Philip Beevers wrote:
I've been musing recently on capacity planning for systems which see
very peaky workloads. For example, in the financial markets, a very busy
day can be much, much busier than an average day; this means you need to
plan for relatively low utilisation on the average day, in order to give
headroom for the peak day.
Of course, I also want my datacentre to be power efficient - even if
need 4 CPUs to cover the peak day, I want lower power consumption on the
average day, where I only really need 2 CPUs. Obviously AMD's PowerNow!
technology delivers this, by reducing CPU power usage when I don't need
peak performance.
I'm interested in how this interacts with monitoring tools - for
example, if my CPUs works out it is only 10% utilised, and thus decides
to slow itself down, will it vmstat still show 90% idle? [I think
ideally it would, so I've got a true idea of how much headroom there is
on my system]
As I understand things, however, this *won't* be the case. For example,
I think microstate accounting works by simply taking timestamps when an
LWP switches between states, which implies it can't cater for the CPU
speed changing dynamically.
Can anyone offer some insight on how this works?
Well, it's more confusing than that, even... because the amount of
speedup on a task may have nothing or a lot to do with clock
frequency. If my workload is always blocked on RAM, the
clock freq of the cpu has far less effect than if my app is
running out of the L1 caches.
Our initial plans for PowerNow support will maintain some
% of idle cpu. If your app is very latency sensitive but
spends most of it's time sleeping, you'll want to use a different
policy.
- Bart
--
Bart Smaalders Solaris Kernel Performance
[EMAIL PROTECTED] http://blogs.sun.com/barts
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
