Mostafa Mokhtar created IMPALA-6359:
---------------------------------------
Summary: Blocking in Service threads due to tcmalloc contention
due to memory allocations for RowBatch::Deserialize
Key: IMPALA-6359
URL: https://issues.apache.org/jira/browse/IMPALA-6359
Project: IMPALA
Issue Type: Sub-task
Reporter: Mostafa Mokhtar
Priority: Critical
This is a regression compared to thrift with KRPC is enabled.
Results are based on running 48 concurrent TPCDS queries on a 40 node cluster,
regression is in the order of 2-4x (Will update the JIRA with a more accurate
number soon).
Service threads stall on memory allocation in the stack below, looking at the
Service threads memory allocation/deallocation I believe they won't benefit
much from the per thread cache as there is more allocations than deallocations
on this threads as a result they go to the CentralFreeList for most
allocations.
{code}
#0 0x00000000020931e7 in base::internal::SpinLockDelay(int volatile*, int,
int) ()
#1 0x00000000020dde29 in SpinLock::SlowLock() ()
#2 0x0000000002121668 in tcmalloc::CentralFreeList::Populate() ()
#3 0x0000000002121768 in tcmalloc::CentralFreeList::FetchFromOneSpansSafe(int,
void**, void**) ()
#4 0x00000000021217f4 in tcmalloc::CentralFreeList::RemoveRange(void**,
void**, int) ()
#5 0x000000000212ef53 in tcmalloc::ThreadCache::FetchFromCentralCache(unsigned
long, unsigned long) ()
#6 0x00000000021d37fb in tc_malloc ()
#7 0x0000000000b27c3a in impala::MemPool::FindChunk(long, bool) ()
#8 0x0000000000b354e6 in impala::MemPool::Allocate(long) ()
#9 0x0000000000b34409 in impala::RowBatch::Deserialize(kudu::Slice const&,
kudu::Slice const&, long, bool) ()
#10 0x0000000000b347fb in impala::RowBatch::RowBatch(impala::RowDescriptor
const*, impala::RowBatchHeaderPB const&, kudu::Slice const&, kudu::Slice
const&, impala::MemTracker*) ()
#11 0x0000000000b15d68 in
impala::KrpcDataStreamRecvr::SenderQueue::AddBatchWork(long,
impala::RowBatchHeaderPB const&, kudu::Slice const&, kudu::Slice const&,
boost::unique_lock<impala::SpinLock>*) ()
#12 0x0000000000b18715 in
impala::KrpcDataStreamRecvr::SenderQueue::AddBatch(impala::TransmitDataRequestPB
const*, kudu::rpc::RpcContext*) ()
#13 0x0000000000b0e750 in
impala::KrpcDataStreamMgr::AddData(impala::TransmitDataRequestPB const*,
kudu::rpc::RpcContext*) ()
#14 0x0000000001183305 in
kudu::rpc::GeneratedServiceIf::Handle(kudu::rpc::InboundCall*) ()
#15 0x00000000011bad7d in impala::ImpalaServicePool::RunThread() ()
#16 0x0000000000d18a73 in
impala::Thread::SuperviseThread(std::basic_string<char, std::char_traits<char>,
std::allocator<char> > const&, std::basic_string<char, std::char_traits<char>,
std::allocator<char> > const&, boost::function<void ()()>,
impala::Promise<long>*) ()
{code}
I added logging to track allocations that take a long time and found that in
the extreme case some allocations take up to 10 seconds when allocating 16KB of
memory.
{code}
MonotonicStopWatch timer;
timer.Start();
SCOPED_TIMER(recvr_->deserialize_row_batch_timer_);
// At this point, the row batch will be inserted into batch_queue_. Close()
will
// handle deleting any unconsumed batches from batch_queue_. Close() cannot
proceed
// until there are no pending insertion to batch_queue_.
batch.reset(new RowBatch(recvr_->row_desc(), header, tuple_offsets,
tuple_data,
recvr_->mem_tracker()));
timer.Stop();
if (timer.ElapsedTime() > 1000000) {
LOG(INFO) << "TTT RowBatch allocation size:" <<
header.uncompressed_size() << " allocation time " << timer.ElapsedTime();
}
{code}
Call stacks from other threads that release RowBatch memory which also are
affected by the tcmalloc contention
{code}
#0 0x000000000212c1f4 in tcmalloc::PageHeap::MayMergeSpans(tcmalloc::Span*,
tcmalloc::Span*) ()
#1 0x000000000212c54d in
tcmalloc::PageHeap::MergeIntoFreeList(tcmalloc::Span*) ()
#2 0x000000000212c874 in tcmalloc::PageHeap::Delete(tcmalloc::Span*) ()
#3 0x0000000002121090 in tcmalloc::CentralFreeList::ReleaseToSpans(void*) ()
#4 0x000000000212111b in tcmalloc::CentralFreeList::ReleaseListToSpans(void*)
()
#5 0x000000000212138c in tcmalloc::CentralFreeList::InsertRange(void*, void*,
int) ()
#6 0x000000000212f0f4 in
tcmalloc::ThreadCache::ReleaseToCentralCache(tcmalloc::ThreadCache::FreeList*,
unsigned long, int) ()
#7 0x000000000212f1bc in
tcmalloc::ThreadCache::ListTooLong(tcmalloc::ThreadCache::FreeList*, unsigned
long) ()
#8 0x00000000021d3d10 in tc_free ()
#9 0x0000000000b26af8 in impala::MemPool::FreeAll() ()
#10 0x0000000000b339b1 in impala::RowBatch::Reset() ()
#11 0x0000000000fea7b1 in
impala::PartitionedAggregationNode::Open(impala::RuntimeState*) ()
#12 0x0000000000b77dfd in impala::FragmentInstanceState::Open() ()
#13 0x0000000000b7943b in impala::FragmentInstanceState::Exec() ()
#14 0x0000000000b68a0a in
impala::QueryState::ExecFInstance(impala::FragmentInstanceState*) ()
{code}
{code}
Thread 682 (Thread 0x7f61ebec4700 (LWP 28762)):
#0 0x00000000020931e7 in base::internal::SpinLockDelay(int volatile*, int,
int) ()
#1 0x00000000020dde29 in SpinLock::SlowLock() ()
#2 0x000000000212140f in tcmalloc::CentralFreeList::InsertRange(void*, void*,
int) ()
#3 0x000000000212f0f4 in
tcmalloc::ThreadCache::ReleaseToCentralCache(tcmalloc::ThreadCache::FreeList*,
unsigned long, int) ()
#4 0x000000000212f1bc in
tcmalloc::ThreadCache::ListTooLong(tcmalloc::ThreadCache::FreeList*, unsigned
long) ()
#5 0x00000000021d3d10 in tc_free ()
#6 0x0000000000b338b1 in impala::RowBatch::~RowBatch() ()
#7 0x0000000000d8d1e3 in
impala::HdfsScanNode::GetNextInternal(impala::RuntimeState*, impala::RowBatch*,
bool*) ()
#8 0x0000000000d8ebf2 in impala::HdfsScanNode::GetNext(impala::RuntimeState*,
impala::RowBatch*, bool*) ()
#9 0x0000000000fee76e in
impala::PartitionedHashJoinNode::NextProbeRowBatch(impala::RuntimeState*,
impala::RowBatch*) ()
#10 0x0000000000ff43b1 in
impala::PartitionedHashJoinNode::GetNext(impala::RuntimeState*,
impala::RowBatch*, bool*) ()
#11 0x0000000000fee76e in
impala::PartitionedHashJoinNode::NextProbeRowBatch(impala::RuntimeState*,
impala::RowBatch*) ()
#12 0x0000000000ff43b1 in
impala::PartitionedHashJoinNode::GetNext(impala::RuntimeState*,
impala::RowBatch*, bool*) ()
#13 0x0000000000fee76e in
impala::PartitionedHashJoinNode::NextProbeRowBatch(impala::RuntimeState*,
impala::RowBatch*) ()
#14 0x0000000000ff43b1 in
impala::PartitionedHashJoinNode::GetNext(impala::RuntimeState*,
impala::RowBatch*, bool*) ()
#15 0x0000000000fea54d in
impala::PartitionedAggregationNode::Open(impala::RuntimeState*) ()
#16 0x0000000000b77dfd in impala::FragmentInstanceState::Open() ()
#17 0x0000000000b7943b in impala::FragmentInstanceState::Exec() ()
#18 0x0000000000b68a0a in
impala::QueryState::ExecFInstance(impala::FragmentInstanceState*) ()
#19 0x0000000000d18a73 in
impala::Thread::SuperviseThread(std::basic_string<char, std::char_traits<char>,
std::allocator<char> > const&, std::basic_string<char, std::char_traits<char>,
std::allocator<char> > const&, boost::function<void ()()>,
impala::Promise<long>*) ()
#20 0x0000000000d19204 in boost::detail::thread_data<boost::_bi::bind_t<void,
void (*)(std::basic_string<char, std::char_traits<char>, std::allocator<char> >
const&, std::basic_string<char, std::char_traits<char>, std::allocator<char> >
const&, boost::function<void ()()>, impala::Promise<long>*),
boost::_bi::list4<boost::_bi::value<std::basic_string<char,
std::char_traits<char>, std::allocator<char> > >,
boost::_bi::value<std::basic_string<char, std::char_traits<char>,
std::allocator<char> > >, boost::_bi::value<boost::function<void ()()> >,
boost::_bi::value<impala::Promise<long>*> > > >::run() ()
#21 0x000000000129bf5a in thread_proxy ()
#22 0x000000326c607aa1 in start_thread () from /lib64/libpthread.so.0
#23 0x000000326c2e893d in clone () from /lib64/libc.so.6
{code}
This regression was observed with the tcmalloc options below
{code}
TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=2147483648
TCMALLOC_TRANSFER_NUM_OBJ=40
{code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
