Hello Alexey Serbin, Dan Burkert, Kudu Jenkins, Adar Dembo,
I'd like you to reexamine a change. Please visit
http://gerrit.cloudera.org:8080/9753
to look at the new patch set (#3).
Change subject: KUDU-1977. Avoid extra reference counting for fast-path
MetaCache lookups
......................................................................
KUDU-1977. Avoid extra reference counting for fast-path MetaCache lookups
This does a bit of refactoring of MetaCache so that fast-path lookups
can occur without allocating a LookupRpc object. This avoids an
allocation/deallocation pair as well as reference counting overhead on
various objects. The reference counting was seen as a bottleneck in a
workload like:
$ kudu perf loadgen -num-threads 16 \
-num-rows-per-thread $[10*1000*1000] -table-num-replicas 3 \
-table-num-buckets=100 <master>
I ran before/after on the above command writing into a large (100+-node
cluster) that I had handy. A few key stats:
Before:
time total : 90676.3 ms
1458961.567442 task-clock (msec) # 15.954 CPUs utilized
Top lines of profile:
27.21% kudu kudu [.]
kudu::client::internal::MetaCache::LookupTabletByKeyFastPath
14.95% kudu kudu [.]
kudu::subtle::RefCountedThreadSafeBase::AddRef
12.52% kudu kudu [.]
kudu::subtle::RefCountedThreadSafeBase::Release
4.75% kudu kudu [.]
kudu::client::internal::MetaCache::LookupTabletByKey
3.04% kudu kudu [.] operator new[]
1.78% rpc reactor-387 kudu [.]
std::unordered_set<InFlightOp*>::erase
After:
time total : 60796.6 ms
693975.355549 task-clock (msec) # 11.154 CPUs utilized
Top lines of profile:
14.05% kudu kudu [.]
kudu::client::internal::MetaCache::LookupTabletByKeyFastPath
7.89% kudu kudu [.] operator new[]
4.96% rpc reactor-397 kudu [.]
std::unordered_set<InFlightOp*>::erase
4.47% kudu libc-2.17.so [.] __memcpy_ssse3_back
3.42% rpc reactor-397 kudu [.]
kudu::client::KuduInsert::~KuduInsert
In other words, this reduced CPU by more than 2x and throughput by about 50%.
Results on the above were relatively stable across multiple runs, with
wall-clock showing more variance than CPU-time.
The new bottleneck seems to be the rw_spinlock in MetaCache.
Change-Id: I5f17ffa88289c766b5b168b22da9781bf78f5592
---
M src/kudu/client/batcher.cc
M src/kudu/client/client-test.cc
M src/kudu/client/meta_cache.cc
M src/kudu/client/meta_cache.h
M src/kudu/client/partitioner-internal.cc
M src/kudu/client/scan_token-internal.cc
M src/kudu/client/scanner-internal.cc
M src/kudu/client/schema.h
8 files changed, 140 insertions(+), 98 deletions(-)
git pull ssh://gerrit.cloudera.org:29418/kudu refs/changes/53/9753/3
--
To view, visit http://gerrit.cloudera.org:8080/9753
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings
Gerrit-Project: kudu
Gerrit-Branch: master
Gerrit-MessageType: newpatchset
Gerrit-Change-Id: I5f17ffa88289c766b5b168b22da9781bf78f5592
Gerrit-Change-Number: 9753
Gerrit-PatchSet: 3
Gerrit-Owner: Todd Lipcon <[email protected]>
Gerrit-Reviewer: Adar Dembo <[email protected]>
Gerrit-Reviewer: Alexey Serbin <[email protected]>
Gerrit-Reviewer: Dan Burkert <[email protected]>
Gerrit-Reviewer: Kudu Jenkins
Gerrit-Reviewer: Todd Lipcon <[email protected]>
