[
https://issues.apache.org/jira/browse/KUDU-1897?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15881469#comment-15881469
]
Todd Lipcon commented on KUDU-1897:
-----------------------------------
https://k5wiki.kerberos.org/wiki/Projects/Replay_cache_improvements seems to
indicate that replay cache performance is a well-known issue with krb5. It
seems they actually fsync() for every connection, while holding this lock... so
that explains why it was only able to do a small number of connections per
second!
> GSSAPI negotiation single-threaded and very slow under concurrency
> ------------------------------------------------------------------
>
> Key: KUDU-1897
> URL: https://issues.apache.org/jira/browse/KUDU-1897
> Project: Kudu
> Issue Type: Bug
> Components: rpc, security
> Affects Versions: 1.3.0
> Reporter: Todd Lipcon
> Priority: Critical
> Attachments: rcache.png
>
>
> I'm running an Impala workload with ~10 concurrent clients submitting queries
> that take a few seconds each on a 9 node cluster. The master is pegging on
> core in negotiation, and has a bunch of other negotiation threads blocked on
> a mutex within SASL.
> Each negotiation is taking 2-3 seconds due to waiting on this lock.
> This severely restricts connection throughput, and probably needs to be
> addressed for 1.3.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
