Hello Adar Dembo,
I'd like you to do a code review. Please visit
http://gerrit.cloudera.org:8080/6472
to review the following change.
Change subject: diskrowset: stop using percpu_rwlock
......................................................................
diskrowset: stop using percpu_rwlock
In a heap analysis of a server with 34K rowsets, I see percpu_rwlocks
inside DiskRowSet taking ~105MB of memory. Each percpu_rwlock is using 48
cache lines (3072 bytes) so the multiplication works out to around the
same value as I see in practice.
Although switching to a normal rwlock might increase contention
marginally, it's probably worth it for the 100MB memory win. If we start
to see contention here in a workload we can figure out other ways to
avoid taking the locks, but my guess is that the improved cache
locality of the lock being stored inline in the DRS class will outweigh
any contention-related issue.
Change-Id: I9e210eb3e8fc13d807f6630e68c68d64902b1cc4
---
M src/kudu/tablet/diskrowset.cc
M src/kudu/tablet/diskrowset.h
2 files changed, 11 insertions(+), 11 deletions(-)
git pull ssh://gerrit.cloudera.org:29418/kudu refs/changes/72/6472/1
--
To view, visit http://gerrit.cloudera.org:8080/6472
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings
Gerrit-MessageType: newchange
Gerrit-Change-Id: I9e210eb3e8fc13d807f6630e68c68d64902b1cc4
Gerrit-PatchSet: 1
Gerrit-Project: kudu
Gerrit-Branch: master
Gerrit-Owner: Todd Lipcon <[email protected]>
Gerrit-Reviewer: Adar Dembo <[email protected]>
