On Nov 3, 2009, at 11:16 AM, Les Mikesell <lesmikesell at gmail.com> wrote: > Ross Walker wrote: >> On Nov 3, 2009, at 8:22 AM, Alan McKay <alan.mckay at gmail.com> wrote: >> >>> Hey folks, >>> >>> We've got some new hardware and are trying to figure out what best >>> to >>> do with it. Either run CentOS right on the bare metal, or >>> virtualize, or several combination options. Mainly looking at : >>> >>> - CentOS on bare metal >>> - CentOS on ESXi 4.0 with local disk >>> - CentOS on ESXi with 1 VM running Openfiler to serve disk to other >>> VMs >>> >>> And want to benchmark these 3 scenarios >>> >>> So far all we have is a dd-based disk IO benchmark. >>> >>> What else can you all recommend. >>> >>> BTW, we also ideally want to try each of the above with a Postgres >>> DB >>> as well (and once without) >> >> Take a look at iozone it is a little dated, but still good and is >> cross platform so you can can have a level playing field between >> Linux, Windows and Solaris. >> >> I'd be interested in hearing your results. I will tell you it will be >> slower then bare metal, that is just a fact of life, but with >> adequate >> hardware it may be more then acceptable. > > Odds are pretty good that the limiting factor will be disk head seek > time regardless of the software layers above, although this may not > show > up in single-tasking benchmarks like it does in typical use. The > disk/partition layout and the way competing tasks are distributed > across > them are likely to have more effect on speed than an extra logical > disk > software wrapper. Disk head seek time is always the limiting factor, but partition misalignment can cause significant artificial limits if blocks overlap page size boundaries and the next block and maybe chunk needs to be read/written. I have seen it limit random throughput by as much a 5 times on small blocks. -Ross