On 08/ 7/16 04:22 PM, Joshua M. Clulow wrote:
On 6 August 2016 at 21:10, Ian Collins <[email protected]> wrote:
Over the weekend I have re-purposed my Solaris build machine to be a compute
node. I was expecting, based on raw numbers and past experience, it to
build the current C++ code base I'm working with in 50-60% of the time taken
by either of my current build slaves.
Reality hasn't matched expectations big time!
An existing system, using the same motherboard but slower CPUs does the
build in about 9 minutes. After 40 minutes the build was only a little over
half done on the "fast" box...
I noticed this comparing prstat on the two hosts:
All the "wait" states look odd to me.
You can get a finer break down of different thread states by using the
"-mL" options to prstat(1M). As per the manual page, the new columns
(LAT, LCK, SLP, etc) are a percentage of time spent in each of the
micro states that the system tracks for a particular thread.
I think I've earned the dunce's hat for the week.
The zone on the new box had a cpu_cap of 100, so it was trying to run as
a one core box... Which is what the "wait" states in prstat where
telling me all along.
Change that to 3200 and it builds in half the time as expected.
--
Ian.
-------------------------------------------
smartos-discuss
Archives: https://www.listbox.com/member/archive/184463/=now
RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00
Modify Your Subscription:
https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb
Powered by Listbox: http://www.listbox.com
