#1735 fix search fail with gpu out of memory using ivf_flag (#1748)

* #1735 fix search fail with gpu out of memory using ivf_flag

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* #1735 add lock for each GpuCacheMgr instance

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* #1735 optimize cache debug info

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* #1735 set smaller cache threshold

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* #1735 update cache debug log

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* fix unittest

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

* #1735 fix cache issue

Signed-off-by: yudong.cai <yudong.cai@zilliz.com>

CHANGELOG.md

@@ -17,6 +17,7 @@ Please mark all change in change log and use the issue from GitHub
 -   \#1708 NSG search crashed
 -   \#1724 Remove unused unittests
 -   \#1734 Opentracing for combined search request
+-   \#1735 Fix search out of memory with ivf_flat 
 
 ## Feature
 -   \#1603 BinaryFlat add 2 Metric: Substructure and Superstructure

core/src/cache/Cache.h

@@ -26,7 +26,7 @@ template <typename ItemObj>
 class Cache {
  public:
     // mem_capacity, units:GB
-    Cache(int64_t capacity_gb, uint64_t cache_max_count);
+    Cache(int64_t capacity_gb, int64_t cache_max_count, const std::string& header = "");
     ~Cache() = default;
 
     int64_t
@@ -80,6 +80,7 @@ class Cache {
     free_memory();
 
  private:
+    std::string header_;
     int64_t usage_;
     int64_t capacity_;
     double freemem_percent_;

core/src/cache/Cache.inl

@@ -12,12 +12,16 @@
 namespace milvus {
 namespace cache {
 
-constexpr double DEFAULT_THRESHHOLD_PERCENT = 0.85;
+constexpr double DEFAULT_THRESHHOLD_PERCENT = 0.7;
+constexpr double WARNING_THRESHHOLD_PERCENT = 0.9;
 
 template <typename ItemObj>
-Cache<ItemObj>::Cache(int64_t capacity, uint64_t cache_max_count)
-    : usage_(0), capacity_(capacity), freemem_percent_(DEFAULT_THRESHHOLD_PERCENT), lru_(cache_max_count) {
-    //    AGENT_LOG_DEBUG << "Construct Cache with capacity " << std::to_string(mem_capacity)
+Cache<ItemObj>::Cache(int64_t capacity, int64_t cache_max_count, const std::string& header)
+    : usage_(0),
+      capacity_(capacity),
+      header_(header),
+      freemem_percent_(DEFAULT_THRESHHOLD_PERCENT),
+      lru_(cache_max_count) {
 }
 
 template <typename ItemObj>
@@ -61,12 +65,7 @@ Cache<ItemObj>::insert(const std::string& key, const ItemObj& item) {
         return;
     }
 
-    //    if(item->size() > capacity_) {
-    //        SERVER_LOG_ERROR << "Item size " << item->size()
-    //                        << " is too large to insert into cache, capacity " << capacity_;
-    //        return;
-    //    }
-
+    size_t item_size = item->Size();
     // calculate usage
     {
         std::lock_guard<std::mutex> lock(mutex_);
@@ -78,13 +77,13 @@ Cache<ItemObj>::insert(const std::string& key, const ItemObj& item) {
         }
 
         // plus new item size
-        usage_ += item->Size();
+        usage_ += item_size;
     }
 
     // if usage exceed capacity, free some items
-    if (usage_ > capacity_) {
-        SERVER_LOG_DEBUG << "Current usage " << usage_ << " exceeds cache capacity " << capacity_
-                         << ", start free memory";
+    if (usage_ > (int64_t)(capacity_ * WARNING_THRESHHOLD_PERCENT)) {
+        SERVER_LOG_DEBUG << header_ << " Current usage " << (usage_ >> 20) << "MB is too high for capacity "
+                         << (capacity_ >> 20) << "MB, start free memory";
         free_memory();
     }
 
@@ -93,8 +92,9 @@ Cache<ItemObj>::insert(const std::string& key, const ItemObj& item) {
         std::lock_guard<std::mutex> lock(mutex_);
 
         lru_.put(key, item);
-        SERVER_LOG_DEBUG << "Insert " << key << " size: " << item->Size() << " bytes into cache, usage: " << usage_
-                         << " bytes," << " capacity: " << capacity_ << " bytes";
+        SERVER_LOG_DEBUG << header_ << " Insert " << key << " size: " << (item_size >> 20) << "MB into cache";
+        SERVER_LOG_DEBUG << header_ << " Count: " << lru_.size() << ", Usage: " << (usage_ >> 20) << "MB, Capacity: "
+                         << (capacity_ >> 20) << "MB";
     }
 }
 
@@ -106,13 +106,15 @@ Cache<ItemObj>::erase(const std::string& key) {
         return;
     }
 
-    const ItemObj& old_item = lru_.get(key);
-    usage_ -= old_item->Size();
-
-    SERVER_LOG_DEBUG << "Erase " << key << " size: " << old_item->Size() << " bytes from cache, usage: " << usage_
-                     << " bytes," << " capacity: " << capacity_ << " bytes";
+    const ItemObj& item = lru_.get(key);
+    size_t item_size = item->Size();
 
     lru_.erase(key);
+
+    usage_ -= item_size;
+    SERVER_LOG_DEBUG << header_ << " Erase " << key << " size: " << (item_size >> 20) << "MB from cache";
+    SERVER_LOG_DEBUG << header_ << " Count: " << lru_.size() << ", Usage: " << (usage_ >> 20) << "MB, Capacity: "
+                     << (capacity_ >> 20) << "MB";
 }
 
 template <typename ItemObj>
@@ -121,15 +123,15 @@ Cache<ItemObj>::clear() {
     std::lock_guard<std::mutex> lock(mutex_);
     lru_.clear();
     usage_ = 0;
-    SERVER_LOG_DEBUG << "Clear cache !";
+    SERVER_LOG_DEBUG << header_ << " Clear cache !";
 }
 
 /* free memory space when CACHE occupation exceed its capacity */
 template <typename ItemObj>
 void
 Cache<ItemObj>::free_memory() {
-    if (usage_ <= capacity_)
-        return;
+    // if (usage_ <= capacity_)
+    //     return;
 
     int64_t threshhold = capacity_ * freemem_percent_;
     int64_t delta_size = usage_ - threshhold;
@@ -154,13 +156,11 @@ Cache<ItemObj>::free_memory() {
         }
     }
 
-    SERVER_LOG_DEBUG << "to be released memory size: " << released_size;
+    SERVER_LOG_DEBUG << header_ << " To be released memory size: " << (released_size >> 20) << "MB";
 
     for (auto& key : key_array) {
         erase(key);
     }
-
-    print();
 }
 
 template <typename ItemObj>
@@ -177,7 +177,8 @@ Cache<ItemObj>::print() {
 #endif
     }
 
-    SERVER_LOG_DEBUG << "[Cache] [item count]: " << cache_count << " [capacity] " << capacity_ << "(bytes) [usage] " << usage_ << "(bytes)";
+    SERVER_LOG_DEBUG << header_ << " [item count]: " << cache_count << ", [usage] " << (usage_ >> 20)
+                     << "MB, [capacity] " << (capacity_ >> 20) << "MB";
 }
 
 }  // namespace cache

core/src/cache/CpuCacheMgr.cpp

@@ -32,7 +32,7 @@ CpuCacheMgr::CpuCacheMgr() {
     int64_t cpu_cache_cap;
     config.GetCacheConfigCpuCacheCapacity(cpu_cache_cap);
     int64_t cap = cpu_cache_cap * unit;
-    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32);
+    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32, "[CACHE CPU]");
 
     float cpu_cache_threshold;
     config.GetCacheConfigCpuCacheThreshold(cpu_cache_threshold);

core/src/cache/GpuCacheMgr.cpp

@@ -21,21 +21,22 @@ namespace milvus {
 namespace cache {
 
 #ifdef MILVUS_GPU_VERSION
-std::mutex GpuCacheMgr::mutex_;
-std::unordered_map<uint64_t, GpuCacheMgrPtr> GpuCacheMgr::instance_;
+std::mutex GpuCacheMgr::global_mutex_;
+std::unordered_map<int64_t, std::pair<GpuCacheMgrPtr, MutexPtr>> GpuCacheMgr::instance_;
 
 namespace {
 constexpr int64_t G_BYTE = 1024 * 1024 * 1024;
 }
 
-GpuCacheMgr::GpuCacheMgr() {
+GpuCacheMgr::GpuCacheMgr(int64_t gpu_id) : gpu_id_(gpu_id) {
     // All config values have been checked in Config::ValidateConfig()
     server::Config& config = server::Config::GetInstance();
 
     int64_t gpu_cache_cap;
     config.GetGpuResourceConfigCacheCapacity(gpu_cache_cap);
     int64_t cap = gpu_cache_cap * G_BYTE;
-    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32);
+    std::string header = "[CACHE GPU" + std::to_string(gpu_id) + "]";
+    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32, header);
 
     float gpu_mem_threshold;
     config.GetGpuResourceConfigCacheThreshold(gpu_mem_threshold);
@@ -51,20 +52,6 @@ GpuCacheMgr::~GpuCacheMgr() {
     config.CancelCallBack(server::CONFIG_GPU_RESOURCE, server::CONFIG_GPU_RESOURCE_ENABLE, identity_);
 }
 
-GpuCacheMgr*
-GpuCacheMgr::GetInstance(uint64_t gpu_id) {
-    if (instance_.find(gpu_id) == instance_.end()) {
-        std::lock_guard<std::mutex> lock(mutex_);
-        if (instance_.find(gpu_id) == instance_.end()) {
-            instance_.insert(std::pair<uint64_t, GpuCacheMgrPtr>(gpu_id, std::make_shared<GpuCacheMgr>()));
-        }
-        return instance_[gpu_id].get();
-    } else {
-        std::lock_guard<std::mutex> lock(mutex_);
-        return instance_[gpu_id].get();
-    }
-}
-
 DataObjPtr
 GpuCacheMgr::GetIndex(const std::string& key) {
     DataObjPtr obj = GetItem(key);
@@ -78,10 +65,33 @@ GpuCacheMgr::InsertItem(const std::string& key, const milvus::cache::DataObjPtr&
     }
 }
 
+GpuCacheMgrPtr
+GpuCacheMgr::GetInstance(int64_t gpu_id) {
+    if (instance_.find(gpu_id) == instance_.end()) {
+        std::lock_guard<std::mutex> lock(global_mutex_);
+        if (instance_.find(gpu_id) == instance_.end()) {
+            instance_[gpu_id] = std::make_pair(std::make_shared<GpuCacheMgr>(gpu_id), std::make_shared<std::mutex>());
+        }
+    }
+    return instance_[gpu_id].first;
+}
+
+MutexPtr
+GpuCacheMgr::GetInstanceMutex(int64_t gpu_id) {
+    if (instance_.find(gpu_id) == instance_.end()) {
+        std::lock_guard<std::mutex> lock(global_mutex_);
+        if (instance_.find(gpu_id) == instance_.end()) {
+            instance_[gpu_id] = std::make_pair(std::make_shared<GpuCacheMgr>(gpu_id), std::make_shared<std::mutex>());
+        }
+    }
+    return instance_[gpu_id].second;
+}
+
 void
 GpuCacheMgr::OnGpuCacheCapacityChanged(int64_t capacity) {
     for (auto& iter : instance_) {
-        iter.second->SetCapacity(capacity * G_BYTE);
+        std::lock_guard<std::mutex> lock(*(iter.second.second));
+        iter.second.first->SetCapacity(capacity * G_BYTE);
     }
 }
 

core/src/cache/GpuCacheMgr.h

@@ -10,8 +10,10 @@
 // or implied. See the License for the specific language governing permissions and limitations under the License.
 
 #include <memory>
+#include <mutex>
 #include <string>
 #include <unordered_map>
+#include <utility>
 
 #include "cache/CacheMgr.h"
 #include "cache/DataObj.h"
@@ -23,31 +25,36 @@ namespace cache {
 #ifdef MILVUS_GPU_VERSION
 class GpuCacheMgr;
 using GpuCacheMgrPtr = std::shared_ptr<GpuCacheMgr>;
+using MutexPtr = std::shared_ptr<std::mutex>;
 
 class GpuCacheMgr : public CacheMgr<DataObjPtr>, public server::GpuResourceConfigHandler {
  public:
-    GpuCacheMgr();
+    explicit GpuCacheMgr(int64_t gpu_id);
 
     ~GpuCacheMgr();
 
-    static GpuCacheMgr*
-    GetInstance(uint64_t gpu_id);
-
     DataObjPtr
     GetIndex(const std::string& key);
 
     void
     InsertItem(const std::string& key, const DataObjPtr& data);
 
+    static GpuCacheMgrPtr
+    GetInstance(int64_t gpu_id);
+
+    static MutexPtr
+    GetInstanceMutex(int64_t gpu_id);
+
  protected:
     void
     OnGpuCacheCapacityChanged(int64_t capacity) override;
 
  private:
     bool gpu_enable_ = true;
+    int64_t gpu_id_;
+    static std::mutex global_mutex_;
+    static std::unordered_map<int64_t, std::pair<GpuCacheMgrPtr, MutexPtr>> instance_;
     std::string identity_;
-    static std::mutex mutex_;
-    static std::unordered_map<uint64_t, GpuCacheMgrPtr> instance_;
 };
 #endif
 

core/src/config/Config.h

@@ -86,7 +86,7 @@ static const char* CONFIG_CACHE = "cache_config";
 static const char* CONFIG_CACHE_CPU_CACHE_CAPACITY = "cpu_cache_capacity";
 static const char* CONFIG_CACHE_CPU_CACHE_CAPACITY_DEFAULT = "4";
 static const char* CONFIG_CACHE_CPU_CACHE_THRESHOLD = "cpu_cache_threshold";
-static const char* CONFIG_CACHE_CPU_CACHE_THRESHOLD_DEFAULT = "0.85";
+static const char* CONFIG_CACHE_CPU_CACHE_THRESHOLD_DEFAULT = "0.7";
 static const char* CONFIG_CACHE_INSERT_BUFFER_SIZE = "insert_buffer_size";
 static const char* CONFIG_CACHE_INSERT_BUFFER_SIZE_DEFAULT = "1";
 static const char* CONFIG_CACHE_CACHE_INSERT_DATA = "cache_insert_data";
@@ -123,7 +123,7 @@ static const char* CONFIG_GPU_RESOURCE_ENABLE_DEFAULT = "false";
 static const char* CONFIG_GPU_RESOURCE_CACHE_CAPACITY = "cache_capacity";
 static const char* CONFIG_GPU_RESOURCE_CACHE_CAPACITY_DEFAULT = "1";
 static const char* CONFIG_GPU_RESOURCE_CACHE_THRESHOLD = "cache_threshold";
-static const char* CONFIG_GPU_RESOURCE_CACHE_THRESHOLD_DEFAULT = "0.85";
+static const char* CONFIG_GPU_RESOURCE_CACHE_THRESHOLD_DEFAULT = "0.7";
 static const char* CONFIG_GPU_RESOURCE_DELIMITER = ",";
 static const char* CONFIG_GPU_RESOURCE_SEARCH_RESOURCES = "search_resources";
 static const char* CONFIG_GPU_RESOURCE_SEARCH_RESOURCES_DEFAULT = "gpu0";

core/src/db/engine/ExecutionEngineImpl.cpp

@@ -544,9 +544,15 @@ ExecutionEngineImpl::CopyToGpu(uint64_t device_id, bool hybrid) {
         }
 
         try {
+            /* Index data is copied to GPU first, then added into GPU cache.
+             * We MUST add a lock here to avoid more than one INDEX are copied to one GPU card at same time,
+             * which will potentially cause GPU out of memory.
+             */
+            std::lock_guard<std::mutex> lock(*(cache::GpuCacheMgr::GetInstanceMutex(device_id)));
+            ENGINE_LOG_DEBUG << "CPU to GPU" << device_id << " start";
             index_ = knowhere::cloner::CopyCpuToGpu(index_, device_id, knowhere::Config());
-            ENGINE_LOG_DEBUG << "CPU to GPU" << device_id;
             GpuCache(device_id);
+            ENGINE_LOG_DEBUG << "CPU to GPU" << device_id << " finished";
         } catch (std::exception& e) {
             ENGINE_LOG_ERROR << e.what();
             return Status(DB_ERROR, e.what());

core/src/scheduler/task/SearchTask.cpp

@@ -148,7 +148,7 @@ XSearchTask::Load(LoadType type, uint8_t device_id) {
                 hybrid = true;
             }
             stat = index_engine_->CopyToGpu(device_id, hybrid);
-            type_str = "CPU2GPU:" + std::to_string(device_id);
+            type_str = "CPU2GPU" + std::to_string(device_id);
         } else if (type == LoadType::GPU2CPU) {
             stat = index_engine_->CopyToCpu();
             type_str = "GPU2CPU";