CRC / Copy offloading to Intel Data Streaming Accelerator

cachelib/CMakeLists.txt

@@ -44,6 +44,9 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
 option(BUILD_TESTS "If enabled, compile the tests." ON)
 
+option(BUILD_WITH_DTO
+  "If enabled, build with the DTO library for Intel DSA offload support." OFF)
+
 
 set(BIN_INSTALL_DIR bin CACHE STRING
     "The subdirectory where binaries should be installed")
@@ -312,6 +315,56 @@ endfunction()
 add_thrift_file(OBJECT_CACHE_PERSISTENCE
                 object_cache/persistence/persistent_data.thrift json)
 
+if (BUILD_WITH_DTO)
+  # DSA offload runtime (DTO). Resolved here so every cachelib library can
+  # use it: DTO::dto is linked PUBLIC into cachelib_common (see
+  # common/CMakeLists.txt), the base library of every other cachelib target,
+  # which propagates the headers, the link dependency, and the
+  # CACHELIB_BUILD_WITH_DTO compile definition to the whole tree and to
+  # consumers of the installed package.
+  #
+  # The DSA checksum offload code requires a DTO with the caller-allocated
+  # async submit/poll API and the raw-CRC32C convention, which upstream
+  # intel/DTO does not have — so when no installed DTO cmake package is
+  # found, fetch and build the pinned revision in-tree rather than linking
+  # whatever libdto happens to be on the system.
+  set(CACHELIB_DTO_GIT_REPOSITORY "https://github.com/byrnedj/DTO.git"
+      CACHE STRING "Git repository for DTO when no installed copy is found")
+  set(CACHELIB_DTO_GIT_TAG "991f4f9084095af78e2d7ca2dcb9a8c6069f8904"
+      CACHE STRING "DTO revision to fetch when no installed copy is found")
+
+  find_package(DTO CONFIG QUIET)
+  if (DTO_FOUND)
+    message(STATUS "Using installed DTO package: ${DTO_DIR}")
+  else()
+    if (CMAKE_VERSION VERSION_LESS 3.14)
+      message(FATAL_ERROR
+        "BUILD_WITH_DTO needs an installed DTO cmake package or "
+        "CMake >= 3.14 to fetch one. Install DTO from "
+        "${CACHELIB_DTO_GIT_REPOSITORY} at ${CACHELIB_DTO_GIT_TAG}, "
+        "or upgrade CMake.")
+    endif()
+    # DTO links against libaccel-config and libnuma; check for them here so
+    # a missing system package fails at configure time with a clear message
+    # instead of at link time inside the fetched project.
+    find_library(ACCEL_CONFIG_LIBRARY accel-config)
+    find_library(NUMA_LIBRARY numa)
+    if (NOT ACCEL_CONFIG_LIBRARY OR NOT NUMA_LIBRARY)
+      message(FATAL_ERROR
+        "BUILD_WITH_DTO requires the libaccel-config and libnuma "
+        "development packages (e.g. apt install libaccel-config-dev "
+        "libnuma-dev)")
+    endif()
+    message(STATUS "DTO not installed; fetching "
+      "${CACHELIB_DTO_GIT_REPOSITORY} @ ${CACHELIB_DTO_GIT_TAG}")
+    include(FetchContent)
+    FetchContent_Declare(dto
+      GIT_REPOSITORY "${CACHELIB_DTO_GIT_REPOSITORY}"
+      GIT_TAG "${CACHELIB_DTO_GIT_TAG}")
+    FetchContent_MakeAvailable(dto)
+  endif()
+endif()
+
 add_subdirectory (common)
 add_subdirectory (shm)
 add_subdirectory (navy)

cachelib/allocator/Cache.cpp

@@ -426,6 +426,12 @@ void CacheBase::updateGlobalCacheStats(const std::string& statPrefix) const {
   if (stats.nvmCacheEnabled) {
     counters_.updateDelta(statPrefix + "nvm.alloc_attempts",
                           stats.numNvmAllocAttempts);
+    counters_.updateDelta(statPrefix + "nvm.copy_out_offloaded",
+                          stats.numNvmCopyOutOffloaded);
+    counters_.updateDelta(statPrefix + "nvm.copy_out_offloaded_bytes",
+                          stats.numNvmCopyOutOffloadedBytes);
+    counters_.updateDelta(statPrefix + "nvm.copy_out_fallbacks",
+                          stats.numNvmCopyOutFallbacks);
     counters_.updateDelta(statPrefix + "nvm.destructor_alloc",
                           stats.numNvmAllocForItemDestructor);
     counters_.updateDelta(statPrefix + "nvm.destructor_alloc_errors",

cachelib/allocator/CacheStats.cpp

@@ -54,10 +54,10 @@ struct SizeVerify {};
 void Stats::populateGlobalCacheStats(GlobalCacheStats& ret) const {
 #ifndef SKIP_SIZE_VERIFY
 #ifdef __GLIBCXX__
-#define EXPECTED_SIZE 16944
+#define EXPECTED_SIZE 17200
 #endif
 #ifdef _LIBCPP_VERSION
-#define EXPECTED_SIZE 16944
+#define EXPECTED_SIZE 17200
 #endif
   SizeVerify<sizeof(Stats)> a = SizeVerify<EXPECTED_SIZE>{};
   std::ignore = a;
@@ -106,6 +106,9 @@ void Stats::populateGlobalCacheStats(GlobalCacheStats& ret) const {
   ret.numInsertOrReplaceInserted = numInsertOrReplaceInserted.get();
   ret.numInsertOrReplaceReplaced = numInsertOrReplaceReplaced.get();
   ret.numNvmAllocAttempts = numNvmAllocAttempts.get();
+  ret.numNvmCopyOutOffloaded = numNvmCopyOutOffloaded.get();
+  ret.numNvmCopyOutOffloadedBytes = numNvmCopyOutOffloadedBytes.get();
+  ret.numNvmCopyOutFallbacks = numNvmCopyOutFallbacks.get();
   ret.numNvmAllocForItemDestructor = numNvmAllocForItemDestructor.get();
   ret.numNvmItemDestructorAllocErrors = numNvmItemDestructorAllocErrors.get();
 

cachelib/allocator/CacheStats.h

@@ -456,6 +456,11 @@ struct GlobalCacheStats {
   // attempts made from nvm cache to allocate an item for promotion
   uint64_t numNvmAllocAttempts{0};
 
+  // flash-hit copy-outs performed by Intel DSA, bytes moved, CPU fallbacks
+  uint64_t numNvmCopyOutOffloaded{0};
+  uint64_t numNvmCopyOutOffloadedBytes{0};
+  uint64_t numNvmCopyOutFallbacks{0};
+
   // attempts made from nvm cache to allocate an item for its destructor
   uint64_t numNvmAllocForItemDestructor{0};
   // heap allocate errors for item destructor

cachelib/allocator/CacheStatsInternal.h

@@ -141,6 +141,12 @@ struct Stats {
   // attempts made from nvm cache to allocate an item for promotion
   TLCounter numNvmAllocAttempts{0};
 
+  // flash-hit copy-outs (Navy buffer -> DRAM item) performed by Intel DSA,
+  // the bytes they moved, and offload attempts that fell back to the CPU
+  TLCounter numNvmCopyOutOffloaded{0};
+  TLCounter numNvmCopyOutOffloadedBytes{0};
+  TLCounter numNvmCopyOutFallbacks{0};
+
   // attempts made from nvm cache to allocate an item for its destructor
   TLCounter numNvmAllocForItemDestructor{0};
   // heap allocate errors for item destructor

cachelib/allocator/nvmcache/NavyConfig.cpp

@@ -323,6 +323,10 @@ std::map<std::string, std::string> EnginesConfig::serialize() const {
       folly::to<std::string>(blockCache().getNumInMemBuffers());
   configMap["navyConfig::blockCacheDataChecksum"] =
       blockCache().getDataChecksum() ? "true" : "false";
+  configMap["navyConfig::blockCacheChecksumOffload"] =
+      blockCache().getChecksumOffload() ? "true" : "false";
+  configMap["navyConfig::blockCacheChecksumOffloadMinSize"] =
+      folly::to<std::string>(blockCache().getChecksumOffloadMinSize());
   configMap["navyConfig::blockCacheSegmentedFifoSegmentRatio"] =
       folly::join(",", blockCache().getSFifoSegmentRatio());
 
@@ -335,6 +339,8 @@ std::map<std::string, std::string> EnginesConfig::serialize() const {
       folly::to<std::string>(bigHash().getBucketBfSize());
   configMap["navyConfig::bigHashSmallItemMaxSize"] =
       folly::to<std::string>(bigHash().getSmallItemMaxSize());
+  configMap["navyConfig::bigHashChecksumOffload"] =
+      bigHash().getChecksumOffload() ? "true" : "false";
   return configMap;
 }
 

cachelib/allocator/nvmcache/NavyConfig.h

@@ -536,6 +536,36 @@ class BlockCacheConfig {
     return *this;
   }
 
+  // Offload data checksumming (fused with the value copy on the write path)
+  // to Intel DSA via the DTO library. Requires data checksum to be enabled
+  // and CacheLib built with BUILD_WITH_DTO. @minSize is the minimum value
+  // size to use the offloaded path; smaller values are checksummed in
+  // software, since submitting and polling a descriptor costs about what the
+  // CPU needs to CRC 16-32 KiB (16 KiB measured as the break-even).
+  BlockCacheConfig& setChecksumOffload(bool checksumOffload,
+                                       uint32_t minSize = 16384) noexcept {
+    checksumOffload_ = checksumOffload;
+    checksumOffloadMinSize_ = minSize;
+    return *this;
+  }
+
+  // Separate minimum size for offloading checksum VERIFICATION on the read,
+  // reclaim and reinsertion paths; 0 = same as the write-path gate. Verifying
+  // has no CPU work to overlap with the accelerator, so its break-even is
+  // higher than the fused write.
+  BlockCacheConfig& setChecksumOffloadReadMinSize(uint32_t minSize) noexcept {
+    checksumOffloadReadMinSize_ = minSize;
+    return *this;
+  }
+
+  // Cache-control hint on the fused write-path copy (default true): steer the
+  // value bytes toward the CPU cache. Turn off when the next reader of the
+  // region buffer is the device (directFlush / flush copy offload).
+  BlockCacheConfig& setChecksumOffloadCacheControl(bool enable) noexcept {
+    checksumOffloadCacheControl_ = enable;
+    return *this;
+  }
+
   BlockCacheConfig& setPreciseRemove(bool preciseRemove) noexcept {
     preciseRemove_ = preciseRemove;
     return *this;
@@ -561,6 +591,17 @@ class BlockCacheConfig {
     return *this;
   }
 
+  // When directFlush is off: copy the region buffer into the flush write
+  // buffer with one Intel DSA Memory Move instead of memcpy. Requires
+  // CacheLib built with BUILD_WITH_DTO; falls back to memcpy if DSA is
+  // unusable. The write buffers are pooled, so after the first flush per
+  // buffer their pages are resident; a work queue with block-on-fault covers
+  // the first touch.
+  BlockCacheConfig& setFlushCopyOffload(bool enable) noexcept {
+    flushCopyOffload_ = enable;
+    return *this;
+  }
+
   BlockCacheConfig& setAllocatorCount(uint32_t numAllocators) noexcept {
     allocatorsPerPriority_ = {numAllocators};
     return *this;
@@ -617,13 +658,26 @@ class BlockCacheConfig {
 
   bool getDataChecksum() const { return dataChecksum_; }
 
+  bool getChecksumOffload() const { return checksumOffload_; }
+
+  uint32_t getChecksumOffloadMinSize() const { return checksumOffloadMinSize_; }
+
+  uint32_t getChecksumOffloadReadMinSize() const {
+    return checksumOffloadReadMinSize_;
+  }
+
+  bool getChecksumOffloadCacheControl() const {
+    return checksumOffloadCacheControl_;
+  }
+
   uint64_t getSize() const { return size_; }
 
   bool isRegionManagerFlushAsync() const { return regionManagerFlushAsync_; }
 
   bool isRecoverEvictionPolicy() const { return recoverEvictionPolicy_; }
 
   bool isDirectFlush() const { return directFlush_; }
+  bool isFlushCopyOffload() const { return flushCopyOffload_; }
 
   bool isCombinedEntryBlockEnabled() const { return useCombinedEntryBlock_; }
 
@@ -660,6 +714,12 @@ class BlockCacheConfig {
   uint32_t regionSize_{16 * 1024 * 1024};
   // Whether enabling data checksum for Navy BlockCache.
   bool dataChecksum_{true};
+  // Whether to offload data checksumming to Intel DSA (fused with the value
+  // copy on the write path), and the minimum value size to do so.
+  bool checksumOffload_{false};
+  uint32_t checksumOffloadMinSize_{16384};
+  uint32_t checksumOffloadReadMinSize_{0};
+  bool checksumOffloadCacheControl_{true};
   // Whether to remove an item by checking the key (true) or only the hash value
   // (false).
   bool preciseRemove_{false};
@@ -677,6 +737,9 @@ class BlockCacheConfig {
   // Whether to write region buffer directly without intermediate copy.
   bool directFlush_{false};
 
+  // Whether to do the flush copy (when not directFlush) on Intel DSA.
+  bool flushCopyOffload_{false};
+
   // Whether to use Combined entry block (For index entries and small sized
   // items).
   // Only FixedSizeIndex will support this and it doesn't work with
@@ -739,6 +802,14 @@ class BigHashConfig {
     return *this;
   }
 
+  // Offload bucket checksumming to Intel DSA via the DTO library. Most
+  // beneficial with large buckets (16KB+). Requires CacheLib built with
+  // BUILD_WITH_DTO.
+  BigHashConfig& setChecksumOffload(bool checksumOffload) noexcept {
+    checksumOffload_ = checksumOffload;
+    return *this;
+  }
+
   bool isBloomFilterEnabled() const { return bucketBfSize_ > 0; }
 
   unsigned int getSizePct() const { return sizePct_; }
@@ -751,6 +822,8 @@ class BigHashConfig {
 
   uint8_t getNumMutexesPower() const { return numMutexesPower_; }
 
+  bool getChecksumOffload() const { return checksumOffload_; }
+
  private:
   // Percentage of how much of the device out of all is given to BigHash
   // engine in Navy, e.g. 50.
@@ -766,6 +839,8 @@ class BigHashConfig {
   uint64_t smallItemMaxSize_{};
   // numMutexes = 1 << numMutexesPower_.
   uint8_t numMutexesPower_{14};
+  // Whether to offload bucket checksumming to Intel DSA.
+  bool checksumOffload_{false};
 };
 
 // Config for a pair of small,large engines.

cachelib/allocator/nvmcache/NavySetup.cpp

@@ -140,6 +140,8 @@ uint64_t setupBigHash(const navy::BigHashConfig& bigHashConfig,
   // Set number of mutexes from config
   bigHash->setNumMutexesPower(bigHashConfig.getNumMutexesPower());
 
+  bigHash->setChecksumOffload(bigHashConfig.getChecksumOffload());
+
   proto.setBigHash(std::move(bigHash), bigHashConfig.getSmallItemMaxSize());
 
   if (bigHashCacheOffset <= bigHashStartOffsetLimit) {
@@ -196,6 +198,12 @@ uint64_t setupBlockCache(const navy::BlockCacheConfig& blockCacheConfig,
   auto blockCache = cachelib::navy::createBlockCacheProto();
   blockCache->setLayout(blockCacheOffset, blockCacheSize, regionSize);
   blockCache->setChecksum(blockCacheConfig.getDataChecksum());
+  blockCache->setChecksumOffload(blockCacheConfig.getChecksumOffload(),
+                                 blockCacheConfig.getChecksumOffloadMinSize());
+  blockCache->setChecksumOffloadReadMinSize(
+      blockCacheConfig.getChecksumOffloadReadMinSize());
+  blockCache->setChecksumOffloadCacheControl(
+      blockCacheConfig.getChecksumOffloadCacheControl());
 
   // set eviction policy
   auto segmentRatio = blockCacheConfig.getSFifoSegmentRatio();
@@ -214,6 +222,7 @@ uint64_t setupBlockCache(const navy::BlockCacheConfig& blockCacheConfig,
   blockCache->setRecoverEvictionPolicy(
       blockCacheConfig.isRecoverEvictionPolicy());
   blockCache->setDirectFlush(blockCacheConfig.isDirectFlush());
+  blockCache->setFlushCopyOffload(blockCacheConfig.isFlushCopyOffload());
   blockCache->setUseCombinedEntryBlock(
       blockCacheConfig.isCombinedEntryBlockEnabled());
   blockCache->setNumAllocatorsPerPriority(