added stats for dsa submit (pmem#6)

Author: guptask

cachelib/allocator/CacheAllocator-inl.h

@@ -2904,7 +2904,7 @@ PoolStats CacheAllocator<CacheTrait>::getPoolStats(PoolId poolId) const {
     for (const ClassId cid : classIds) {
       uint64_t allocAttempts, evictionAttempts, allocFailures,
                fragmentationSize, classHits, chainedItemEvictions,
-               regularItemEvictions, numWritebacks = 0;
+               dsaEvictionSubmits, regularItemEvictions, numWritebacks = 0;
       MMContainerStat mmContainerStats;
       for (TierId tid = 0; tid < getNumTiers(); tid++) {
         allocAttempts += (*stats_.allocAttempts)[tid][poolId][cid].get();
@@ -2913,6 +2913,7 @@ PoolStats CacheAllocator<CacheTrait>::getPoolStats(PoolId poolId) const {
         fragmentationSize += (*stats_.fragmentationSize)[tid][poolId][cid].get();
         classHits += (*stats_.cacheHits)[tid][poolId][cid].get();
         chainedItemEvictions += (*stats_.chainedItemEvictions)[tid][poolId][cid].get();
+        dsaEvictionSubmits += (*stats_.dsaEvictionSubmits)[tid][poolId][cid].get();
         regularItemEvictions += (*stats_.regularItemEvictions)[tid][poolId][cid].get();
         numWritebacks += (*stats_.numWritebacks)[tid][poolId][cid].get();
         mmContainerStats += getMMContainerStat(tid, poolId, cid);
@@ -2928,6 +2929,7 @@ PoolStats CacheAllocator<CacheTrait>::getPoolStats(PoolId poolId) const {
             fragmentationSize,
             classHits,
             chainedItemEvictions,
+            dsaEvictionSubmits,
             regularItemEvictions,
             numWritebacks,
             mmContainerStats}});
@@ -2982,6 +2984,7 @@ PoolStats CacheAllocator<CacheTrait>::getPoolStats(TierId tid, PoolId poolId) co
             (*stats_.fragmentationSize)[tid][poolId][cid].get(),
             classHits,
             (*stats_.chainedItemEvictions)[tid][poolId][cid].get(),
+            (*stats_.dsaEvictionSubmits)[tid][poolId][cid].get(),
             (*stats_.regularItemEvictions)[tid][poolId][cid].get(),
             (*stats_.numWritebacks)[tid][poolId][cid].get(),
             getMMContainerStat(tid, poolId, cid)}});

cachelib/allocator/CacheAllocator.h

@@ -2238,6 +2238,7 @@ auto& mmContainer = getMMContainer(tid, pid, cid);
       dml::data_view dstView = dml::make_view(
           reinterpret_cast<uint8_t*>(newItemHandles[index]->getMemory()),
           newItemHandles[index]->getSize());
+      (*stats_.dsaEvictionSubmits)[tid][pid][cid].inc();
       if (config_.dsaAsync) {
         dmlHandles.emplace_back(
             dml::submit<dml::hardware>(dml::mem_copy, srcView, dstView));

cachelib/allocator/CacheStats.cpp

@@ -29,6 +29,7 @@ void Stats::init() {
   fragmentationSize = std::make_unique<PerTierPerPoolClassAtomicCounters>();
   allocFailures = std::make_unique<PerTierPerPoolClassAtomicCounters>();
   chainedItemEvictions = std::make_unique<PerTierPerPoolClassAtomicCounters>();
+  dsaEvictionSubmits = std::make_unique<PerTierPerPoolClassAtomicCounters>();
   regularItemEvictions = std::make_unique<PerTierPerPoolClassAtomicCounters>();
   numWritebacks = std::make_unique<PerTierPerPoolClassAtomicCounters>();
   auto initToZero = [](auto& a) {
@@ -46,6 +47,7 @@ void Stats::init() {
   initToZero(*allocFailures);
   initToZero(*fragmentationSize);
   initToZero(*chainedItemEvictions);
+  initToZero(*dsaEvictionSubmits);
   initToZero(*regularItemEvictions);
   initToZero(*numWritebacks);
 
@@ -57,7 +59,7 @@ struct SizeVerify {};
 
 void Stats::populateGlobalCacheStats(GlobalCacheStats& ret) const {
 #ifndef SKIP_SIZE_VERIFY
-  SizeVerify<sizeof(Stats)> a = SizeVerify<16192>{};
+  SizeVerify<sizeof(Stats)> a = SizeVerify<16208>{};
   std::ignore = a;
 #endif
   ret.numCacheGets = numCacheGets.get();
@@ -151,6 +153,7 @@ void Stats::populateGlobalCacheStats(GlobalCacheStats& ret) const {
   ret.evictionAttempts = accum(*evictionAttempts);
   ret.allocFailures = accum(*allocFailures);
   auto chainedEvictions = accum(*chainedItemEvictions);
+  ret.dsaEvictionSubmits = accum(*dsaEvictionSubmits);
   auto regularEvictions = accum(*regularItemEvictions);
   for (TierId tid = 0; tid < chainedEvictions.size(); tid++) {
     ret.numEvictions.push_back(chainedEvictions[tid] + regularEvictions[tid]);
@@ -223,6 +226,7 @@ PoolStats& PoolStats::operator+=(const PoolStats& other) {
       d.numHits += s.numHits;
       d.numWritebacks += s.numWritebacks;
       d.chainedItemEvictions += s.chainedItemEvictions;
+      d.dsaEvictionSubmits += s.dsaEvictionSubmits;
       d.regularItemEvictions += s.regularItemEvictions;
     }
 

cachelib/allocator/CacheStats.h

@@ -122,6 +122,9 @@ struct CacheStat {
   // number of evictions from this class id that was of a chained item
   uint64_t chainedItemEvictions{0};
 
+  // number of items submitted via DSA for eviction
+  uint64_t dsaEvictionSubmits{0};
+
   // number of regular items that were evicted from this classId
   uint64_t regularItemEvictions{0};
 
@@ -462,6 +465,9 @@ struct GlobalCacheStats {
   // number of failures to allocate an item due to internal error
   std::vector<uint64_t> allocFailures;
 
+  // number of dsa evictions submits across all the pools in the cache.
+  std::vector<uint64_t> dsaEvictionSubmits;
+
   // number of evictions across all the pools in the cache.
   std::vector<uint64_t> numEvictions;
 

cachelib/allocator/CacheStatsInternal.h

@@ -230,6 +230,7 @@ struct Stats {
   std::unique_ptr<PerTierPerPoolClassAtomicCounters> allocFailures{};
   std::unique_ptr<PerTierPerPoolClassAtomicCounters> fragmentationSize{};
   std::unique_ptr<PerTierPerPoolClassAtomicCounters> chainedItemEvictions{};
+  std::unique_ptr<PerTierPerPoolClassAtomicCounters> dsaEvictionSubmits{};
   std::unique_ptr<PerTierPerPoolClassAtomicCounters> regularItemEvictions{};
   std::unique_ptr<PerTierPerPoolClassAtomicCounters> numWritebacks{};
 

cachelib/cachebench/cache/Cache-inl.h

@@ -677,6 +677,7 @@ Stats Cache<Allocator>::getStats() const {
             cacheStats.promotionStats.runCount;
 
   ret.evictAttempts = cacheStats.evictionAttempts;
+  ret.dsaEvictionSubmits = cacheStats.dsaEvictionSubmits;
   ret.allocAttempts = cacheStats.allocAttempts;
   ret.allocFailures = cacheStats.allocFailures;
 

cachelib/cachebench/cache/CacheStats.h

@@ -54,6 +54,7 @@ struct Stats {
   BackgroundPromotionStats backgndPromoStats;
   ReaperStats reaperStats;
 
+  std::vector<uint64_t> dsaEvictionSubmits;
   std::vector<uint64_t> numEvictions;
   std::vector<uint64_t> numWritebacks;
   std::vector<uint64_t> numCacheHits;
@@ -167,8 +168,8 @@ struct Stats {
             << std::endl;
     }
     for (TierId tid = 0; tid < nTiers; tid++) {
-        out << folly::sformat("Tier {} Evictions : {:,} Writebacks: {:,} Success: {:.2f}%",
-                tid, numEvictions[tid], numWritebacks[tid],
+        out << folly::sformat("Tier {} Evictions : {:,} DSA submits: {:,} Writebacks: {:,} Success: {:.2f}%",
+                tid, numEvictions[tid], dsaEvictionSubmits[tid], numWritebacks[tid],
                 invertPctFn(numEvictions[tid] - numWritebacks[tid], numEvictions[tid])) << std::endl;
     }
     auto foreachAC = [&](auto &map, auto cb) {