[
https://issues.apache.org/jira/browse/KUDU-1977?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16399753#comment-16399753
]
Todd Lipcon commented on KUDU-1977:
-----------------------------------
A nice way to repro this is the following against a 6-node cluster:
$ kudu perf loadgen -keep-auto-table -num-threads 16 -num-rows-per-thread
$[1000*1000*1000] -table-num-replicas 3 -table-num-buckets=20 localhost
A profile shows a ton of CPU in the above-referenced code
(kudu::client::internal::MetaCache::LookupTabletByKeyFastPath is the top
consumer)
> MetaCache cache hits should avoid ref count incr/decr
> -----------------------------------------------------
>
> Key: KUDU-1977
> URL: https://issues.apache.org/jira/browse/KUDU-1977
> Project: Kudu
> Issue Type: Improvement
> Components: client, perf
> Affects Versions: 1.3.0
> Reporter: Todd Lipcon
> Priority: Major
>
> Looking at CPU consumption of a client using perf tools (including the new
> perf-c2c tool) I see that metacache hits in multi-threaded writers are rather
> slow because of the reference count increments and decrements on the
> MetaCache object.
> In the case of the cache hit, there's no real need to allocate the LookupRpc
> object, increment ref counts, and then decrement them. Instead we could move
> the fast-path check out of LookupRpc and save some CPU.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
