feat: Add eviction based on rss memory

Signed-off-by: Stepan Bagritsevich <[email protected]>

Author: BagritsevichStepan

src/server/dfly_main.cc

@@ -823,7 +823,12 @@ Usage: dragonfly [FLAGS]
   // export MIMALLOC_VERBOSE=1 to see the options before the override.
   mi_option_enable(mi_option_show_errors);
   mi_option_set(mi_option_max_warnings, 0);
+
   mi_option_enable(mi_option_purge_decommits);
+  DCHECK(mi_option_get(mi_option_reset_decommits) == 1);
+
+  mi_option_set(mi_option_purge_delay, 0);
+  DCHECK(!mi_option_get(mi_option_reset_delay));
 
   fb2::SetDefaultStackResource(&fb2::std_malloc_resource, kFiberDefaultStackSize);
 

src/server/engine_shard.cc

@@ -19,10 +19,12 @@ extern "C" {
 #include "server/search/doc_index.h"
 #include "server/server_state.h"
 #include "server/tiered_storage.h"
+#include "server/tiering/common.h"
 #include "server/transaction.h"
 #include "util/fibers/proactor_base.h"
 
 using namespace std;
+using namespace ::dfly::tiering::literals;
 
 ABSL_FLAG(float, mem_defrag_threshold, 0.7,
           "Minimum percentage of used memory relative to maxmemory cap before running "
@@ -65,6 +67,9 @@ ABSL_FLAG(double, eviction_memory_budget_threshold, 0.1,
           "Eviction starts when the free memory (including RSS memory) drops below "
           "eviction_memory_budget_threshold * max_memory_limit.");
 
+ABSL_FLAG(uint64_t, force_decommit_threshold, 8_MB,
+          "The threshold of memory to force decommit when memory is under pressure.");
+
 ABSL_DECLARE_FLAG(uint32_t, max_eviction_per_heartbeat);
 
 namespace dfly {
@@ -216,45 +221,6 @@ size_t CalculateHowManyBytesToEvictOnShard(size_t global_memory_limit, size_t gl
   return shard_budget < shard_memory_threshold ? (shard_memory_threshold - shard_budget) : 0;
 }
 
-/* Calculates the number of bytes to evict based on memory and rss memory usage. */
-size_t CalculateEvictionBytes() {
-  const size_t shards_count = shard_set->size();
-  const double eviction_memory_budget_threshold = GetFlag(FLAGS_eviction_memory_budget_threshold);
-
-  const size_t shard_memory_budget_threshold =
-      size_t(max_memory_limit * eviction_memory_budget_threshold) / shards_count;
-
-  const size_t global_used_memory = used_mem_current.load(memory_order_relaxed);
-
-  // Calculate how many bytes we need to evict on this shard
-  size_t goal_bytes = CalculateHowManyBytesToEvictOnShard(max_memory_limit, global_used_memory,
-                                                          shard_memory_budget_threshold);
-
-  // TODO: Eviction due to rss usage is not working well as it causes eviction
-  // of to many keys untill we finally see decrease in rss. We need to improve
-  // this logic before we enable it.
-  /*
-  const double rss_oom_deny_ratio = ServerState::tlocal()->rss_oom_deny_ratio;
-  // If rss_oom_deny_ratio is set, we should evict depending on rss memory too
-  if (rss_oom_deny_ratio > 0.0) {
-    const size_t max_rss_memory = size_t(rss_oom_deny_ratio * max_memory_limit);
-    // We start eviction when we have less than eviction_memory_budget_threshold * 100% of free rss
-    memory const size_t shard_rss_memory_budget_threshold =
-        size_t(max_rss_memory * eviction_memory_budget_threshold) / shards_count;
-
-    // Calculate how much rss memory is used by all shards
-    const size_t global_used_rss_memory = rss_mem_current.load(memory_order_relaxed);
-
-    // Try to evict more bytes if we are close to the rss memory limit
-    goal_bytes = std::max(
-        goal_bytes, CalculateHowManyBytesToEvictOnShard(max_rss_memory, global_used_rss_memory,
-                                                        shard_rss_memory_budget_threshold));
-  }
-  */
-
-  return goal_bytes;
-}
-
 }  // namespace
 
 __thread EngineShard* EngineShard::shard_ = nullptr;
@@ -359,6 +325,18 @@ bool EngineShard::DefragTaskState::CheckRequired() {
     return false;
   }
 
+  std::optional<ShardMemUsage> shard_mem_usage;
+
+  if (GetFlag(FLAGS_enable_heartbeat_eviction)) {
+    shard_mem_usage = ReadShardMemUsage(GetFlag(FLAGS_mem_defrag_page_utilization_threshold));
+    const static double eviction_waste_threshold = 0.05;
+    if (shard_mem_usage->wasted_mem >
+        (uint64_t(shard_mem_usage->commited * eviction_waste_threshold))) {
+      VLOG(1) << "memory issue found for memory " << shard_mem_usage.value();
+      return true;
+    }
+  }
+
   const std::size_t global_threshold = max_memory_limit * GetFlag(FLAGS_mem_defrag_threshold);
   if (global_threshold > rss_mem_current.load(memory_order_relaxed)) {
     return false;
@@ -373,11 +351,15 @@ bool EngineShard::DefragTaskState::CheckRequired() {
   }
   last_check_time = now;
 
-  ShardMemUsage usage = ReadShardMemUsage(GetFlag(FLAGS_mem_defrag_page_utilization_threshold));
+  if (!shard_mem_usage) {
+    shard_mem_usage = ReadShardMemUsage(GetFlag(FLAGS_mem_defrag_page_utilization_threshold));
+  }
+
+  DCHECK(shard_mem_usage.has_value());
 
   const double waste_threshold = GetFlag(FLAGS_mem_defrag_waste_threshold);
-  if (usage.wasted_mem > (uint64_t(usage.commited * waste_threshold))) {
-    VLOG(1) << "memory issue found for memory " << usage;
+  if (shard_mem_usage->wasted_mem > (uint64_t(shard_mem_usage->commited * waste_threshold))) {
+    VLOG(1) << "memory issue found for memory " << shard_mem_usage.value();
     return true;
   }
 
@@ -812,6 +794,7 @@ void EngineShard::RetireExpiredAndEvict() {
   DbContext db_cntx;
   db_cntx.time_now_ms = GetCurrentTimeMs();
 
+  size_t deleted_bytes = 0;
   size_t eviction_goal = GetFlag(FLAGS_enable_heartbeat_eviction) ? CalculateEvictionBytes() : 0;
 
   for (unsigned i = 0; i < db_slice.db_array_size(); ++i) {
@@ -823,6 +806,7 @@ void EngineShard::RetireExpiredAndEvict() {
     if (expt->size() > 0) {
       DbSlice::DeleteExpiredStats stats = db_slice.DeleteExpiredStep(db_cntx, ttl_delete_target);
 
+      deleted_bytes += stats.deleted_bytes;
       eviction_goal -= std::min(eviction_goal, size_t(stats.deleted_bytes));
       counter_[TTL_TRAVERSE].IncBy(stats.traversed);
       counter_[TTL_DELETE].IncBy(stats.deleted);
@@ -844,9 +828,71 @@ void EngineShard::RetireExpiredAndEvict() {
                << " bytes. Max eviction per heartbeat: "
                << GetFlag(FLAGS_max_eviction_per_heartbeat);
 
+      deleted_bytes += evicted_bytes;
       eviction_goal -= std::min(eviction_goal, evicted_bytes);
     }
   }
+
+  eviction_state_.deleted_bytes_before_rss_update += deleted_bytes;
+}
+
+size_t EngineShard::CalculateEvictionBytes() {
+  const size_t shards_count = shard_set->size();
+  const double eviction_memory_budget_threshold = GetFlag(FLAGS_eviction_memory_budget_threshold);
+
+  const size_t shard_memory_budget_threshold =
+      size_t(max_memory_limit * eviction_memory_budget_threshold) / shards_count;
+
+  const size_t global_used_memory = used_mem_current.load(memory_order_relaxed);
+
+  // Calculate how many bytes we need to evict on this shard
+  size_t goal_bytes = CalculateHowManyBytesToEvictOnShard(max_memory_limit, global_used_memory,
+                                                          shard_memory_budget_threshold);
+
+  LOG_IF_EVERY_N(INFO, goal_bytes > 0, 50)
+      << "Memory goal bytes: " << goal_bytes << ", used memory: " << global_used_memory
+      << ", memory limit: " << max_memory_limit;
+
+  // If rss_oom_deny_ratio is set, we should evict depending on rss memory too
+  const double rss_oom_deny_ratio = ServerState::tlocal()->rss_oom_deny_ratio;
+  if (rss_oom_deny_ratio > 0.0) {
+    const size_t max_rss_memory = size_t(rss_oom_deny_ratio * max_memory_limit);
+    /* We start eviction when we have less than eviction_memory_budget_threshold * 100% of free rss
+     * memory */
+    const size_t shard_rss_memory_budget_threshold =
+        size_t(max_rss_memory * eviction_memory_budget_threshold) / shards_count;
+
+    // Calculate how much rss memory is used by all shards
+    const size_t global_used_rss_memory = rss_mem_current.load(memory_order_relaxed);
+
+    auto& global_rss_memory_at_prev_eviction = eviction_state_.global_rss_memory_at_prev_eviction;
+    auto& deleted_bytes_before_rss_update = eviction_state_.deleted_bytes_before_rss_update;
+    if (global_used_rss_memory < eviction_state_.global_rss_memory_at_prev_eviction) {
+      deleted_bytes_before_rss_update -=
+          std::min(deleted_bytes_before_rss_update,
+                   (global_rss_memory_at_prev_eviction - global_used_rss_memory) / shards_count);
+    }
+
+    /* if (global_used_rss_memory == global_rss_memory_at_prev_eviction) {
+
+    } */
+
+    global_rss_memory_at_prev_eviction = global_used_rss_memory;
+
+    // Try to evict more bytes if we are close to the rss memory limit
+    const size_t rss_goal_bytes = CalculateHowManyBytesToEvictOnShard(
+        max_rss_memory, global_used_rss_memory - deleted_bytes_before_rss_update * shards_count,
+        shard_rss_memory_budget_threshold);
+
+    LOG_IF_EVERY_N(INFO, rss_goal_bytes > 0, 50)
+        << "Rss memory goal bytes: " << rss_goal_bytes
+        << ", rss used memory: " << global_used_rss_memory
+        << ", rss memory limit: " << max_rss_memory
+        << ", deleted_bytes_before_rss_update: " << deleted_bytes_before_rss_update;
+
+    goal_bytes = std::max(goal_bytes, rss_goal_bytes);
+  }
+  return goal_bytes;
 }
 
 void EngineShard::CacheStats() {

src/server/engine_shard.h

@@ -219,6 +219,11 @@ class EngineShard {
     void ResetScanState();
   };
 
+  struct EvictionTaskState {
+    size_t deleted_bytes_before_rss_update = 0;
+    size_t global_rss_memory_at_prev_eviction = 0;
+  };
+
   EngineShard(util::ProactorBase* pb, mi_heap_t* heap);
 
   // blocks the calling fiber.
@@ -230,6 +235,9 @@ class EngineShard {
   void Heartbeat();
   void RetireExpiredAndEvict();
 
+  /* Calculates the number of bytes to evict based on memory and rss memory usage. */
+  size_t CalculateEvictionBytes();
+
   void CacheStats();
 
   // We are running a task that checks whether we need to
@@ -270,6 +278,7 @@ class EngineShard {
   IntentLock shard_lock_;
 
   uint32_t defrag_task_ = 0;
+  EvictionTaskState eviction_state_;  // Used on eviction fiber
   util::fb2::Fiber fiber_heartbeat_periodic_;
   util::fb2::Done fiber_heartbeat_periodic_done_;