Hbase 0.90.1, hadoop0.20.2+append Put op use putlist(2000). About 5000 regions, 512MB max size, 50 ipc handler each RS. Totally 8 RS. Both put and scan have 500 threads.
Because we use random start key to scan, cache hitratio is very low. 2011-04-19 20:16:52,038 DEBUG org.apache.hadoop.hbase.io.hfile.LruBlockCache: LRU Stats: total=1.23 GB, free=378.14 MB, max=1.59 GB, blocks=19741, accesses=10958888, hits=2870, hitRatio=0.02%%, cachingAccesses=227733, cachingHits=2870, cachingHitsRatio=1.26%%, evictions=79, evicted=205121, evictedPerRun=2596.46826171875 2011-04-19 20:21:52,038 DEBUG org.apache.hadoop.hbase.io.hfile.LruBlockCache: LRU Stats: total=1.27 GB, free=329.64 MB, max=1.59 GB, blocks=20511, accesses=11097514, hits=4399, hitRatio=0.03%%, cachingAccesses=326115, cachingHits=4399, cachingHitsRatio=1.34%%, evictions=116, evicted=301193, evictedPerRun=2596.491455078125 Zhou Shuaifeng(Frank) -----邮件原件----- 发件人: [email protected] [mailto:[email protected]] 代表 Stack 发送时间: 2011年4月22日 11:11 收件人: [email protected] 抄送: Yanlijun 主题: Re: Is there any influence to put performance when "Block cache LRU eviction" occur? On Thu, Apr 21, 2011 at 7:00 PM, Zhoushuaifeng <[email protected]> wrote: > I'm running random scan and put performance test at the same time, found > that put performance descend a lot. Don't know why. > Looking into the logs, found that "Block cache LRU eviction" occured almost > once per second. But it only last about 10 ms, so, will it bring so much > influence to put performance? And how to improve it? > I would think that there would be little to no relation between put and block cache evictions. > hbase-root-regionserver-c3s5.log.2011-04-19:2011-04-19 21:00:00,067 DEBUG > org.apache.hadoop.hbase.io.hfile.LruBlockCache: Block cache LRU eviction > started; Attempting to free 163.35 MB of total=1.36 GB > hbase-root-regionserver-c3s5.log.2011-04-19:2011-04-19 21:00:00,081 DEBUG > org.apache.hadoop.hbase.io.hfile.LruBlockCache: Block cache LRU eviction > completed; freed=163.39 MB, total=1.2 GB, single=973.98 MB, multi=400.97 MB, > memory=0 KB > It looks like each time it runs its evictly ~10 percent. You are doing lots of scans that go via the block cache? Regards the puts, tell us more. How many clients, what type of clients, regions, regionservers, versions of hbase and hadoop, etc. You've seen this page? http://hbase.apache.org/book.html#performance St.Ack
