chore: schedule chains (#3819)

Use intrusive queue that allows batching of scheduling calls instead of handling each call separately.
This optimizations improves latency and throughput by 3-5%
In addition, we expose batching statistics in info transaction block.

Signed-off-by: Roman Gershman <roman@dragonflydb.io>

src/server/engine_shard.cc

@@ -210,15 +210,20 @@ ShardId Shard(string_view v, ShardId shard_num) {
 }
 
 EngineShard::Stats& EngineShard::Stats::operator+=(const EngineShard::Stats& o) {
-  static_assert(sizeof(Stats) == 48);
+  static_assert(sizeof(Stats) == 64);
 
-  defrag_attempt_total += o.defrag_attempt_total;
+#define ADD(x) x += o.x
-  defrag_realloc_total += o.defrag_realloc_total;
-  defrag_task_invocation_total += o.defrag_task_invocation_total;
-  poll_execution_total += o.poll_execution_total;
-  tx_ooo_total += o.tx_ooo_total;
-  tx_optimistic_total += o.tx_optimistic_total;
 
+  ADD(defrag_attempt_total);
+  ADD(defrag_realloc_total);
+  ADD(defrag_task_invocation_total);
+  ADD(poll_execution_total);
+  ADD(tx_ooo_total);
+  ADD(tx_optimistic_total);
+  ADD(tx_batch_schedule_calls_total);
+  ADD(tx_batch_scheduled_items_total);
+
+#undef ADD
   return *this;
 }
 

src/server/engine_shard.h

@@ -38,6 +38,12 @@ class EngineShard {
     uint64_t tx_optimistic_total = 0;
     uint64_t tx_ooo_total = 0;
 
+    // Number of ScheduleBatchInShard calls.
+    uint64_t tx_batch_schedule_calls_total = 0;
+
+    // Number of transactions scheduled via ScheduleBatchInShard.
+    uint64_t tx_batch_scheduled_items_total = 0;
+
     Stats& operator+=(const Stats&);
   };
 

src/server/main_service.cc

@@ -979,6 +979,8 @@ void Service::Init(util::AcceptServer* acceptor, std::vector<facade::Listener*>
     server_family_.GetDflyCmd()->BreakStalledFlowsInShard();
     server_family_.UpdateMemoryGlobalStats();
   });
+  Transaction::Init(shard_num);
+
   SetOomDenyRatioOnAllThreads(absl::GetFlag(FLAGS_oom_deny_ratio));
   SetRssOomDenyRatioOnAllThreads(absl::GetFlag(FLAGS_rss_oom_deny_ratio));
 
@@ -1010,6 +1012,7 @@ void Service::Shutdown() {
   shard_set->PreShutdown();
   namespaces.Clear();
   shard_set->Shutdown();
+  Transaction::Shutdown();
 
   pp_.Await([](ProactorBase* pb) { ServerState::tlocal()->Destroy(); });
 

src/server/server_family.cc

@@ -2440,7 +2440,8 @@ void ServerFamily::Info(CmdArgList args, ConnectionContext* cntx) {
     append("tx_normal_total", m.coordinator_stats.tx_normal_cnt);
     append("tx_inline_runs_total", m.coordinator_stats.tx_inline_runs);
     append("tx_schedule_cancel_total", m.coordinator_stats.tx_schedule_cancel_cnt);
-
+    append("tx_batch_scheduled_items_total", m.shard_stats.tx_batch_scheduled_items_total);
+    append("tx_batch_schedule_calls_total", m.shard_stats.tx_batch_schedule_calls_total);
     append("tx_with_freq", absl::StrJoin(m.coordinator_stats.tx_width_freq_arr, ","));
     append("tx_queue_len", m.tx_queue_len);
 

src/server/transaction.cc

@@ -6,6 +6,8 @@
 
 #include <absl/strings/match.h>
 
+#include <new>
+
 #include "base/flags.h"
 #include "base/logging.h"
 #include "facade/op_status.h"
@@ -86,12 +88,32 @@ uint16_t trans_id(const Transaction* ptr) {
 struct ScheduleContext {
   Transaction* trans;
   bool optimistic_execution = false;
+
+  std::atomic<ScheduleContext*> next{nullptr};
+
   std::atomic_uint32_t fail_cnt{0};
 
   ScheduleContext(Transaction* t, bool optimistic) : trans(t), optimistic_execution(optimistic) {
   }
 };
 
+constexpr size_t kAvoidFalseSharingSize = 64;
+struct ScheduleQ {
+  alignas(kAvoidFalseSharingSize) base::MPSCIntrusiveQueue<ScheduleContext> queue;
+  alignas(kAvoidFalseSharingSize) atomic_bool armed{false};
+};
+
+void MPSC_intrusive_store_next(ScheduleContext* dest, ScheduleContext* next_node) {
+  dest->next.store(next_node, std::memory_order_relaxed);
+}
+
+ScheduleContext* MPSC_intrusive_load_next(const ScheduleContext& src) {
+  return src.next.load(std::memory_order_acquire);
+}
+
+// of shard_num arity.
+ScheduleQ* schedule_queues = nullptr;
+
 }  // namespace
 
 bool Transaction::BatonBarrier::IsClaimed() const {
@@ -139,6 +161,17 @@ Transaction::Guard::~Guard() {
   tx->Refurbish();
 }
 
+void Transaction::Init(unsigned num_shards) {
+  DCHECK(schedule_queues == nullptr);
+  schedule_queues = new ScheduleQ[num_shards];
+}
+
+void Transaction::Shutdown() {
+  DCHECK(schedule_queues);
+  delete[] schedule_queues;
+  schedule_queues = nullptr;
+}
+
 Transaction::Transaction(const CommandId* cid) : cid_{cid} {
   InitTxTime();
   string_view cmd_name(cid_->name());
@@ -685,11 +718,11 @@ void Transaction::ScheduleInternal() {
   // Try running immediately (during scheduling) if we're concluding and either:
   // - have a single shard, and thus never have to cancel scheduling due to reordering
   // - run as an idempotent command, meaning we can safely repeat the operation if scheduling fails
-  bool can_run_immediately = !IsGlobal() && (coordinator_state_ & COORD_CONCLUDING) &&
+  bool optimistic_exec = !IsGlobal() && (coordinator_state_ & COORD_CONCLUDING) &&
                          (unique_shard_cnt_ == 1 || (cid_->opt_mask() & CO::IDEMPOTENT));
 
   DVLOG(1) << "ScheduleInternal " << cid_->name() << " on " << unique_shard_cnt_ << " shards "
-           << " immediate run: " << can_run_immediately;
+           << " optimistic_execution: " << optimistic_exec;
 
   auto is_active = [this](uint32_t i) { return IsActive(i); };
 
@@ -711,13 +744,23 @@ void Transaction::ScheduleInternal() {
       // in the lower-level code. It's not really needed otherwise because we run inline.
 
       // single shard schedule operation can't fail
-      CHECK(ScheduleInShard(EngineShard::tlocal(), can_run_immediately));
+      CHECK(ScheduleInShard(EngineShard::tlocal(), optimistic_exec));
       run_barrier_.Dec();
       break;
     }
 
-    ScheduleContext schedule_ctx{this, can_run_immediately};
+    ScheduleContext schedule_ctx{this, optimistic_exec};
 
+    if (unique_shard_cnt_ == 1) {
+      // Single shard optimization. Note: we could apply the same optimization
+      // to multi-shard transactions as well by creating a vector of ScheduleContext.
+      schedule_queues[unique_shard_id_].queue.Push(&schedule_ctx);
+      bool current_val = false;
+      if (schedule_queues[unique_shard_id_].armed.compare_exchange_strong(current_val, true,
+                                                                          memory_order_acq_rel)) {
+        shard_set->Add(unique_shard_id_, &Transaction::ScheduleBatchInShard);
+      }
+    } else {
       auto cb = [&schedule_ctx]() {
         if (!schedule_ctx.trans->ScheduleInShard(EngineShard::tlocal(),
                                                  schedule_ctx.optimistic_execution)) {
@@ -734,6 +777,7 @@ void Transaction::ScheduleInternal() {
         return absl::StrCat("unique_shard_cnt_: ", unique_shard_cnt_,
                             " run_barrier_cnt: ", run_barrier_.DEBUG_Count(), "\n");
       });
+    }
     run_barrier_.Wait();
 
     if (schedule_ctx.fail_cnt.load(memory_order_relaxed) == 0) {
@@ -1115,6 +1159,45 @@ bool Transaction::ScheduleInShard(EngineShard* shard, bool execute_optimistic) {
   return true;
 }
 
+void Transaction::ScheduleBatchInShard() {
+  EngineShard* shard = EngineShard::tlocal();
+  auto& stats = shard->stats();
+  stats.tx_batch_schedule_calls_total++;
+
+  ShardId sid = shard->shard_id();
+  auto& sq = schedule_queues[sid];
+
+  for (unsigned j = 0;; ++j) {
+    // We pull the items from the queue in a loop until we reach the stop condition.
+    // TODO: we may have fairness problem here, where transactions being added up all the time
+    // and we never break from the loop. It is possible to break early but it's not trivial
+    // because we must ensure that there is another ScheduleBatchInShard callback in the queue.
+    // Can be checked with testing sq.armed is true when j == 1.
+    while (true) {
+      ScheduleContext* item = sq.queue.Pop();
+      if (!item)
+        break;
+
+      if (!item->trans->ScheduleInShard(shard, item->optimistic_execution)) {
+        item->fail_cnt.fetch_add(1, memory_order_relaxed);
+      }
+      item->trans->FinishHop();
+      stats.tx_batch_scheduled_items_total++;
+    };
+
+    // j==1 means we already signalled that we're done with the current batch.
+    if (j == 1)
+      break;
+
+    // We signal that we're done with the current batch but then we check if there are more
+    // transactions to fetch in the next iteration.
+    // We do this to avoid the situation where we have a data race, where
+    // a transaction is added to the queue, we've checked that sq.armed is true and skipped
+    // adding the callback that fetches the transaction.
+    sq.armed.store(false, memory_order_release);
+  }
+}
+
 bool Transaction::CancelShardCb(EngineShard* shard) {
   ShardId idx = SidToId(shard->shard_id());
   auto& sd = shard_data_[idx];

src/server/transaction.h

@@ -178,6 +178,9 @@ class Transaction {
     Transaction* tx;
   };
 
+  static void Init(unsigned num_shards);
+  static void Shutdown();
+
   explicit Transaction(const CommandId* cid);
 
   // Initialize transaction for squashing placed on a specific shard with a given parent tx
@@ -515,6 +518,9 @@ class Transaction {
   // subject to uncontended keys.
   bool ScheduleInShard(EngineShard* shard, bool execute_optimistic);
 
+  // Optimized extension of ScheduleInShard. Pulls several transactions queued for scheduling.
+  static void ScheduleBatchInShard();
+
   // Set ARMED flags, start run barrier and submit poll tasks. Doesn't wait for the run barrier
   void DispatchHop();