On Tue, Mar 27, 2012 at 1:29 AM, Stack <[email protected]> wrote: > On Mon, Mar 19, 2012 at 3:41 AM, Juhani Connolly <[email protected]> wrote: >> Hi, >> >> We're running into a brick wall where our throughput numbers will not >> scale as we increase server counts both using custom inhouse tests and >> ycsb. >> > > Does the above statement still hold? We've moved past the above and > we are now on to 'writes are slow'? > >> We're using hbase 0.92 on hadoop 0.20.2(we also experience the same >> issues using 0.90 before switching our testing to this version). >> >> Our cluster consists of: >> - Namenode and hmaster on separate servers, 24 core, 64gb >> - up to 11 datanode/regionservers. 24 core, 64gb, 4 * 1tb disks(hope >> to get this changed) >> > > You can put the master and namenode on the same machine. > > Yes, more disks are better (see the GBIF blog cited in another thread). > > > >> - load 10m rows > > Are the 10m rows for sure spread across all regions? > > >> Delaying WAL flushes gives a small throughput bump but it doesn't >> scale. >> > > Why does it not scale? > > St.Ack
This was on our old setup, things weren't scaling because there weren't enough regions. I had originally meant to make the other thread because the problem was fundamentally different, sorry for the confusion. In summary the problem now is no longer "not scaling"(because as we increase regions to match the available cpu's it seemingly does, just the base numbers are miserable). Instead it is now "since switching to hdfs 0.23 reads are good and scaling but writes are miserably slow(approx 2000 per region)"
