use weak wait, add options to tweak for caching mode

qdp_project/src/Benchmark.cpp

@@ -32,11 +32,13 @@ constexpr uint32_t ITERATION_COUNT = 5;
 #ifdef MODE_PREFETCH
 constexpr uint32_t GROUP_COUNT = 16;
 constexpr size_t CHUNK_SIZE_B = WL_SIZE_B / GROUP_COUNT;
-constexpr uint32_t TC_SCANA = 1;
+constexpr uint32_t TC_SCANA = 2;
 constexpr uint32_t TC_SCANB = 1;
-constexpr uint32_t TC_AGGRJ = 4;
+constexpr uint32_t TC_AGGRJ = 8;
 constexpr bool PERFORM_CACHING = true;
+constexpr bool YIELD_ON_CACHE_MISS = false;
 constexpr bool DATA_IN_HBM = false;
+constexpr bool AGGRJ_ITERATIVE = true;
 constexpr char MODE_STRING[] = "prefetch";
 #endif
 #ifdef MODE_DRAM
@@ -46,7 +48,9 @@ constexpr uint32_t TC_SCANA = 4;
 constexpr uint32_t TC_SCANB = 0;
 constexpr uint32_t TC_AGGRJ = 2;
 constexpr bool PERFORM_CACHING = false;
+constexpr bool YIELD_ON_CACHE_MISS = false;
 constexpr bool DATA_IN_HBM = false;
+constexpr bool AGGRJ_ITERATIVE = false;
 constexpr char MODE_STRING[] = "dram";
 #endif
 #ifdef MODE_HBM
@@ -56,7 +60,9 @@ constexpr uint32_t TC_SCANA = 4;
 constexpr uint32_t TC_SCANB = 0;
 constexpr uint32_t TC_AGGRJ = 2;
 constexpr bool PERFORM_CACHING = false;
+constexpr bool YIELD_ON_CACHE_MISS = false;
 constexpr bool DATA_IN_HBM = true;
+constexpr bool AGGRJ_ITERATIVE = false;
 constexpr char MODE_STRING[] = "hbm";
 #endif
 
@@ -92,6 +98,8 @@ constexpr size_t TIME_STAMP_END = 2;
 // THREAD_TIMING_[TYPE][TID][ITERATION][STAMP] = TIMEPOINT
 std::array<std::vector<std::vector<std::array<std::chrono::steady_clock::time_point, 3>>>, 3> THREAD_TIMING_;
 
+std::array<uint32_t, GROUP_COUNT * TC_AGGRJ> CACHE_HITS_;
+
 std::vector<std::barrier<NopStruct>*> BARRIERS_;
 std::shared_future<void> LAUNCH_;
 
@@ -114,7 +122,16 @@ template<size_t TC>
 inline uint16_t* get_mask(uint16_t* base, const size_t chunk_index, const size_t tid) {
     size_t offset = chunk_index * CHUNK_SIZE_ELEMENTS + tid * (CHUNK_SIZE_ELEMENTS / TC);
     return base + (offset / 16);
+}
+
+double process_cache_hitrate() {
+    double hr = 0.0;
 
+    for (const uint32_t& e : CACHE_HITS_) {
+        hr += e;
+    }
+
+    return hr / (double)(TC_AGGRJ * GROUP_COUNT * RUN_COUNT);
 }
 
 void process_timings(
@@ -226,8 +243,10 @@ void scan_a(size_t gid, size_t tid) {
 }
 
 void aggr_j(size_t gid, size_t tid) {
-    THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid  * gid].clear();
-    THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid  * gid].resize(1);
+    CACHE_HITS_[gid * tid] = 0;
+
+    THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid * gid].clear();
+    THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid * gid].resize(1);
 
     LAUNCH_.wait();
 
@@ -249,11 +268,15 @@ void aggr_j(size_t gid, size_t tid) {
 
         if constexpr (PERFORM_CACHING) {
             data = CACHE_.Access(reinterpret_cast<uint8_t *>(chunk_ptr), CHUNK_SIZE_B / TC_AGGRJ);
-            data->WaitOnCompletion();
+            data->WaitOnCompletion(dsacache::WAIT_WEAK);
             data_ptr = reinterpret_cast<uint64_t*>(data->GetDataLocation());
+
             if (data_ptr == nullptr) {
-                std::cerr << "[x] Cache Miss!" << std::endl;
-                exit(-1);
+                data_ptr = chunk_ptr;
+                if constexpr (YIELD_ON_CACHE_MISS) sched_yield();
+            }
+            else {
+                CACHE_HITS_[gid * tid]++;
             }
         }
         else {
@@ -269,6 +292,10 @@ void aggr_j(size_t gid, size_t tid) {
     BARRIERS_[gid]->arrive_and_drop();
 
     aggregation::happly(DATA_DST_ + (tid * GROUP_COUNT + gid), aggregator);
+
+    if constexpr (AGGRJ_ITERATIVE) {
+        if (++tid < TC_AGGRJ) aggr_j(gid, tid);
+    }
 }
 
 int main() {
@@ -283,7 +310,7 @@ int main() {
     const std::string ofname = "results/qdp-xeonmax-simpleq-" + std::string(MODE_STRING) + "-tca" + std::to_string(TC_SCANA) + "-tcb" + std::to_string(TC_SCANB) + "-tcj" + std::to_string(TC_AGGRJ) + "-tmul" + std::to_string(GROUP_COUNT) + "-wl" + std::to_string(WL_SIZE_B) + "-cs" + std::to_string(CHUNK_SIZE_B) + ".csv";
     std::ofstream fout(ofname);
 
-    fout << "run;rt-ns;rt-s;result[0];scana-run;scana-wait;scanb-run;scanb-wait;aggrj-run;aggrj-wait;" << std::endl;
+    fout << "run;rt-ns;rt-s;result[0];scana-run;scana-wait;scanb-run;scanb-wait;aggrj-run;aggrj-wait;cache-hr;" << std::endl;
 
     if constexpr (DATA_IN_HBM) {
         DATA_A_ = (uint64_t*) numa_alloc_onnode(WL_SIZE_B, cache_node);
@@ -324,8 +351,13 @@ int main() {
                 copy_pool.emplace_back(scan_b, gid, tid);
             }
 
-            for(uint32_t tid = 0; tid < TC_AGGRJ; ++tid) {
-                agg_pool.emplace_back(aggr_j, gid,  tid);
+            if constexpr (AGGRJ_ITERATIVE) {
+                agg_pool.emplace_back(aggr_j, gid, 0);
+            }
+            else {
+                for(uint32_t tid = 0; tid < TC_AGGRJ; ++tid) {
+                    agg_pool.emplace_back(aggr_j, gid,  tid);
+                }
             }
         }
 
@@ -354,6 +386,7 @@ int main() {
                 << nanos << ";" << seconds << ";"
                 << std::hex << DATA_DST_[0] << std::dec << ";"
                 << scana_run << ";" << scana_wait << ";" << scanb_run << ";" << scanb_wait << ";" << aggrj_run << ";" << aggrj_wait << ";"
+                << process_cache_hitrate() << ";"
                 << std::endl;
         }