I can think of at least 2 clusters running 32GB boxes with single Cassandra processes on each. (16 seems to be more common.) At 64 I would seriously consider multiple processes per machine. You'd want to configure a Snitch such that same-machine boxes were considered the same rack, there is no separate closeness level of same machine.
At 32 I think you're fine with one process. Watch for latency spikes and see how it goes. I would run raid 10 on the data disks if you can afford giving up the space, otherwise raid0. I don't know that anyone's tested raid5. On Sun, May 23, 2010 at 3:30 PM, Aaron McCurry <amccu...@gmail.com> wrote: > I am planning on setting up a Cassandra cluster on a small 16 node cluster > (possibly 32 way). Each machine has 8 cores 32 Gig of ram and 8 hds. My > first thought is to setup one of those hds for the commit log, 6 for data > and leave one for the OS. However I do have a concern about best utilizing > my memory, should I run a larger heap? Should I run several cassandra > processes on the same box? > My concern about the larger heap is because GC's typically get slower. And > if I run several procs, does cassandra realize that it's the same box for > replication purposes? > I do have other hd conf options, hardware RAID 0,1,or 5. > Just looking for some general configuration options as well as some real > world successes with similarly sized hardware. Thanks! > Aaron -- Jonathan Ellis Project Chair, Apache Cassandra co-founder of Riptano, the source for professional Cassandra support http://riptano.com
