hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Anoop Sam John (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-18002) Investigate why bucket cache filling up in file mode in an exisitng file is slower
Date Mon, 08 May 2017 06:21:04 GMT

    [ https://issues.apache.org/jira/browse/HBASE-18002?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16000320#comment-16000320

Anoop Sam John commented on HBASE-18002:

I mean this.  From SSD, the reads/writes happen as pages and update as blocks.  See here http://codecapsule.com/2014/02/12/coding-for-ssds-part-6-a-summary-what-every-programmer-should-know-about-solid-state-drives/.
When the file is already having data, we will overhead of writing new cache entries.   When
our blocks are not aligned with pages in SSD (mostly yes this way), we will end up have to
rewrite more pages/blocks to add an HFile block entry!
bq.bigger concern is that if there are lot of evictions and new blocks keeps getting filled
up we may still end up with the same problem right?
Hmm true.. Need tests!

> Investigate why bucket cache filling up in file mode in an exisitng file  is slower
> -----------------------------------------------------------------------------------
>                 Key: HBASE-18002
>                 URL: https://issues.apache.org/jira/browse/HBASE-18002
>             Project: HBase
>          Issue Type: Sub-task
>          Components: BucketCache
>    Affects Versions: 2.0.0
>            Reporter: ramkrishna.s.vasudevan
>             Fix For: 2.0.0
> This issue was observed when we recently did some tests with SSD based bucket cache.
Similar thing was also reported by @stack and [~danielpol] while doing some of these bucket
cache related testing.
> When we try to preload a bucket cache (in file mode) with a new file the bucket cache
fills up quite faster and there not much 'failedBlockAdditions'. But when the same bucket
cache is filled up with a preexisitng file ( that had already some entries filled up) this
time it has more 'failedBlockAdditions' and the cache does not fill up faster. Investigate
why this happens. 

This message was sent by Atlassian JIRA

View raw message