Benchmarking 6.0

[John Baldwin]

On Saturday 04 June 2005 04:02 pm, Robert Watson wrote:
> On Thu, 2 Jun 2005, Singh, Vijay wrote:
> > Hello, I am trying to run network benchmarking using netperf. I have a
> > snapshot of "systat -vmstat 1". Is it OK to see those large number of
> > interrupts (~ 1998) for the lapics?
>
> That's a property of the increase in HZ to 1000 from 100, and that on 6.x,
> each lapic generates HZ interrupts/sec to the individual processors,
> rather than the programmable timer deliverable at HZ to a particular (or
> round-robin'd) CPU, then that CPU broadcasting the clock tick to the other
> CPUs using IPIs.  The rate you see should be approximately HZ * #cpus, so
> about 2000 for a dual-processor system (which is indeed about 1998).
> While lowering HZ will lower the overhead associated with processing timer
> ticks, it will also reduce timer granularity, which can affect performance
> in other ways (for example, TCP may behave better with finer granularity
> timeouts).

Actually, the lapic timer runs at hz * 2 on each CPU.  The interrupts are 
local to each CPU though.

> One of the things we don't show in vmstat -systat that we probably should
> is per-CPU IPI rates.  I've CC'd John in the hopes of convincing him to
> add that to his todo list :-).

Ah yes.

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org