Feisiyicl 2190 memory usage (#2191)

* rename MILVUS_ENABLE_PROFILING to ENABLE_CPU_PROFILING, and add ENABLE_MEM_PROFILING
Signed-off-by: Nfeisiyicl <7764126@qq.com>

* optimize ReadOnlyArrayInvertedLists
Signed-off-by: Nfeisiyicl <7764126@qq.com>

* update changelog
Signed-off-by: Nfeisiyicl <7764126@qq.com>

* retry ci
Signed-off-by: Nfeisiyicl <7764126@qq.com>

* fix build issue
Signed-off-by: Nfeisiyicl <7764126@qq.com>

CHANGELOG.md

@@ -35,9 +35,11 @@ Please mark all change in change log and use the issue from GitHub
 -   \#2167 Merge log_config.conf with server_config.yaml
 -   \#2173 Check storage permission
 -   \#2178 Using elkan K-Means to improve IVF
+-   \#2190 Fix memory usage is twice of index size when using GPU searching
 
 ## Task
 
+
 # Milvus 0.8.0 (2020-04-15)
 
 ## Bug

core/src/CMakeLists.txt

@@ -212,9 +212,10 @@ if (MILVUS_GPU_VERSION)
             )
 endif ()
 
-if (MILVUS_ENABLE_PROFILING)
+if (ENABLE_CPU_PROFILING OR ENABLE_MEM_PROFILING)
     set(third_party_libs ${third_party_libs}
             gperftools
+            tcmalloc
             libunwind
             )
 endif ()

core/src/index/knowhere/knowhere/index/vector_index/IndexIVF.cpp

@@ -310,8 +310,6 @@ IVF::SealImpl() {
     faiss::Index* index = index_.get();
     auto idx = dynamic_cast<faiss::IndexIVF*>(index);
     if (idx != nullptr) {
-        // To be deleted
-        LOG_KNOWHERE_DEBUG_ << "Test before to_readonly: IVF READONLY " << std::boolalpha << idx->is_readonly();
         idx->to_readonly();
     }
 #endif

core/src/index/knowhere/knowhere/index/vector_index/helpers/FaissGpuResourceMgr.cpp

@@ -124,7 +124,7 @@ void
 FaissGpuResourceMgr::Dump() {
     for (auto& item : idle_map_) {
         auto& bq = item.second;
-        // std::cout << "DEVICEID: " << item.first << ", resource count:" << bq.Size();
+        LOG_KNOWHERE_DEBUG_ << "DEVICEID: " << item.first << ", resource count:" << bq.Size();
     }
 }
 

core/src/index/thirdparty/faiss/InvertedLists.cpp

@@ -29,25 +29,32 @@ namespace faiss {
  */
 
 PageLockMemory::PageLockMemory(size_t size) : nbytes(size) {
-    CUDA_VERIFY(cudaHostAlloc(&data, size, 0));
+    auto err = cudaHostAlloc(&(this->data), size, 0);
+    if (err) {
+        FAISS_THROW_MSG("Fail to alloc page lock memory " + std::to_string(size));
+    }
 }
 
 PageLockMemory::~PageLockMemory() {
-    CUDA_VERIFY(cudaFreeHost((void*)data));
+    CUDA_VERIFY(cudaFreeHost((void*)(this->data)));
 }
 
 PageLockMemory::PageLockMemory(const PageLockMemory& other) {
-    CUDA_VERIFY(cudaHostAlloc(&data, other.nbytes, 0));
-    memcpy(data, other.data, other.nbytes);
-    nbytes = other.nbytes;
+    auto err = cudaHostAlloc(&(this->data), other.nbytes, 0);
+    if (err) {
+        FAISS_THROW_MSG("Fail to alloc page lock memory " + std::to_string(other.nbytes));
+    }
+    memcpy(this->data, other.data, other.nbytes);
+    this->nbytes = other.nbytes;
 }
 
 PageLockMemory::PageLockMemory(PageLockMemory &&other) {
-    data = other.data;
-    nbytes = other.nbytes;
+    this->data = other.data;
+    this->nbytes = other.nbytes;
     other.data = nullptr;
     other.nbytes = 0;
 }
+
 }
 #endif
 
@@ -274,34 +281,42 @@ ReadOnlyArrayInvertedLists::ReadOnlyArrayInvertedLists(size_t nlist,
 
 ReadOnlyArrayInvertedLists::ReadOnlyArrayInvertedLists(const ArrayInvertedLists& other)
         : InvertedLists (other.nlist, other.code_size) {
-#ifndef USE_CPU
-    std::vector <uint8_t> readonly_codes;
-    std::vector <idx_t> readonly_ids;
-#endif
-    readonly_length.reserve(nlist);
+    readonly_length.resize(nlist);
+    readonly_offset.resize(nlist);
     size_t offset = 0;
-    for (auto& list_ids : other.ids) {
-        readonly_length.emplace_back(list_ids.size());
-        readonly_offset.emplace_back(offset);
+    for (auto i = 0; i < other.ids.size(); i++) {
+        auto& list_ids = other.ids[i];
+        readonly_length[i] = list_ids.size();
+        readonly_offset[i] = offset;
         offset += list_ids.size();
-        readonly_ids.insert(readonly_ids.end(), list_ids.begin(), list_ids.end());
     }
 
-    for(auto& list_codes : other.codes) {
+#ifdef USE_CPU
+    for (auto i = 0; i < other.ids.size(); i++) {
+        auto& list_ids = other.ids[i];
+        readonly_ids.insert(readonly_ids.end(), list_ids.begin(), list_ids.end());
+
+        auto& list_codes = other.codes[i];
         readonly_codes.insert(readonly_codes.end(), list_codes.begin(), list_codes.end());
     }
+#else
+    size_t ids_size = offset * sizeof(idx_t);
+    size_t codes_size = offset * (this->code_size) * sizeof(uint8_t);
+    pin_readonly_codes = std::make_shared<PageLockMemory>(codes_size);
+    pin_readonly_ids = std::make_shared<PageLockMemory>(ids_size);
 
-#ifndef USE_CPU
-    // convert to page-lock memory
-    {
-        size_t size = readonly_codes.size() * sizeof(uint8_t);
-        pin_readonly_codes = std::make_shared<PageLockMemory>(size);
-        memcpy(pin_readonly_codes->data, readonly_codes.data(), size);
-    }
-    {
-        size_t size = readonly_ids.size() * sizeof(idx_t);
-        pin_readonly_ids = std::make_shared<PageLockMemory>(size);
-        memcpy(pin_readonly_ids->data, readonly_ids.data(), size);
+    offset = 0;
+    for (auto i = 0; i < other.ids.size(); i++) {
+        auto& list_ids = other.ids[i];
+        auto& list_codes = other.codes[i];
+
+        uint8_t* ids_ptr = (uint8_t*)(pin_readonly_ids->data) + offset * sizeof(idx_t);
+        memcpy(ids_ptr, list_ids.data(), list_ids.size() * sizeof(idx_t));
+
+        uint8_t* codes_ptr = (uint8_t*)(pin_readonly_codes->data) + offset * (this->code_size) * sizeof(uint8_t);
+        memcpy(codes_ptr, list_codes.data(), list_codes.size() * sizeof(uint8_t));
+
+        offset += list_ids.size();
     }
 #endif
 

core/src/server/delivery/hybrid_request/HybridSearchRequest.cpp

@@ -23,7 +23,7 @@
 #include <unordered_map>
 #include <utility>
 #include <vector>
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 #include <gperftools/profiler.h>
 #endif
 
@@ -98,7 +98,7 @@ HybridSearchRequest::OnExecute() {
         status = DBWrapper::DB()->HybridQuery(context_, collection_name_, partition_list_, hybrid_search_contxt_,
                                               general_query_, attr_type, nq, result_ids, result_distances);
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         ProfilerStop();
 #endif
 

core/src/server/delivery/hybrid_request/InsertEntityRequest.cpp

@@ -23,7 +23,7 @@
 #include <unordered_map>
 #include <utility>
 #include <vector>
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 #include <gperftools/profiler.h>
 #endif
 
@@ -123,7 +123,7 @@ InsertEntityRequest::OnExecute() {
 
         rc.RecordSection("check validation");
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         std::string fname = "/tmp/insert_" + CommonUtil::GetCurrentTimeStr() + ".profiling";
         ProfilerStart(fname.c_str());
 #endif
@@ -162,7 +162,7 @@ InsertEntityRequest::OnExecute() {
                          : collection_schema.flag_ |= engine::meta::FLAG_MASK_NO_USERID;
         status = DBWrapper::DB()->UpdateCollectionFlag(collection_name_, collection_schema.flag_);
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         ProfilerStop();
 #endif
 

core/src/server/delivery/request/InsertRequest.cpp

@@ -22,7 +22,7 @@
 #include <string>
 #include <vector>
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 #include <gperftools/profiler.h>
 #endif
 
@@ -124,7 +124,7 @@ InsertRequest::OnExecute() {
 
         rc.RecordSection("check validation");
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         std::string fname = "/tmp/insert_" + CommonUtil::GetCurrentTimeStr() + ".profiling";
         ProfilerStart(fname.c_str());
 #endif
@@ -167,7 +167,7 @@ InsertRequest::OnExecute() {
                          : collection_schema.flag_ |= engine::meta::FLAG_MASK_NO_USERID;
         status = DBWrapper::DB()->UpdateCollectionFlag(collection_name_, collection_schema.flag_);
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         ProfilerStop();
 #endif
 

core/src/server/delivery/request/SearchByIDRequest.cpp

@@ -26,7 +26,7 @@
 #include "utils/TimeRecorder.h"
 #include "utils/ValidationUtil.h"
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 #include <gperftools/profiler.h>
 #endif
 
@@ -122,7 +122,7 @@ SearchByIDRequest::OnExecute() {
         engine::ResultIds result_ids;
         engine::ResultDistances result_distances;
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         std::string fname = "/tmp/search_by_id_" + CommonUtil::GetCurrentTimeStr() + ".profiling";
         ProfilerStart(fname.c_str());
 #endif
@@ -132,7 +132,7 @@ SearchByIDRequest::OnExecute() {
         status = DBWrapper::DB()->QueryByIDs(context_, collection_name_, partition_list_, (size_t)topk_, extra_params_,
                                              id_array_, result_ids, result_distances);
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         ProfilerStop();
 #endif
 

core/src/server/delivery/request/SearchRequest.cpp

@@ -22,7 +22,7 @@
 #include "utils/TimeRecorder.h"
 #include "utils/ValidationUtil.h"
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 #include <gperftools/profiler.h>
 #endif
 
@@ -136,7 +136,7 @@ SearchRequest::OnExecute() {
         rc.RecordSection("check validation");
 
         // step 7: search vectors
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         std::string fname = "/tmp/search_" + CommonUtil::GetCurrentTimeStr() + ".profiling";
         ProfilerStart(fname.c_str());
 #endif
@@ -154,7 +154,7 @@ SearchRequest::OnExecute() {
 
         rc.RecordSection("query vectors from engine");
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
         ProfilerStop();
 #endif
         fiu_do_on("SearchRequest.OnExecute.query_fail", status = Status(milvus::SERVER_UNEXPECTED_ERROR, ""));

core/src/utils/CommonUtil.cpp

@@ -228,7 +228,7 @@ CommonUtil::ConvertTime(tm time_struct, time_t& time_integer) {
     time_integer = mktime(&time_struct);
 }
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
 std::string
 CommonUtil::GetCurrentTimeStr() {
     time_t tt;

core/src/utils/CommonUtil.h

@@ -51,7 +51,7 @@ class CommonUtil {
     static void
     ConvertTime(tm time_struct, time_t& time_integer);
 
-#ifdef MILVUS_ENABLE_PROFILING
+#ifdef ENABLE_CPU_PROFILING
     static std::string
     GetCurrentTimeStr();
 #endif