"
		#include "scheduler/task/SearchTask.h"
		#include "scheduler/tasklabel/SpecResLabel.h"
		#include "server/Config.h"
		#include "utils/Log.h"

		namespace milvus {
		namespace scheduler {

		void
		HybridPass::Init() {
		FaissFlatPass::Init() {
		server::Config& config = server::Config::GetInstance();
		Status s = config.GetEngineConfigGpuSearchThreshold(threshold_);
		if (!s.ok()) {
		threshold_ = std::numeric_limits<int32_t>::max();
		}
		s = config.GetGpuResourceConfigSearchResources(gpus);
		if (!s.ok()) {
		throw;
		}
		}

		bool
		HybridPass::Run(const TaskPtr& task) {
		// TODO: future, Index::IVFSQ8H, if nq < threshold set cpu, else set gpu
		if (task->Type() != TaskType::SearchTask)
		FaissFlatPass::Run(const TaskPtr& task) {
		if (task->Type() != TaskType::SearchTask) {
		return false;
		}

		auto search_task = std::static_pointer_cast<XSearchTask>(task);
		if (search_task->file_->engine_type_ == (int)engine::EngineType::FAISS_IVFSQ8H) {
		// TODO: remove "cpu" hardcode
		ResourcePtr res_ptr = ResMgrInst::GetInstance()->GetResource("cpu");
		auto label = std::make_shared<SpecResLabel>(std::weak_ptr<Resource>(res_ptr));
		if (search_task->file_->engine_type_ != (int)engine::EngineType::FAISS_IDMAP) {
		return false;
		}

		auto search_job = std::static_pointer_cast<SearchJob>(search_task->job_.lock());
		ResourcePtr res_ptr;
		if (search_job->nq() < threshold_) {
		res_ptr = ResMgrInst::GetInstance()->GetResource("cpu");
		} else {
		auto best_device_id = count_ % gpus.size();
		count_++;
		res_ptr = ResMgrInst::GetInstance()->GetResource(ResourceType::GPU, best_device_id);
		}
		auto label = std::make_shared<SpecResLabel>(res_ptr);
		task->label() = label;
		return true;
		}
		return false;
		}

		} // namespace scheduler
		} // namespace milvus

core/src/scheduler/optimizer/LargeSQ8HPass.h→core/src/scheduler/optimizer/FaissFlatPass.h

+3 −3

Original line number	Diff line number	Diff line
		@@ -33,9 +33,9 @@
		namespace milvus {
		namespace scheduler {

		class LargeSQ8HPass : public Pass {
		class FaissFlatPass : public Pass {
		public:
		LargeSQ8HPass() = default;
		FaissFlatPass() = default;

		public:
		void
		@@ -50,7 +50,7 @@ class LargeSQ8HPass : public Pass {
		std::vector<int32_t> gpus;
		};

		using LargeSQ8HPassPtr = std::shared_ptr<LargeSQ8HPass>;
		using FaissFlatPassPtr = std::shared_ptr<FaissFlatPass>;

		} // namespace scheduler
		} // namespace milvus