Vldb

cicada/Makefile

@@ -5,17 +5,18 @@ REL := ../common/
 include $(REL)Makefile
 CICADA_ALLSRC = $(CICADA_SRCS1) $(wildcard include/*.hh)
 
-KEY_SIZE=8
-VAL_SIZE=4
 ADD_ANALYSIS=1
 BACK_OFF=0
-INLINE_VERSION_PROMOTION=1
 INLINE_VERSION_OPT=1
+INLINE_VERSION_PROMOTION=1
 MASSTREE_USE=1
-PARTITION_TABLE=1
+PARTITION_TABLE=0
 REUSE_VERSION=1
 SINGLE_EXEC=0
+KEY_SIZE=8
+VAL_SIZE=4
 WRITE_LATEST_ONLY=0
+WORKER1_INSERT_DELAY_RPHASE=1
 
 CC = g++
 CFLAGS = -c -pipe -g -O3 -std=c++17 -march=native \
@@ -33,6 +34,7 @@ CFLAGS = -c -pipe -g -O3 -std=c++17 -march=native \
 				 -DREUSE_VERSION=$(REUSE_VERSION) \
 				 -DSINGLE_EXEC=$(SINGLE_EXEC) \
 				 -DWRITE_LATEST_ONLY=$(WRITE_LATEST_ONLY) \
+				 -DWORKER1_INSERT_DELAY_RPHASE=$(WORKER1_INSERT_DELAY_RPHASE) \
 
 # WRITE_LATEST_ONLY : for all write. 
 # original model adapt write-latest-only-rule for RMW

cicada/cicada.cc

@@ -19,6 +19,7 @@
 #include "../include/compiler.hh"
 #include "../include/cpu.hh"
 #include "../include/debug.hh"
+#include "../include/delay.hh"
 #include "../include/int64byte.hh"
 #include "../include/procedure.hh"
 #include "../include/random.hh"
@@ -98,10 +99,20 @@ void worker(size_t thid, char& ready, const bool& start, const bool& quit,
 
       if (trans.status_ == TransactionStatus::abort) {
         trans.earlyAbort();
+#if SINGLE_EXEC
+#else
+        trans.mainte();
+#endif
         goto RETRY;
       }
     }
 
+#if WORKER1_INSERT_DELAY_RPHASE
+    if (unlikely(thid == 1) && WORKER1_INSERT_DELAY_RPHASE_US != 0) {
+      clock_delay(WORKER1_INSERT_DELAY_RPHASE_US * CLOCKS_PER_US);
+    }
+#endif
+
     // read only tx doesn't collect read set and doesn't validate.
     // write phase execute logging and commit pending versions, but r-only tx
     // can skip it.
@@ -111,6 +122,10 @@ void worker(size_t thid, char& ready, const bool& start, const bool& quit,
       // Validation phase
       if (!trans.validation()) {
         trans.abort();
+#if SINGLE_EXEC
+#else
+        trans.mainte();
+#endif
         goto RETRY;
       }
 
@@ -156,6 +171,7 @@ int main(int argc, char* argv[]) try {
   for (unsigned int i = 0; i < THREAD_NUM; ++i) {
     res[0].addLocalAllResult(res[i]);
   }
+  ShowOptParameters();
   res[0].displayAllResult(CLOCKS_PER_US, EXTIME, THREAD_NUM);
   deleteDB();
 

cicada/include/common.hh

@@ -47,6 +47,10 @@ GLOBAL size_t IO_TIME_NS;               // nano second
 GLOBAL size_t GROUP_COMMIT_TIMEOUT_US;  // micro seconds
 GLOBAL size_t GC_INTER_US;              // garbage collection interval
 GLOBAL size_t PRE_RESERVE_VERSION;
+/**
+ * worker 1 insert delay in the end of read phase[us].
+ */
+GLOBAL size_t WORKER1_INSERT_DELAY_RPHASE_US; 
 GLOBAL size_t EXTIME;
 
 alignas(CACHE_LINE_SIZE) GLOBAL uint64_t_64byte *ThreadWtsArray;

cicada/include/transaction.hh

@@ -172,6 +172,9 @@ class TxExecutor {
 #if INLINE_VERSION_OPT
     if (tuple->getInlineVersionRight()) {
       tuple->inline_ver_.set(0, this->wts_.ts_);
+#if ADD_ANALYSIS
+      ++cres_->local_version_reuse_;
+#endif
       return &(tuple->inline_ver_);
     }
 #endif  // if INLINE_VERSION_OPT

cicada/include/util.hh

@@ -6,6 +6,8 @@
 
 extern void chkArg(const int argc, char *argv[]);
 
+extern void deleteDB();
+
 extern void displayDB();
 
 extern void displayMinRts();
@@ -18,12 +20,12 @@ extern void displayThreadRtsArray();
 
 extern void displaySLogSet();
 
-extern void partTableInit([[maybe_unused]] size_t thid, uint64_t initts, uint64_t start, uint64_t end);
+extern void leaderWork([[maybe_unused]] Backoff &backoff, [[maybe_unused]] std::vector<Result> &res);
 
-extern void partTableDelete([[maybe_unused]] size_t thid, uint64_t start, uint64_t end);
+extern void makeDB(uint64_t *initial_wts);
 
-extern void deleteDB();
+extern void partTableDelete([[maybe_unused]] size_t thid, uint64_t start, uint64_t end);
 
-extern void makeDB(uint64_t *initial_wts);
+extern void partTableInit([[maybe_unused]] size_t thid, uint64_t initts, uint64_t start, uint64_t end);
 
-extern void leaderWork([[maybe_unused]] Backoff &backoff, [[maybe_unused]] std::vector<Result> &res);
+extern void ShowOptParameters();

cicada/script/ycsb-xgci.sh

@@ -1,7 +1,7 @@
 #ycsb-xrs.sh(cicada)
-tuple=100000000
+tuple=1000000
 maxope=10
-rratioary=(50 95 100)
+rratioary=(50)
 rmw=off
 skew=0
 ycsb=on
@@ -10,10 +10,11 @@ group_commit=off
 cpu_mhz=2100
 io_time_ns=5
 group_commit_timeout_us=2
-gci=10
+gci=1
 prv=0
-extime=3
-epoch=3
+w1idr=10000 # worker 1 insert delay in the end of read phase[us]
+extime=5
+epoch=5
 
 host=`hostname`
 chris41="chris41.omni.hpcc.jp"
@@ -26,13 +27,13 @@ if  test $host = $dbs11 ; then
 fi
 
 cd ../
-make clean; make -j KEY_SIZE=8 VAL_SIZE=4
+make clean; make -j
 cd script/
 
 for rratio in "${rratioary[@]}"
 do
   if test $rratio = 50 ; then
-    result=result_cicada-part_ycsbA_tuple100m_gci10us_224th.dat
+    result=result_cicada_ycsbA_tuple1m_w1idr10ms_gci1us-1s.dat
   elif test $rratio = 95 ; then
     result=result_cicada-part_ycsbB_tuple100m_gci10us_224th.dat
   elif test $rratio = 100 ; then
@@ -44,11 +45,14 @@ do
   rm $result
 
   echo "#tuple num, avg-tps, min-tps, max-tps, avg-ar, min-ar, max-ar, avg-camiss, min-camiss, max-camiss, avg-make_procedure_latency_rate, avg-read_latency_rate, avg-write_latency_rate, avg-vali_latency_rate, avg-gc_latency_rate avg-other_work_latency_rate avg-maxrss version_malloc version_reuse" >> $result
-  echo "#sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $extime" >> $result
+  echo "#sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $w1idr $extime" >> $result
+  ../cicada.exe > exp.txt
+  tmpStr=`grep ShowOptParameters ./exp.txt`
+  echo "#$tmpStr" >> $result
 
-  for ((gci=10; gci<=10; gci*=10))
+  for ((gci=1; gci<=1000000; gci*=10))
   do
-    echo "sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $extime"
+    echo "sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $w1idr $extime"
     echo "Thread number $thread"
 
     sumTH=0
@@ -72,10 +76,10 @@ do
     for ((i = 1; i <= epoch; ++i))
     do
       if test $host = $dbs11 ; then
-        sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $extime > exp.txt
+        sudo perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $w1idr $extime > exp.txt
       fi
       if test $host = $chris41 ; then
-        perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $extime > exp.txt
+        perf stat -e cache-misses,cache-references -o ana.txt numactl --interleave=all ../cicada.exe $tuple $maxope $thread $rratio $rmw $skew $ycsb $wal $group_commit $cpu_mhz $io_time_ns $group_commit_timeout_us $gci $prv $w1idr $extime > exp.txt
       fi
 
       tmpTH=`grep throughput ./exp.txt | awk '{print $2}'`

cicada/transaction.cc

@@ -597,8 +597,9 @@ void TxExecutor::mainte() {
   }
 
   this->gcstop_ = rdtscp();
-  if (chkClkSpan(this->gcstart_, this->gcstop_, GC_INTER_US * CLOCKS_PER_US)) {
-    __atomic_store_n(&(GCFlag[thid_].obj_), 1, __ATOMIC_RELEASE);
+  if (chkClkSpan(this->gcstart_, this->gcstop_, GC_INTER_US * CLOCKS_PER_US)
+      && (loadAcquire(GCFlag[thid_].obj_) == 0)) {
+    storeRelease(GCFlag[thid_].obj_, 1);
     this->gcstart_ = this->gcstop_;
   }
   //-----

cicada/util.cc

@@ -25,11 +25,12 @@
 #include "include/time_stamp.hh"
 #include "include/transaction.hh"
 #include "include/tuple.hh"
+#include "include/util.hh"
 
 using std::cout, std::endl;
 
 void chkArg(const int argc, char *argv[]) {
-  if (argc != 16) {
+  if (argc != 17) {
     cout << "usage: ./cicada.exe TUPLE_NUM MAX_OPE THREAD_NUM RRATIO RMW "
             "ZIPF_SKEW YCSB WAL GROUP_COMMIT CPU_MHZ IO_TIME_NS "
             "GROUP_COMMIT_TIMEOUT_US GC_INTER_US PRE_RESERVE_VERSION EXTIME"
@@ -62,24 +63,9 @@ void chkArg(const int argc, char *argv[]) {
     cout << "GC_INTER_US: garbage collection interval [usec]" << endl;
     cout << "PRE_RESERVE_VERSION: pre-prepare memory for version generation."
          << endl;
+    cout << "WORKER1_INSERT_DELAY_RPHASE_US : worker 1 insert delay in the end of read phase[us]." << endl;
     cout << "EXTIME: execution time [sec]" << endl << endl;
-
-    cout << "Tuple " << sizeof(Tuple) << endl;
-    cout << "Version " << sizeof(Version) << endl;
-    cout << "TimeStamp " << sizeof(TimeStamp) << endl;
-    cout << "Procedure " << sizeof(Procedure) << endl;
-    cout << "KEY_SIZE : " << KEY_SIZE << endl;
-    cout << "VAL_SIZE : " << VAL_SIZE << endl;
-    cout << "CACHE_LINE_SIZE - ((17 + KEY_SIZE + sizeof(Version)) % "
-            "(CACHE_LINE_SIZE)) : "
-         << CACHE_LINE_SIZE -
-                ((17 + KEY_SIZE + sizeof(Version)) % (CACHE_LINE_SIZE))
-         << endl;
-    cout << "CACHE_LINE_SIZE - ((25 + VAL_SIZE) % (CACHE_LINE_SIZE)) : "
-         << CACHE_LINE_SIZE - ((25 + VAL_SIZE) % (CACHE_LINE_SIZE)) << endl;
-    cout << "MASSTREE_USE : " << MASSTREE_USE << endl;
-    cout << "Result:\t" << sizeof(Result) << endl;
-    cout << "uint64_t_64byte: " << sizeof(uint64_t_64byte) << endl;
+    ShowOptParameters();
     exit(0);
   }
 
@@ -108,7 +94,8 @@ void chkArg(const int argc, char *argv[]) {
   GROUP_COMMIT_TIMEOUT_US = atoi(argv[12]);
   GC_INTER_US = atoi(argv[13]);
   PRE_RESERVE_VERSION = atoi(argv[14]);
-  EXTIME = atoi(argv[15]);
+  WORKER1_INSERT_DELAY_RPHASE_US = atoi(argv[15]);
+  EXTIME = atoi(argv[16]);
 
   if (RRATIO > 100) {
     cout << "rratio [%%] must be 0 ~ 100)" << endl;
@@ -436,3 +423,22 @@ void leaderWork([[maybe_unused]] Backoff &backoff,
   leaderBackoffWork(backoff, res);
 #endif
 }
+
+void
+ShowOptParameters()
+{
+  cout << "ShowOptParameters()"
+    << ": ADD_ANALYSIS " << ADD_ANALYSIS
+    << ": BACK_OFF " << BACK_OFF
+    << ": INLINE_VERSION_OPT " << INLINE_VERSION_OPT
+    << ": INLINE_VERSION_PROMOTION " << INLINE_VERSION_PROMOTION
+    << ": MASSTREE_USE " << MASSTREE_USE
+    << ": PARTITION_TABLE " << PARTITION_TABLE
+    << ": REUSE_VERSION " << REUSE_VERSION
+    << ": SINGLE_EXEC " << SINGLE_EXEC
+    << ": KEY_SIZE " << KEY_SIZE
+    << ": VAL_SIZE " << VAL_SIZE
+    << ": WRITE_LATEST_ONLY " << WRITE_LATEST_ONLY
+    << ": WORKER1_INSERT_DELAY_RPHASE " << WORKER1_INSERT_DELAY_RPHASE
+    << endl;
+}

common/result.cc

@@ -285,7 +285,6 @@ void Result::displayValidationFailureByWritelockRate() {
 }
 
 void Result::displayVersionMalloc() {
-  if (total_version_malloc_)
     cout << "version_malloc:\t" << total_version_malloc_ << endl;
 }
 

include/delay.hh

@@ -0,0 +1,24 @@
+#pragma once
+
+#include <x86intrin.h>
+
+#include <iostream>
+
+#include "tsc.hh"
+
+using namespace std;
+
+[[maybe_unused]] static void 
+clock_delay(size_t clocks)
+{
+  std::size_t start(rdtscp()), stop;
+
+  for (;;) {
+    stop = rdtscp();
+    if (stop - start > clocks) {
+      break;
+    } else {
+      _mm_pause();
+    }
+  }
+}

include/tsc.hh

@@ -2,7 +2,9 @@
 
 #include <stdint.h>
 
-[[maybe_unused]] static uint64_t rdtsc() {
+[[maybe_unused]] static uint64_t 
+rdtsc() 
+{
   uint64_t rax;
   uint64_t rdx;
 
@@ -14,7 +16,9 @@
   return (rdx << 32) | rax;
 }
 
-[[maybe_unused]] static uint64_t rdtsc_serial() {
+[[maybe_unused]] static uint64_t 
+rdtsc_serial() 
+{
   uint64_t rax;
   uint64_t rdx;
 
@@ -24,7 +28,9 @@
   return (rdx << 32) | rax;
 }
 
-[[maybe_unused]] static uint64_t rdtscp() {
+[[maybe_unused]] static uint64_t 
+rdtscp() 
+{
   uint64_t rax;
   uint64_t rdx;
   uint32_t aux;

silo/Makefile

@@ -16,9 +16,8 @@ MASSTREE_USE=1
 NO_WAIT_LOCKING_IN_VALIDATION=1
 PARTITION_TABLE=0
 PROCEDURE_SORT=0
-WAL=1
-
-SLEEP_READ_PHASE=0 # # of tics
+WAL=0
+SLEEP_READ_PHASE=0 # of tics
 # end of initialization
 
 CC = g++
@@ -40,7 +39,6 @@ CFLAGS = -c -pipe -g -O3 -std=c++17 -march=native \
 LDFLAGS = -L../third_party/mimalloc/out/release -L../third_party/mimalloc/out/debug -L/user/lib
 INCLUDE = -I/usr/include
 
-#LDLIBS = -lpthread -lmimalloc
 LDLIBS = -lpthread -lboost_filesystem -lboost_system
 
 OBJS1 = $(SILO_SRCS1:.cc=.o)

silo/include/util.hh

@@ -15,3 +15,5 @@ extern void leaderWork(uint64_t &epoch_timer_start, uint64_t &epoch_timer_stop);
 extern void makeDB();
 
 extern void partTableInit([[maybe_unused]] size_t thid, uint64_t start, uint64_t end);
+
+extern void ShowOptParameters();

silo/replayTest.cc

@@ -29,9 +29,9 @@ int main() {
   for (unsigned int i = 0; i < loadhd.logRecNum_; ++i) {
     loadfile.read((void *)&logrec, sizeof(LogRecord));
     chkSum_ += logrec.computeChkSum();
-    // cout << "tid : " << logrec.tid << endl;
-    // cout << "key : " << logrec.key << endl;
-    // cout << "val : " << logrec.val << endl << endl;
+    cout << "tid : " << logrec.tid_ << endl;
+    cout << "key : " << logrec.key_ << endl;
+    cout << "val : " << logrec.val_ << endl << endl;
   }
 
   cout << "computed chkSum_ : " << chkSum_ << endl;