[ https://issues.apache.org/jira/browse/HBASE-11331?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Nick Dimiduk updated HBASE-11331: --------------------------------- Attachment: lazy-decompress.02.0.pdf Attaching some metrics after running v02 with the new flag enabled/disabled. The test cluster is a single machine, configured with 12g heap and 0.6 for blockcache, so slighly over 7g mem available for cache. The TestTable is snappy compressed, written with PE --size=100. Shell {{status 'simple'}} reports {{compressionRatio=0.2437}}. RandomRead test is run with --size=25, so with feature enabled, we should be at ~6g of compressed blocks. h3. client perceived perf With flag=false, I see slightly higher throughput (~7.5k vs ~8k) and slightly less gc (360ms vs 325ms). h3. system load CPU load looks slightly higher with flag=true (user time diff is negligible, system time more noticeable, no difference in iowait). h3. blockcache utilization There are definitely cache evictions happening with flag=false that do not happen with flag=true. The number of blocks cached is definitely higher with flag=true (~17.5k vs ~21.5k). h3. IO No difference in reported IO requests. Probably it's the OS buffer cache to the rescue here. > [blockcache] lazy block decompression > ------------------------------------- > > Key: HBASE-11331 > URL: https://issues.apache.org/jira/browse/HBASE-11331 > Project: HBase > Issue Type: Improvement > Components: regionserver > Reporter: Nick Dimiduk > Assignee: Nick Dimiduk > Attachments: HBASE-11331.00.patch, HBASE-11331.01.patch, > HBASE-11331.02.patch, HBASE-11331LazyBlockDecompressperfcompare.pdf, > lazy-decompress.02.0.pdf > > > Maintaining data in its compressed form in the block cache will greatly > increase our effective blockcache size and should show a meaning improvement > in cache hit rates in well designed applications. The idea here is to lazily > decompress/decrypt blocks when they're consumed, rather than as soon as > they're pulled off of disk. > This is related to but less invasive than HBASE-8894. -- This message was sent by Atlassian JIRA (v6.2#6252)