c

CMakeLists.txt

@@ -10,7 +10,7 @@ option(BUILD_EXAMPLE "Build examples." OFF)
 option(LEGACY_PAPI "Use Vikram's PAPI stuff to do performance monitering." OFF)
 option(HIGH_LEVEL_PAPI "Use PAPI high-level monitoring" OFF)
 option(VTUNE "Use VTUNE to do performance monitering." OFF)
-option(AGGR "Use aggregation hashtable (histogram)" OFF)
+option(AGGR "Use aggregation hashtable (histogram)" ON)
 option(BQUEUE "Enable bqueue tests" OFF)
 option(XORWOW "Xorwow" OFF)
 option(BQ_ZIPFIAN "Enable global zipfian distribution generation" ON)

include/hashtables/array_kht.hpp

@@ -89,14 +89,14 @@ class ArrayHashTable : public BaseHashTable {
 #endif
 
     uint64_t hash = this->hash((const char *)data);
-    size_t idx = hash;
-
+    size_t idx = hash & (this->capacity - 1);  // modulo
+    // size_t idx = fastrange32(hash, this->capacity);  // modulo
+
     KVQ *elem = const_cast<KVQ *>(reinterpret_cast<const KVQ *>(data));
 
     KV *curr = &this->hashtable[idx];
     if (curr->is_empty()) {
-      PLOGV.printf("inserting key %llu at idx %llu", elem->key, idx);
-      bool cas_res = curr->insert(elem);
+      curr->insert(elem);
     } else {
       curr->update(elem);
     }
@@ -124,6 +124,8 @@ class ArrayHashTable : public BaseHashTable {
       KVQ q;
       q.idx = this->hash((const char *)&data.key);
       q.value = data.value;
+
+      PLOG_INFO.printf("key insertion: %d", data.key);
       __insert_one(&q, collector);
     }
   }

include/tests/KmerTest.hpp

@@ -8,31 +8,30 @@
 #include "input_reader/fastq.hpp"
 #include "types.hpp"
 
-
 namespace kmercounter {
 
 class KmerTest {
-private:
-    Kmer packkmer(const char* s, int k);
-    uint64_t radix_partition( RadixContext* context, 
-                          BufferedPartition** local_partitions,
-                          input_reader::FastqReader& reader,
-                           int id, int k);
- public:
-  void count_kmer(Shard *sh, const Configuration &config, BaseHashTable *ht,
-                  std::barrier<VoidFn> *barrier);
-
-    void count_kmer_radix_partition_global(Shard* sh, const Configuration& config,
-                            std::barrier<VoidFn>* barrier,
-                            RadixContext* context, BaseHashTable* ht);
+ private:
+  Kmer packkmer(const char* s, int k);
+  uint64_t radix_partition(RadixContext* context,
+                           BufferedPartition** local_partitions,
+                           input_reader::FastqReader& reader, int id, int k);
 
-    void count_kmer_baseline(Shard* sh, const Configuration& config,
-                                std::barrier<VoidFn>* barrier,
-                                BaseHashTable* ht);
+ public:
+  void count_kmer(Shard* sh, const Configuration& config, BaseHashTable* ht,
+                  std::barrier<VoidFn>* barrier);
+
+  void count_kmer_radix_partition_global(Shard* sh, const Configuration& config,
+                                         std::barrier<VoidFn>* barrier,
+                                         RadixContext* context,
+                                         BaseHashTable* ht);
+
+  void count_kmer_baseline(Shard* sh, const Configuration& config,
+                           std::barrier<VoidFn>* barrier, BaseHashTable* ht);
+  void count_kmer_partition(Shard* sh, const Configuration& config,
+                            std::barrier<VoidFn>* barrier);
 };
 
 }  // namespace kmercounter
 
-
-
 #endif  // TESTS_KMERTEST_HPP

include/types.hpp

@@ -84,6 +84,7 @@ typedef enum {
   HASHJOIN = 13,
   FASTQ_WITH_INSERT_RADIX = 14,
   BASELINE = 15,
+  FASTQ_INSERT_PARTITION = 16,
 } run_mode_t;
 
 // XXX: If you add/modify a mode, update the `ht_type_strings` in
@@ -360,6 +361,10 @@ struct Configuration {
   uint32_t data_size;
   uint32_t workload_size;
 
+  //kmer partition
+  double max_fill_factor;
+  uint32_t num_ht;
+
   // hashtable configuration
   // different hashtable types
   uint32_t ht_type;

jerry-benchmark.sh

@@ -29,23 +29,46 @@ function command_baseline()
 {
     sudo ./build/dramhit \
     --mode 15 \
-    --ht-type 3 \
+    --ht-type 4 \
     --numa-split 1 \
     --num-threads 1 \
     --ht-size $1 \
     --datasize $2 \
     --workload $3 > logs/baseline_d${2}_ht${1}.log
 }
 
+function command_partition() 
+{
+    sudo ./build/dramhit \
+    --mode 16 \
+    --ht-type 3 \
+    --numa-split 1 \
+    --num-threads 32 \
+    --in-file $FASTA_FILE \
+    --ht-size $1 \
+    --num-ht $2 \
+    --max-fill-factor $3 > logs/partition_sz${1}_num${2}_fill${3}.log
+}
+
+function bench_partition() 
+{
+    #ht=300
+    fill=0.6
+    for (( i=10; i<=10; i+=1 )); do
+        num=$((1 << i))
+        size=$(($HT_SIZE/$num/32))
+        command_partition $size $num $fill
+    done    
+}
 
 function bench_baseline() 
 {
     #ht=300
     #d=100
-    w=100000
-    for (( i=5; i<=25; i+=1 )); do
+    w=1000000
+    for (( i=5; i<=15; i+=1 )); do
         d=$((1<<i))
-        ht=$((1*d))
+        ht=$((2*d))
         command_baseline $ht $d $w
     done    
 }
@@ -75,14 +98,22 @@ function bench_radix()
 
 function debug() 
 {
-    sudo gdb --args ./build/dramhit --mode 15 --ht-type 3 --numa-split 1 --num-threads 1 --ht-size 300 --datasize 100 --workload 10000 
+    sudo gdb --args ./build/dramhit \
+    --mode 15 \
+    --ht-type 4 \
+    --numa-split 1 \
+    --num-threads 1 \
+    --ht-size 64 \
+    --datasize 32 \
+    --workload 1 
 }
 
 
 function bench() {
     #bench_radix
     #command_regular
     bench_baseline
+    # bench_partition
 }
 
 function build() {

jerry/baseline.c

@@ -0,0 +1,21 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <x86intrin.h> 
+#include <stdint.h>
+
+uint64_t rdtsc() {
+    uint32_t low, high;
+    asm volatile ("mfence\n\trdtsc" : "=a" (low), "=d" (high));
+    return ((uint64_t)high << 32) | low;
+}
+
+int main(int argc, char** argv) 
+{
+    int workload = atoi(argv[1]); 
+
+    unsigned long long a = __rdtsc();
+   // for(int i=0; i < workload; i++);
+    unsigned long long b = __rdtsc(); 
+
+    printf("Cycle: %llu Workload: %d \n", b-a, workload);
+}