[
https://issues.apache.org/jira/browse/HBASE-2451?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12857592#action_12857592
]
stack commented on HBASE-2451:
------------------------------
I've been running this patch up on cluster. Here is what I see now:
{code}
2010-04-15 16:23:47,892 DEBUG org.apache.hadoop.hbase.io.hfile.LruBlockCache:
Cache Stats: Sizes: Total=6.0434265MB (6336992), Free=602.95654MB (632245792),
Max=609.0MB (638582784), Counts: Blocks=177, Access=1022416, Hit=356196,
Miss=666220, Evictions=0, Evicted=0, Ratios: Hit Ratio=34.83865559101105%, Miss
Ratio=65.16134142875671%, Evicted/Run=NaN
{code}
First it takes a while for the ratios to build up... it start out at 100%
misses. The above is skewed some by the fact that the cache is so big relative
to the amount of data that needs to caching (this is a big upload and only meta
data is being cached). Its also skewed because meta changing a lot durinng
this upload. This is better than what we had previous but looks like more work
can be done in here (this table is pegged in_memory). Will commit this after
review for 0.20.4. We can do other improvements elsewhere.
> .META. by-passes cache; BLOCKCACHE=>'false'
> -------------------------------------------
>
> Key: HBASE-2451
> URL: https://issues.apache.org/jira/browse/HBASE-2451
> Project: Hadoop HBase
> Issue Type: Bug
> Reporter: stack
> Priority: Blocker
> Fix For: 0.20.4, 0.20.5, 0.21.0
>
> Attachments: meta.txt, meta3.txt
>
>
> In a new install, if I describe '.META.', it says:
> {code}
> DESCRIPTION
> ENABLED
> {NAME => '.META.', IS_META => 'true', MEMSTORE_FLUSHSIZE => '16384', F true
>
> AMILIES => [{NAME => 'historian', COMPRESSION => 'NONE', VERSIONS => '
>
> 2147483647', TTL => '604800', BLOCKSIZE => '8192', IN_MEMORY => 'false
>
> ', BLOCKCACHE => 'false'}, {NAME => 'info', COMPRESSION => 'NONE', VER
>
> SIONS => '10', TTL => '2147483647', BLOCKSIZE => '8192', IN_MEMORY =>
>
> 'false', BLOCKCACHE => 'true'}]}
> {code}
> BLOCKCACHE is 'true' for the 'info' family (Yes historian is still in 0.20
> branch).
> But, if I add logging to hfile and storefile and store, blockcache is 'false'
> -- there is no cache constructed for use by the hfile.
> This is killing cluster performance.
> It looks like a problem parsing the 'true' value in columnfamily. I'll put
> up a patch in the morning. Meantime, marking as blocker on 0.20.4.
--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
https://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira
