database/mariadb-columnstore-engine
1
0
Fork 0
mirror of https://github.com/mariadb-corporation/mariadb-columnstore-engine.git synced 2025-07-30 19:23:07 +03:00
Code Activity

MCOL-5044 This patch replaces PriorityThreadPool with FairThreadPool that uses a simple

Browse Source 
operations + morsel size weight model to equally allocate CPU b/w parallel query morsels.
This patch delivers better parallel query timings distribution(timings graph resembles normal
distribution with a bigger left side thus more queries runs faster comparing with PrioThreadPool-based
single-node installation).
See changes in batchprimitiveprocessor-jl.h and comments in fair_threadpool.h for
important implementation details
This commit is contained in:
Roman Nozdrin
2022-05-24 17:57:40 +00:00
committed by Roman Nozdrin
parent 0f0b3a2bed
commit fd8ba33f21
12 changed files with 173 additions and 254 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
dbcon/joblist/batchprimitiveprocessor-jl.cpp
View File

@ -1274,6 +1274,10 @@ void BatchPrimitiveProcessorJL::runBPP(ByteStream& bs, uint32_t pmNum)
bs << uniqueID;
bs << _priority;
// The weight is used by PrimProc thread pool algo
uint32_t weight = calculateBPPWeight();
bs << weight;
bs << dbRoot;
bs << count;

21
dbcon/joblist/batchprimitiveprocessor-jl.h
View File

@ -252,8 +252,27 @@ class BatchPrimitiveProcessorJL
}
private:
// void setLBIDForScan(uint64_t rid, uint32_t dbroot);
const size_t perColumnProjectWeight_ = 10;
const size_t perColumnFilteringWeight_ = 10;
const size_t fe1Weight_ = 10;
const size_t fe2Weight_ = 10;
const size_t joinWeight_ = 500;
const size_t aggregationWeight_ = 500;
// This is simple SQL operations-based model leveraged by
// FairThreadPool run by PP facility.
// Every operation mentioned in this calculation spends
// some CPU so the morsel uses this op weights more.
uint32_t calculateBPPWeight() const
{
uint32_t weight = perColumnProjectWeight_ * projectCount;
weight += filterCount * perColumnFilteringWeight_;
weight += tJoiners.size() * joinWeight_;
weight += (aggregatorPM) ? aggregationWeight_ : 0;
weight += (fe1) ? fe1Weight_ : 0;
weight += (fe2) ? fe2Weight_ : 0;
return weight;
}
BPSOutputType ot;
bool needToSetLBID;

1
dbcon/joblist/diskjoinstep.h
View File

@ -20,7 +20,6 @@
#include "tuplehashjoin.h"
#include "joinpartition.h"
#include "threadnaming.h"
#include "../../utils/threadpool/prioritythreadpool.h"
#pragma once

3
primitives/primproc/batchprimitiveprocessor.cpp
View File

@ -140,6 +140,7 @@ BatchPrimitiveProcessor::BatchPrimitiveProcessor()
, ptMask(0)
, firstInstance(false)
, valuesLBID(0)
, weight_(0)
{
pp.setLogicalBlockMode(true);
pp.setBlockPtr((int*)blockData);
@ -193,6 +194,7 @@ BatchPrimitiveProcessor::BatchPrimitiveProcessor(ByteStream& b, double prefetch,
// ptMask(processorThreads - 1),
, firstInstance(true)
, valuesLBID(0)
, weight_(0)
{
// promote processorThreads to next power of 2. also need to change the name to bucketCount or similar
processorThreads = nextPowOf2(processorThreads);
@ -542,6 +544,7 @@ void BatchPrimitiveProcessor::resetBPP(ByteStream& bs, const SP_UM_MUTEX& w, con
// skip the header, sessionID, stepID, uniqueID, and priority
bs.advance(sizeof(ISMPacketHeader) + 16);
bs >> weight_;
bs >> dbRoot;
bs >> count;
bs >> ridCount;

8
primitives/primproc/batchprimitiveprocessor.h
View File

@ -137,6 +137,11 @@ class BatchPrimitiveProcessor
fBusy = b;
}
size_t getWeight() const
{
return weight_;
}
uint16_t FilterCount() const
{
return filterCount;
@ -433,6 +438,9 @@ class BatchPrimitiveProcessor
uint ptMask;
bool firstInstance;
uint64_t valuesLBID;
uint32_t weight_;
static const uint64_t maxResultCount = 1048576; // 2^20
friend class Command;
friend class ColumnCommand;

7
primitives/primproc/bppseeder.h
View File

@ -47,7 +47,7 @@
namespace primitiveprocessor
{
class BPPSeeder : public threadpool::PriorityThreadPool::Functor
class BPPSeeder : public threadpool::FairThreadPool::Functor
{
public:
BPPSeeder(const messageqcpp::SBS&, const SP_UM_MUTEX& wLock, const SP_UM_IOSOCK& ios, const int pmThreads,
@ -71,6 +71,11 @@ class BPPSeeder : public threadpool::PriorityThreadPool::Functor
{
return _priority;
}
size_t getWeight() const
{
assert(bpp);
return bpp->getWeight();
}
private:
BPPSeeder();

274
primitives/primproc/primitiveserver.cpp
View File

@ -124,7 +124,7 @@ oam::OamCache* oamCache = oam::OamCache::makeOamCache();
// FIXME: there is an anon ns burried later in between 2 named namespaces...
namespace primitiveprocessor
{
boost::shared_ptr<threadpool::PriorityThreadPool> OOBPool;
boost::shared_ptr<threadpool::FairThreadPool> OOBPool;
BlockRequestProcessor** BRPp;
#ifndef _MSC_VER
@ -1050,7 +1050,7 @@ using namespace primitiveprocessor;
/** @brief The job type to process a dictionary scan (pDictionaryScan class on the UM)
* TODO: Move this & the impl into different files
*/
class DictScanJob : public threadpool::PriorityThreadPool::Functor
class DictScanJob : public threadpool::FairThreadPool::Functor
{
public:
DictScanJob(SP_UM_IOSOCK ios, SBS bs, SP_UM_MUTEX writeLock);
@ -1242,7 +1242,7 @@ struct BPPHandler
scoped.unlock();
}
struct BPPHandlerFunctor : public PriorityThreadPool::Functor
struct BPPHandlerFunctor : public FairThreadPool::Functor
{
BPPHandlerFunctor(boost::shared_ptr<BPPHandler> r, SBS b) : bs(b)
{
@ -1710,7 +1710,7 @@ return 0;
PrimitiveServer* fPrimitiveServerPtr;
};
class DictionaryOp : public PriorityThreadPool::Functor
class DictionaryOp : public FairThreadPool::Functor
{
public:
DictionaryOp(SBS cmd) : bs(cmd)
@ -1947,8 +1947,7 @@ struct ReadThread
void operator()()
{
utils::setThreadName("PPReadThread");
boost::shared_ptr<threadpool::PriorityThreadPool> procPoolPtr =
fPrimitiveServerPtr->getProcessorThreadPool();
threadpool::FairThreadPool* procPoolPtr = fPrimitiveServerPtr->getProcessorThreadPool();
SBS bs;
UmSocketSelector* pUmSocketSelector = UmSocketSelector::instance();
@ -2044,35 +2043,69 @@ struct ReadThread
switch (ismHdr->Command)
{
case DICT_CREATE_EQUALITY_FILTER:
{
PriorityThreadPool::Job job;
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
job.functor = boost::shared_ptr<PriorityThreadPool::Functor>(new CreateEqualityFilter(bs));
OOBPool->addJob(job);
break;
}
case DICT_DESTROY_EQUALITY_FILTER:
case BATCH_PRIMITIVE_CREATE:
case BATCH_PRIMITIVE_ADD_JOINER:
case BATCH_PRIMITIVE_END_JOINER:
case BATCH_PRIMITIVE_DESTROY:
case BATCH_PRIMITIVE_ABORT:
{
PriorityThreadPool::Job job;
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
job.functor = boost::shared_ptr<PriorityThreadPool::Functor>(new DestroyEqualityFilter(bs));
const uint32_t txnId = *((uint32_t*)&buf[pos + 2]);
const uint32_t stepID = *((uint32_t*)&buf[pos + 6]);
const uint32_t uniqueID = *((uint32_t*)&buf[pos + 10]);
const uint32_t weight = 1;
const uint32_t priority = 0;
uint32_t id = 0;
boost::shared_ptr<FairThreadPool::Functor> functor;
if (ismHdr->Command == DICT_CREATE_EQUALITY_FILTER)
{
functor.reset(new CreateEqualityFilter(bs));
}
else if (ismHdr->Command == DICT_DESTROY_EQUALITY_FILTER)
{
functor.reset(new DestroyEqualityFilter(bs));
}
else if (ismHdr->Command == BATCH_PRIMITIVE_CREATE)
{
functor.reset(new BPPHandler::Create(fBPPHandler, bs));
}
else if (ismHdr->Command == BATCH_PRIMITIVE_ADD_JOINER)
{
functor.reset(new BPPHandler::AddJoiner(fBPPHandler, bs));
}
else if (ismHdr->Command == BATCH_PRIMITIVE_END_JOINER)
{
id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
functor.reset(new BPPHandler::LastJoiner(fBPPHandler, bs));
}
else if (ismHdr->Command == BATCH_PRIMITIVE_DESTROY)
{
functor.reset(new BPPHandler::Destroy(fBPPHandler, bs));
}
else if (ismHdr->Command == BATCH_PRIMITIVE_ABORT)
{
id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
functor.reset(new BPPHandler::Abort(fBPPHandler, bs));
}
FairThreadPool::Job job(uniqueID, stepID, txnId, functor, outIos, weight, priority, id);
OOBPool->addJob(job);
break;
}
case DICT_TOKEN_BY_SCAN_COMPARE:
case BATCH_PRIMITIVE_RUN:
{
idbassert(bs->length() >= sizeof(TokenByScanRequestHeader));
TokenByScanRequestHeader* hdr = (TokenByScanRequestHeader*)ismHdr;
TokenByScanRequestHeader* hdr = nullptr;
boost::shared_ptr<FairThreadPool::Functor> functor;
uint32_t id = 0;
uint32_t weight = 0;
uint32_t priority = 0;
uint32_t txnId = 0;
uint32_t stepID = 0;
uint32_t uniqueID = 0;
bool isSyscat = false;
if (bRotateDest)
{
@ -2090,23 +2123,41 @@ struct ReadThread
}
}
PriorityThreadPool::Job job;
job.functor = boost::shared_ptr<DictScanJob>(new DictScanJob(outIos, bs, writeLock));
job.id = hdr->Hdr.UniqueID;
job.weight = LOGICAL_BLOCK_RIDS;
job.priority = hdr->Hdr.Priority;
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
if (hdr->flags & IS_SYSCAT)
if (ismHdr->Command == DICT_TOKEN_BY_SCAN_COMPARE)
{
idbassert(bs->length() >= sizeof(TokenByScanRequestHeader));
hdr = (TokenByScanRequestHeader*)ismHdr;
functor.reset(new DictScanJob(outIos, bs, writeLock));
id = hdr->Hdr.UniqueID;
weight = LOGICAL_BLOCK_RIDS;
priority = hdr->Hdr.Priority;
const uint8_t* buf = bs->buf();
const uint32_t pos = sizeof(ISMPacketHeader) - 2;
txnId = *((uint32_t*)&buf[pos + 2]);
stepID = *((uint32_t*)&buf[pos + 6]);
uniqueID = *((uint32_t*)&buf[pos + 10]);
isSyscat = hdr->flags & IS_SYSCAT;
}
else if (ismHdr->Command == BATCH_PRIMITIVE_RUN)
{
functor.reset(new BPPSeeder(bs, writeLock, outIos,
fPrimitiveServerPtr->ProcessorThreads(),
fPrimitiveServerPtr->PTTrace()));
BPPSeeder* bpps = dynamic_cast<BPPSeeder*>(functor.get());
id = bpps->getID();
priority = bpps->priority();
const uint8_t* buf = bs->buf();
const uint32_t pos = sizeof(ISMPacketHeader) - 2;
txnId = *((uint32_t*)&buf[pos + 2]);
stepID = *((uint32_t*)&buf[pos + 6]);
uniqueID = *((uint32_t*)&buf[pos + 10]);
weight = ismHdr->Size + *((uint32_t*)&buf[pos + 18]);
isSyscat = bpps->isSysCat();
}
FairThreadPool::Job job(uniqueID, stepID, txnId, functor, outIos, weight, priority, id);
if (isSyscat)
{
// boost::thread t(DictScanJob(outIos, bs, writeLock));
// using already-existing threads may cut latency
// if it's changed back to running in an independent thread
// change the issyscat() checks in BPPSeeder as well
OOBPool->addJob(job);
}
else
@ -2117,146 +2168,11 @@ struct ReadThread
break;
}
case BATCH_PRIMITIVE_RUN:
{
if (bRotateDest)
{
if (!pUmSocketSelector->nextIOSocket(fIos, outIos, writeLock))
{
// If we ever fall into this part of the
// code we have a "bug" of some sort.
// See handleUmSockSelErr() for more info.
// We reset ios and mutex to defaults.
handleUmSockSelErr(string("BPR cmd"));
outIos = outIosDefault;
writeLock = writeLockDefault;
pUmSocketSelector->delConnection(fIos);
bRotateDest = false;
}
}
/* Decide whether this is a syscat call and run
right away instead of queueing */
boost::shared_ptr<BPPSeeder> bpps(new BPPSeeder(bs, writeLock, outIos,
fPrimitiveServerPtr->ProcessorThreads(),
fPrimitiveServerPtr->PTTrace()));
PriorityThreadPool::Job job;
job.functor = bpps;
job.id = bpps->getID();
job.weight = ismHdr->Size;
job.priority = bpps->priority();
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
if (bpps->isSysCat())
{
// boost::thread t(*bpps);
// using already-existing threads may cut latency
// if it's changed back to running in an independent thread
// change the issyscat() checks in BPPSeeder as well
OOBPool->addJob(job);
}
else
{
procPoolPtr->addJob(job);
}
break;
}
case BATCH_PRIMITIVE_CREATE:
{
PriorityThreadPool::Job job;
job.functor =
boost::shared_ptr<PriorityThreadPool::Functor>(new BPPHandler::Create(fBPPHandler, bs));
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
OOBPool->addJob(job);
// fBPPHandler->createBPP(*bs);
break;
}
case BATCH_PRIMITIVE_ADD_JOINER:
{
PriorityThreadPool::Job job;
job.functor =
boost::shared_ptr<PriorityThreadPool::Functor>(new BPPHandler::AddJoiner(fBPPHandler, bs));
job.id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
OOBPool->addJob(job);
// fBPPHandler->addJoinerToBPP(*bs);
break;
}
case BATCH_PRIMITIVE_END_JOINER:
{
// lastJoinerMsg can block; must do this in a different thread
// OOBPool->invoke(BPPHandler::LastJoiner(fBPPHandler, bs)); // needs a threadpool that can
// resched boost::thread tmp(BPPHandler::LastJoiner(fBPPHandler, bs));
PriorityThreadPool::Job job;
job.functor =
boost::shared_ptr<PriorityThreadPool::Functor>(new BPPHandler::LastJoiner(fBPPHandler, bs));
job.id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
OOBPool->addJob(job);
break;
}
case BATCH_PRIMITIVE_DESTROY:
{
// OOBPool->invoke(BPPHandler::Destroy(fBPPHandler, bs)); // needs a threadpool that can
// resched boost::thread tmp(BPPHandler::Destroy(fBPPHandler, bs));
PriorityThreadPool::Job job;
job.functor =
boost::shared_ptr<PriorityThreadPool::Functor>(new BPPHandler::Destroy(fBPPHandler, bs));
job.id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
OOBPool->addJob(job);
// fBPPHandler->destroyBPP(*bs);
break;
}
case BATCH_PRIMITIVE_ACK:
{
fBPPHandler->doAck(*bs);
break;
}
case BATCH_PRIMITIVE_ABORT:
{
// OBPool->invoke(BPPHandler::Abort(fBPPHandler, bs));
// fBPPHandler->doAbort(*bs);
PriorityThreadPool::Job job;
job.functor =
boost::shared_ptr<PriorityThreadPool::Functor>(new BPPHandler::Abort(fBPPHandler, bs));
job.id = fBPPHandler->getUniqueID(bs, ismHdr->Command);
const uint8_t* buf = bs->buf();
uint32_t pos = sizeof(ISMPacketHeader) - 2;
job.stepID = *((uint32_t*)&buf[pos + 6]);
job.uniqueID = *((uint32_t*)&buf[pos + 10]);
job.sock = outIos;
OOBPool->addJob(job);
break;
}
default:
{
std::ostringstream os;
@ -2406,12 +2322,12 @@ PrimitiveServer::PrimitiveServer(int serverThreads, int serverQueueSize, int pro
fServerpool.setQueueSize(fServerQueueSize);
fServerpool.setName("PrimitiveServer");
fProcessorPool.reset(new threadpool::PriorityThreadPool(fProcessorWeight, highPriorityThreads,
medPriorityThreads, lowPriorityThreads, 0));
fProcessorPool = new threadpool::FairThreadPool(fProcessorWeight, highPriorityThreads,
medPriorityThreads, lowPriorityThreads, 0);
// We're not using either the priority or the job-clustering features, just need a threadpool
// that can reschedule jobs, and an unlimited non-blocking queue
OOBPool.reset(new threadpool::PriorityThreadPool(1, 5, 0, 0, 1));
OOBPool.reset(new threadpool::FairThreadPool(1, 5, 0, 0, 1));
asyncCounter = 0;

7
primitives/primproc/primitiveserver.h
View File

@ -37,6 +37,7 @@
#include "threadpool.h"
#include "../../utils/threadpool/prioritythreadpool.h"
#include "fair_threadpool.h"
#include "messagequeue.h"
#include "blockrequestprocessor.h"
#include "batchprimitiveprocessor.h"
@ -48,7 +49,7 @@ extern oam::OamCache* oamCache;
namespace primitiveprocessor
{
extern boost::shared_ptr<threadpool::PriorityThreadPool> OOBPool;
extern boost::shared_ptr<threadpool::FairThreadPool> OOBPool;
extern dbbc::BlockRequestProcessor** BRPp;
extern BRM::DBRM* brm;
extern boost::mutex bppLock;
@ -128,7 +129,7 @@ class PrimitiveServer
/** @brief get a pointer the shared processor thread pool
*/
inline boost::shared_ptr<threadpool::PriorityThreadPool> getProcessorThreadPool() const
inline threadpool::FairThreadPool* getProcessorThreadPool() const
{
return fProcessorPool;
}
@ -165,7 +166,7 @@ class PrimitiveServer
/** @brief the thread pool used to process
* primitive commands
*/
boost::shared_ptr<threadpool::PriorityThreadPool> fProcessorPool;
threadpool::FairThreadPool* fProcessorPool;
int fServerThreads;
int fServerQueueSize;

9
tests/CMakeLists.txt
View File

@ -51,11 +51,10 @@ if (WITH_UNITTESTS)
target_link_libraries(simd_processors ${ENGINE_LDFLAGS} ${MARIADB_CLIENT_LIBS} ${ENGINE_WRITE_LIBS} ${GTEST_LIBRARIES} processor dbbc)
gtest_discover_tests(simd_processors TEST_PREFIX columnstore:)
# Comment this out b/c FairThreadPoolRemove segfaults in containers. Moreover this code isn't
# in production yet.
# add_executable(fair_threadpool_test fair_threadpool.cpp)
# target_link_libraries(fair_threadpool_test ${ENGINE_LDFLAGS} ${MARIADB_CLIENT_LIBS} ${ENGINE_WRITE_LIBS} ${GTEST_LIBRARIES} processor dbbc)
# gtest_discover_tests(fair_threadpool_test TEST_PREFIX columnstore:)
add_executable(fair_threadpool_test fair_threadpool.cpp)
add_dependencies(fair_threadpool_test googletest)
target_link_libraries(fair_threadpool_test ${ENGINE_LDFLAGS} ${MARIADB_CLIENT_LIBS} ${ENGINE_WRITE_LIBS} ${GTEST_LIBRARIES} processor dbbc)
gtest_discover_tests(fair_threadpool_test TEST_PREFIX columnstore:)
# CPPUNIT TESTS
add_executable(we_shared_components_tests shared_components_tests.cpp)

19
tests/fair_threadpool.cpp
View File

@ -111,12 +111,11 @@ TEST_F(FairThreadPoolTest, FairThreadPoolAdd)
while (threadPool->queueSize())
{
usleep(2500000);
usleep(250000);
}
usleep(2500000);
EXPECT_EQ(threadPool->queueSize(), 0);
EXPECT_EQ(results.size(), 3);
EXPECT_EQ(threadPool->queueSize(), 0ULL);
EXPECT_EQ(results.size(), 3ULL);
EXPECT_EQ(results[0], 1);
EXPECT_EQ(results[1], 3);
EXPECT_EQ(results[2], 2);
@ -139,11 +138,11 @@ TEST_F(FairThreadPoolTest, FairThreadPoolRemove)
while (threadPool->queueSize())
{
usleep(1500000);
usleep(250000);
}
EXPECT_EQ(threadPool->queueSize(), 0);
EXPECT_EQ(results.size(), 2);
EXPECT_EQ(threadPool->queueSize(), 0ULL);
EXPECT_EQ(results.size(), 2ULL);
EXPECT_EQ(results[0], 1);
EXPECT_EQ(results[1], 3);
}
@ -164,11 +163,11 @@ TEST_F(FairThreadPoolTest, FairThreadPoolReschedule)
while (threadPool->queueSize())
{
usleep(1500000);
usleep(250000);
}
EXPECT_EQ(threadPool->queueSize(), 0);
EXPECT_EQ(results.size(), 3);
EXPECT_EQ(threadPool->queueSize(), 0ULL);
EXPECT_EQ(results.size(), 3ULL);
EXPECT_EQ(results[0], 1);
EXPECT_TRUE(isThisOrThat(results, 1, 2, 2, 3));
}

48
utils/threadpool/fair_threadpool.cpp
View File

@ -34,7 +34,7 @@ namespace threadpool
{
FairThreadPool::FairThreadPool(uint targetWeightPerRun, uint highThreads, uint midThreads, uint lowThreads,
uint ID)
: _stop(false), weightPerRun(targetWeightPerRun), id(ID), blockedThreads(0), extraThreads(0), stopExtra(true)
: _stop(false), weightPerRun(targetWeightPerRun), id(ID)
{
boost::thread* newThread;
size_t numberOfThreads = highThreads + midThreads + lowThreads;
@ -73,32 +73,20 @@ void FairThreadPool::addJob_(const Job& job, bool useLock)
++threadCounts;
}
// If some threads have blocked (because of output queue full)
// Temporarily add some extra worker threads to make up for the blocked threads.
if (blockedThreads > extraThreads)
{
stopExtra = false;
newThread = threads.create_thread(ThreadHelper(this, PriorityThreadPool::Priority::EXTRA));
newThread->detach();
extraThreads++;
}
else if (blockedThreads == 0)
{
// Release the temporary threads -- some threads have become unblocked.
stopExtra = true;
}
auto jobsListMapIter = txn2JobsListMap_.find(job.txnIdx_);
if (jobsListMapIter == txn2JobsListMap_.end())
if (jobsListMapIter == txn2JobsListMap_.end()) // there is no txn in the map
{
ThreadPoolJobsList* jobsList = new ThreadPoolJobsList;
jobsList->push_back(job);
txn2JobsListMap_[job.txnIdx_] = jobsList;
WeightT currentTopWeight = weightedTxnsQueue_.empty() ? 0 : weightedTxnsQueue_.top().first;
weightedTxnsQueue_.push({currentTopWeight, job.txnIdx_});
weightedTxnsQueue_.push({job.weight_, job.txnIdx_});
}
else
else // txn is in the map
{
if (jobsListMapIter->second->empty()) // there are no jobs for the txn
{
weightedTxnsQueue_.push({job.weight_, job.txnIdx_});
}
jobsListMapIter->second->push_back(job);
}
@ -135,11 +123,8 @@ void FairThreadPool::removeJobs(uint32_t id)
void FairThreadPool::threadFcn(const PriorityThreadPool::Priority preferredQueue)
{
if (preferredQueue == PriorityThreadPool::Priority::EXTRA)
utils::setThreadName("Extra");
else
utils::setThreadName("Idle");
RunListT runList; // This is a vector to allow to grab multiple jobs
utils::setThreadName("Idle");
RunListT runList(1); // This is a vector to allow to grab multiple jobs
RescheduleVecType reschedule;
bool running = false;
bool rescheduleJob = false;
@ -151,12 +136,6 @@ void FairThreadPool::threadFcn(const PriorityThreadPool::Priority preferredQueue
runList.clear(); // remove the job
std::unique_lock<std::mutex> lk(mutex);
if (preferredQueue == PriorityThreadPool::Priority::EXTRA && stopExtra)
{
--extraThreads;
return;
}
if (weightedTxnsQueue_.empty())
{
newJob.wait(lk);
@ -166,7 +145,7 @@ void FairThreadPool::threadFcn(const PriorityThreadPool::Priority preferredQueue
WeightedTxnT weightedTxn = weightedTxnsQueue_.top();
auto txnAndJobListPair = txn2JobsListMap_.find(weightedTxn.second);
// Looking for non-empty jobsList in a loop
// Waiting on cond_var if PQ is empty(no jobs in this thread pool)
// The loop waits on newJob cond_var if PQ is empty(no jobs in this thread pool)
while (txnAndJobListPair == txn2JobsListMap_.end() || txnAndJobListPair->second->empty())
{
// JobList is empty. This can happen when this method pops the last Job.
@ -196,7 +175,6 @@ void FairThreadPool::threadFcn(const PriorityThreadPool::Priority preferredQueue
weightedTxnsQueue_.pop();
TransactionIdxT txnIdx = txnAndJobListPair->first;
ThreadPoolJobsList* jobsList = txnAndJobListPair->second;
// Job& job = jobsList->front();
runList.push_back(jobsList->front());
jobsList->pop_front();
@ -210,13 +188,15 @@ void FairThreadPool::threadFcn(const PriorityThreadPool::Priority preferredQueue
lk.unlock();
running = true;
rescheduleJob = (*(runList[0].functor_))();
rescheduleJob = (*(runList[0].functor_))(); // run the functor
running = false;
utils::setThreadName("Idle");
if (rescheduleJob)
{
// to avoid excessive CPU usage waiting for data from storage
usleep(500);
lk.lock();
addJob_(runList[0], false);
newJob.notify_one();

26
utils/threadpool/fair_threadpool.h
View File

@ -39,6 +39,12 @@
namespace threadpool
{
// The idea of this thread pool is to run morsel jobs(primitive job) is to equaly distribute CPU time
// b/w multiple parallel queries(thread maps morsel to query using txnId). Query(txnId) has its weight
// stored in PriorityQueue that thread increases before run another morsel for the query. When query is
// done(ThreadPoolJobsList is empty) it is removed from PQ and the Map(txn to ThreadPoolJobsList).
// I tested multiple morsels per one loop iteration in ::threadFcn. This approach reduces CPU consumption
// and increases query timings.
class FairThreadPool
{
public:
@ -92,23 +98,6 @@ class FairThreadPool
*/
void dump();
// If a job is blocked, we want to temporarily increase the number of threads managed by the pool
// A problem can occur if all threads are running long or blocked for a single query. Other
// queries won't get serviced, even though there are cpu cycles available.
// These calls are currently protected by respondLock in sendThread(). If you call from other
// places, you need to consider atomicity.
void incBlockedThreads()
{
blockedThreads++;
}
void decBlockedThreads()
{
blockedThreads--;
}
uint32_t blockedThreadCount() const
{
return blockedThreads;
}
size_t queueSize() const
{
return weightedTxnsQueue_.size();
@ -165,9 +154,6 @@ class FairThreadPool
using Txn2ThreadPoolJobsListMap = std::unordered_map<TransactionIdxT, ThreadPoolJobsList*>;
Txn2ThreadPoolJobsListMap txn2JobsListMap_;
WeightedTxnPrioQueue weightedTxnsQueue_;
std::atomic<uint32_t> blockedThreads;
std::atomic<uint32_t> extraThreads;
bool stopExtra;
};
} // namespace threadpool