What Happens To Application Performance When CPU Utilisation Increases?
October 26th, 2004 by Mark Rittman
There was an interesting
question on Oracle-L the other day about system performance when CPU
utilisation goes up towards 100%:
"I’m not a hardware guy or sys admin person so forgive me if this
is a stupid question. Leaving out all other variables(such as IO), should I
expect performance to be the same in a databse if the server it is riding at
is at 90% cpu usages as opposed to 10%? since there would still be spare
cycles? Or is there a declining returns as you get closing to the maximum
available cpu usage?"
Cary Millsap actually answered in the end and came
back with this answer:
"Good question actually. It depends on how many CPUs are in there.
If you have a copy of "Optimizing Oracle Performance" handy, check
out table 9-3 on page 260. What it says is that you can expect performance to
vary unpredictably on a 1-CPU machine when CPU utilization exceeds 50%. On a
2-CPU machine, the magic number is about 57.7%. On a 4-CPU machine, it’s
66.5%.The table lists "knee in the curve" numbers for CPU counts =
ranging from 1 to 128. The whole section called "Behavior of M/M/m
Systems" (pp254-266) should = be of interest to you."
From looking through the relevant
pages in Cary’s book, it appears that, once you get to the level of CPU
usage specified for each CPU count (50% for a single CPU, 57.7% for 2 CPUs and
so on) average response time not only goes up dramatically, but the variation in
response time differs greatly as well. The chapter also answers the question
about whether it’s better to get a one single fast CPU, or two (or more) slower
CPUs – the answer being that it depends on the nature of your system; single CPU
systems are good for dealing with applications with low request ‘arrival rates’,
and in these circumstances will give superior performance. If your application
typically has lots of concurrent users, with lots of ‘arrivals’ all jockeying
for CPU, multi-CPU systems are usually faster as they reduce the amount of ‘queueing’
for CPU resources. All interesting stuff.
