Yes, that's with 250 threads running, and yes, cpu/memory are mostly around 25% and 40% percent. So we are not pegged there. There are 2 phyical cpus and 4 core each, with hyper-threading on, which gives us 16 logical cpus.
The hbase cluster has 3 nodes, each with the same cpu/hardware configuration above. The processor is intel E5620 i think. What's the best way to do network packet capture? thanks On Mon, Feb 28, 2011 at 10:39 AM, tsuna <tsuna...@gmail.com> wrote: > On Mon, Feb 28, 2011 at 10:22 AM, benchmark <zsl...@gmail.com> wrote: > > Here is the vmstat 1 results. Is that a lot of context switches? this is > a > > box with 8 cpus and hyperthreading enabled on centOS. > > Is this output with 250 threads running? > When you say "8 cpus", do you mean 1 physical CPU with 4 cores with > HyperThreading enabled? What's the model number of that CPU? > > You're seeing only ~1000 context switches per second per core, that's > not completely unreasonable. Your server seems more idle than > anything else... I'd suggest looking at a packet capture to see what > the network latency looks like, I'm guessing you'll see that the slow > responses you're seeing are indeed really slow and that the problem > comes from the server-side. > > What's your HBase cluster like? How many nodes, what hardware, what > configuration? What's you block cache hit ratio? > > -- > Benoit "tsuna" Sigoure > Software Engineer @ www.StumbleUpon.com >