I have couple of standalone Zookeeper servers that perform very
differently. I collected a block io trace using the following:
perf record -a -e "block:block_rq_issue,block:block_rq_complete"
On the fast server, I see lots of sequences that look like this:
java 32429 [005] 786479.507611: block_rq_issue: 8,0 WS 0 ()
187774376 + 8 [java]
init 1 [001] 786479.514222: block_rq_complete: 8,0 WS ()
187774376 + 8 [0]
On the slow server I see these additional events:
java 2209 [007] 254523.025322: block:block_rq_issue: 8,16 FWS
0 () 18446744073709551615 + 0 [java]
init 1 [007] 254523.044166: block:block_rq_complete: 8,16
WS () 0 + 0 [0]
How do I determine what is triggering the flushes on the slow system?
For what it's worth, the slow server is following the best practice
recommendation of putting the ZK log on a dedicated drive. The fast
server is not, because we only have a single disk on that system.
Henry May
IBM InfoSphere Streams Performance
hjmay@us.ibm.com
720-342-8873
Tie: 963-8873
|