At Wed, 02 May 2012 13:31:16 +0900, MORITA Kazutaka wrote: > > At Tue, 1 May 2012 20:06:12 +0800, > Yibin Shen wrote: > > > > Hi, list: > > > > last weeks, > > we use 96 physical server to simulate 1k nodes sheepdog cluster's > > performance (with zookeeper driver) > > and here is the report: > > > > http://sheepdog.taobao.org/people/zituan/sheepdog1k.html > > Great report, thanks! Can you add the report to the Sheepdog wiki? > https://github.com/collie/sheepdog/wiki > > I'd like to see how the result will change if we remove the '-a' > option from the sheep options because the option doesn't follow block > device semantics. Another acceptable option would be: - use both '-a' and '-D' for sheep options, and - disable write-caching of the local disk This ensures that the flushed data is written to the local disk at least, though the data is not replicated to multiple nodes until the backgroud thread replicates it. Thanks, Kazutaka |