perf: support system transaction

Signed-off-by: Jian Zhang <[email protected]>

benchmarks/micro-benchmarks/InsertUpdateBench.cpp

@@ -22,7 +22,7 @@ static void BenchUpdateInsert(benchmark::State& state) {
   std::filesystem::remove_all(dirPath);
   std::filesystem::create_directories(dirPath);
 
-  StoreOption* option = CreateStoreOption("/tmp/InsertUpdateBench");
+  StoreOption* option = CreateStoreOption("/tmp/leanstore/InsertUpdateBench");
   option->mCreateFromScratch = true;
   option->mWorkerThreads = 4;
   auto sLeanStore = std::make_unique<leanstore::LeanStore>(option);

benchmarks/ycsb/YcsbLeanStore.hpp

@@ -41,6 +41,7 @@ class YcsbLeanStore : public YcsbExecutor {
     auto dataDirStr = FLAGS_ycsb_data_dir + std::string("/leanstore");
     StoreOption* option = CreateStoreOption(dataDirStr.c_str());
     option->mCreateFromScratch = createFromScratch;
+    option->mEnableEagerGc = true;
     option->mWorkerThreads = FLAGS_ycsb_threads;
     option->mBufferPoolSize = FLAGS_ycsb_mem_kb * 1024;
     option->mEnableMetrics = true;

benchmarks/ycsb/ycsb-config.flags

@@ -3,8 +3,8 @@
 --ycsb_val_size=200
 --ycsb_record_count=100000
 --ycsb_mem_kb=1048576
---ycsb_run_for_seconds=5
---ycsb_target=basickv
---ycsb_workload=c
---ycsb_threads=2
---ycsb_cmd=load
+--ycsb_run_for_seconds=600
+--ycsb_target=transactionkv
+--ycsb_workload=b
+--ycsb_threads=8
+--ycsb_cmd=run

include/leanstore/LeanStore.hpp

@@ -73,10 +73,13 @@ class LeanStore {
   //! NOTE: Ownerd by LeanStore instance, should be destroyed together with it
   cr::CRManager* mCRManager;
 
-  //! The global timestamp oracle, used to generate start and commit timestamps
-  //! for all transactions in the store. Start from a positive number, 0
-  //! indicates invalid timestamp
-  std::atomic<uint64_t> mTimestampOracle = 1;
+  //! The global timestamp oracle for user transactions. Used to generate start and commit
+  //! timestamps for user transactions. Start from a positive number, 0 indicates invalid timestamp
+  std::atomic<uint64_t> mUsrTso = 1;
+
+  //! The global timestamp oracle for system transactions. Used to generate timestamps for system
+  //! transactions. Start from a positive number, 0 indicates invalid timestamp
+  std::atomic<uint64_t> mSysTso = 1;
 
   //! The metrics manager
   std::unique_ptr<leanstore::telemetry::MetricsManager> mMetricsManager;
@@ -118,13 +121,21 @@ class LeanStore {
   //! Unregister a TransactionKV
   void DropTransactionKV(const std::string& name);
 
-  uint64_t GetTs() {
-    return mTimestampOracle.load();
+  uint64_t GetUsrTxTs() {
+    return mUsrTso.load();
+  }
+
+  uint64_t GetSysTxTs() {
+    return mSysTso.load();
   }
 
   //! Alloc a new timestamp from the timestamp oracle
-  uint64_t AllocTs() {
-    return mTimestampOracle.fetch_add(1);
+  uint64_t AllocUsrTxTs() {
+    return mUsrTso.fetch_add(1);
+  }
+
+  uint64_t AllocSysTxTs() {
+    return mSysTso.fetch_add(1);
   }
 
   //! Execute a custom user function on a worker thread.

include/leanstore/btree/core/BTreeGeneric.hpp

@@ -10,6 +10,8 @@
 #include "leanstore/sync/HybridLatch.hpp"
 #include "leanstore/utils/Log.hpp"
 
+#include <rapidjson/document.h>
+
 #include <atomic>
 #include <limits>
 
@@ -55,9 +57,9 @@ class BTreeGeneric : public leanstore::storage::BufferManagedTree {
 
   //! Try to merge the current node with its left or right sibling, reclaim the merged left or right
   //! sibling if successful.
-  bool TryMergeMayJump(BufferFrame& toMerge, bool swizzleSibling = true);
+  bool TryMergeMayJump(TXID sysTxId, BufferFrame& toMerge, bool swizzleSibling = true);
 
-  void TrySplitMayJump(BufferFrame& toSplit, int16_t pos = -1);
+  void TrySplitMayJump(TXID sysTxId, BufferFrame& toSplit, int16_t pos = -1);
 
   XMergeReturnCode XMerge(GuardedBufferFrame<BTreeNode>& guardedParent,
                           GuardedBufferFrame<BTreeNode>& guardedChild,
@@ -141,7 +143,7 @@ class BTreeGeneric : public leanstore::storage::BufferManagedTree {
   //!              |     |
   //!           newLeft toSplit
   ///
-  void splitRootMayJump(GuardedBufferFrame<BTreeNode>& guardedParent,
+  void splitRootMayJump(TXID sysTxId, GuardedBufferFrame<BTreeNode>& guardedParent,
                         GuardedBufferFrame<BTreeNode>& guardedChild,
                         const BTreeNode::SeparatorInfo& sepInfo);
 
@@ -152,7 +154,7 @@ class BTreeGeneric : public leanstore::storage::BufferManagedTree {
   //!   |            |   |
   //! toSplit   newLeft toSplit
   ///
-  void splitNonRootMayJump(GuardedBufferFrame<BTreeNode>& guardedParent,
+  void splitNonRootMayJump(TXID sysTxId, GuardedBufferFrame<BTreeNode>& guardedParent,
                            GuardedBufferFrame<BTreeNode>& guardedChild,
                            const BTreeNode::SeparatorInfo& sepInfo,
                            uint16_t spaceNeededForSeparator);
@@ -198,15 +200,15 @@ class BTreeGeneric : public leanstore::storage::BufferManagedTree {
     xGuardedMeta.Reclaim();
   }
 
-  // static void ToJson(BTreeGeneric& btree, rapidjson::Document* resultDoc);
+  static void ToJson(BTreeGeneric& btree, rapidjson::Document* resultDoc);
 
 private:
   static void freeBTreeNodesRecursive(BTreeGeneric& btree,
                                       GuardedBufferFrame<BTreeNode>& guardedNode);
 
-  // static void toJsonRecursive(BTreeGeneric& btree, GuardedBufferFrame<BTreeNode>& guardedNode,
-  //                             rapidjson::Value* resultObj,
-  //                             rapidjson::Value::AllocatorType& allocator);
+  static void toJsonRecursive(BTreeGeneric& btree, GuardedBufferFrame<BTreeNode>& guardedNode,
+                              rapidjson::Value* resultObj,
+                              rapidjson::Value::AllocatorType& allocator);
 
   static ParentSwipHandler findParentMayJump(BTreeGeneric& btree, BufferFrame& bfToFind) {
     return FindParent<true>(btree, bfToFind);

include/leanstore/btree/core/PessimisticExclusiveIterator.hpp

@@ -56,22 +56,23 @@ class PessimisticExclusiveIterator : public PessimisticIterator {
   virtual void InsertToCurrentNode(Slice key, Slice val) {
     LS_DCHECK(KeyInCurrentNode(key));
     LS_DCHECK(HasEnoughSpaceFor(key.size(), val.size()));
-    LS_DCHECK(mSlotId != -1);
+    LS_DCHECK(Valid());
     mSlotId = mGuardedLeaf->InsertDoNotCopyPayload(key, val.size(), mSlotId);
     std::memcpy(mGuardedLeaf->ValData(mSlotId), val.data(), val.size());
   }
 
   void SplitForKey(Slice key) {
+    auto sysTxId = mBTree.mStore->AllocSysTxTs();
     while (true) {
       JUMPMU_TRY() {
-        if (mSlotId == -1 || !KeyInCurrentNode(key)) {
+        if (!Valid() || !KeyInCurrentNode(key)) {
           mBTree.FindLeafCanJump(key, mGuardedLeaf);
         }
         BufferFrame* bf = mGuardedLeaf.mBf;
         mGuardedLeaf.unlock();
-        mSlotId = -1;
+        SetToInvalid();
 
-        mBTree.TrySplitMayJump(*bf);
+        mBTree.TrySplitMayJump(sysTxId, *bf);
         COUNTERS_BLOCK() {
           WorkerCounters::MyCounters().dt_split[mBTree.mTreeId]++;
         }
@@ -162,7 +163,8 @@ class PessimisticExclusiveIterator : public PessimisticIterator {
 
       mSlotId = -1;
       JUMPMU_TRY() {
-        mBTree.TrySplitMayJump(*mGuardedLeaf.mBf, splitSlot);
+        TXID sysTxId = mBTree.mStore->AllocSysTxTs();
+        mBTree.TrySplitMayJump(sysTxId, *mGuardedLeaf.mBf, splitSlot);
 
         LS_DLOG("[Contention Split] succeed, pageId={}, contention pct={}, split "
                 "slot={}",
@@ -201,7 +203,8 @@ class PessimisticExclusiveIterator : public PessimisticIterator {
       mGuardedLeaf.unlock();
       mSlotId = -1;
       JUMPMU_TRY() {
-        mBTree.TryMergeMayJump(*mGuardedLeaf.mBf);
+        TXID sysTxId = mBTree.mStore->AllocSysTxTs();
+        mBTree.TryMergeMayJump(sysTxId, *mGuardedLeaf.mBf);
       }
       JUMPMU_CATCH() {
         LS_DLOG("TryMergeIfNeeded failed, pageId={}", mGuardedLeaf.mBf->mHeader.mPageId);

include/leanstore/btree/core/PessimisticIterator.hpp

@@ -12,6 +12,8 @@
 
 #include <functional>
 
+#include <sys/syscall.h>
+
 namespace leanstore::storage::btree {
 
 using LeafCallback = std::function<void(GuardedBufferFrame<BTreeNode>& guardedLeaf)>;
@@ -401,7 +403,8 @@ inline void PessimisticIterator::Next() {
     if (mGuardedLeaf->mNumSlots == 0) {
       SetCleanUpCallback([&, toMerge = mGuardedLeaf.mBf]() {
         JUMPMU_TRY() {
-          mBTree.TryMergeMayJump(*toMerge, true);
+          TXID sysTxId = mBTree.mStore->AllocSysTxTs();
+          mBTree.TryMergeMayJump(sysTxId, *toMerge, true);
         }
         JUMPMU_CATCH() {
         }

include/leanstore/buffer-manager/AsyncWriteBuffer.hpp

@@ -25,7 +25,7 @@ namespace leanstore::storage {
 //!  writeBuffer.SubmitAll();
 //!  writeBuffer.WaitAll();
 //!  writeBuffer.IterateFlushedBfs([](BufferFrame& flushedBf, uint64_t
-//!  flushedGsn) {
+//!  flushedPsn) {
 //!    // do something with flushedBf
 //!  }, numFlushedBfs);
 ///
@@ -72,7 +72,7 @@ class AsyncWriteBuffer {
     return mAIo.GetNumRequests();
   }
 
-  void IterateFlushedBfs(std::function<void(BufferFrame& flushedBf, uint64_t flushedGsn)> callback,
+  void IterateFlushedBfs(std::function<void(BufferFrame& flushedBf, uint64_t flushedPsn)> callback,
                          uint64_t numFlushedBfs);
 
 private:

include/leanstore/buffer-manager/BufferFrame.hpp

@@ -9,6 +9,7 @@
 #include "leanstore/utils/UserThread.hpp"
 
 #include <atomic>
+#include <cstdint>
 #include <limits>
 
 namespace leanstore::storage {
@@ -66,14 +67,14 @@ class BufferFrameHeader {
   //! ID of page resides in this buffer frame.
   PID mPageId = std::numeric_limits<PID>::max();
 
-  //! ID of the last worker who has modified the containing page.  For remote
-  //! flush avoidance (RFA), see "Rethinking Logging, Checkpoints, and Recovery
-  //! for High-Performance Storage Engines, SIGMOD 2020" for details.
+  //! ID of the last worker who has modified the containing page. For remote flush avoidance (RFA),
+  //! see "Rethinking Logging, Checkpoints, and Recovery for High-Performance Storage Engines,
+  //! SIGMOD 2020" for details.
   WORKERID mLastWriterWorker = std::numeric_limits<uint8_t>::max();
 
-  //! The flushed global sequence number of the containing page. Initialized to
-  //! the page GSN when loaded from disk.
-  uint64_t mFlushedGsn = 0;
+  //! The flushed page sequence number of the containing page. Initialized when the containing page
+  //! is loaded from disk.
+  uint64_t mFlushedPsn = 0;
 
   //! Whether the containing page is being written back to disk.
   std::atomic<bool> mIsBeingWrittenBack = false;
@@ -97,7 +98,7 @@ class BufferFrameHeader {
 
     mPageId = std::numeric_limits<PID>::max();
     mLastWriterWorker = std::numeric_limits<uint8_t>::max();
-    mFlushedGsn = 0;
+    mFlushedPsn = 0;
     mIsBeingWrittenBack.store(false, std::memory_order_release);
     mContentionStats.Reset();
     mCrc = 0;
@@ -130,9 +131,15 @@ class Page {
   //! Short for "global sequence number", increased when a page is modified.
   //! It's used to check whether the page has been read or written by
   //! transactions in other workers.
-  //! A page is "dirty" when mPage.mGSN > mHeader.mFlushedGsn.
   uint64_t mGSN = 0;
 
+  //! Short for "system transaction id", increased when a system transaction modifies the page.
+  uint64_t mSysTxId = 0;
+
+  //! Short for "page sequence number", increased when a page is modified by any user or system
+  //! transaction. A page is "dirty" when mPage.mPsn > mHeader.mFlushedPsn.
+  uint64_t mPsn = 0;
+
   //! The btree ID it belongs to.
   TREEID mBTreeId = std::numeric_limits<TREEID>::max();
 
@@ -176,7 +183,7 @@ class BufferFrame {
   }
 
   bool IsDirty() const {
-    return mPage.mGSN != mHeader.mFlushedGsn;
+    return mPage.mPsn != mHeader.mFlushedPsn;
   }
 
   bool IsFree() const {
@@ -192,9 +199,11 @@ class BufferFrame {
     LS_DCHECK(mHeader.mState == State::kFree);
     mHeader.mPageId = pageId;
     mHeader.mState = State::kHot;
-    mHeader.mFlushedGsn = 0;
+    mHeader.mFlushedPsn = 0;
 
     mPage.mGSN = 0;
+    mPage.mSysTxId = 0;
+    mPage.mPsn = 0;
   }
 
   // Pre: bf is exclusively locked