Can I set the vmstat collection to average less
and collect more data points? The users here
don't agree with the CPU Utilization Graph,
they don't like the way the Average doesn't
fluctuate much.
Basically, they are comparing the Load graph
with the Utilization Graph and want to see the
graph correlate more with the Load Graph.
My understanding is that the Utilization Graph
is using a 5 minute cpu idle average from vmstat.
Can I change vmstat to take a 15 second average
as an example, and do more counts so more datapoints
are graphed?
Any suggestions would be appreciated.
Thanks.James
Why do users think that the load average graph (which measures the average length of the CPU run queue) should correlate with the CPU utilization (busy time divided by available time)? The values in each graph are calculated very differently.
Just wondering...
GLH
From: James Wade [mailto:jkwade at futurefrontiers.com]
Sent: Friday, February 16, 2007 12:31 PM
To: hobbit at hswn.dk
Subject: [hobbit] CPU Utilization, Less Averaging
Can I set the vmstat collection to average less
and collect more data points? The users here
don't agree with the CPU Utilization Graph,
they don't like the way the Average doesn't
fluctuate much.
Basically, they are comparing the Load graph
with the Utilization Graph and want to see the
graph correlate more with the Load Graph.
My understanding is that the Utilization Graph
is using a 5 minute cpu idle average from vmstat.
Can I change vmstat to take a 15 second average
as an example, and do more counts so more datapoints
are graphed?
Any suggestions would be appreciated.
Thanks...James
True, and I've checked, by looking at what the cpu utilization
should be compared to the load average. Everything looks fine.
To be honest, it's user perception. They want to see a
graph that doesn't average but shows the peaks and troughs.
There is another tool here that does cpu utilization, and it
doesn't average and shows more peaks and troughs.
So, I just want to drop the averaging so that it takes more samples,
and less averaging.
As an example, they are looking at load doubling,
but the cpu utilization graph shows no change. However, because
the box is a mutli-cpu box, the load doubling in this case only
means an average increase in utilization by 2%. However, if the
load goes down within 5 minutes, the 2% isn't registered, so
the graph gives a constant cpu average.
What I have seen is the Utilization may go up by 10% or 15%
for only a few seconds, then drop down, (As seen by TOP)
This normally occurs when a large job hits. Because the cpu
utilization is an average on a 5 minute basis, there is not
normally a change in the graph.
I figure if I drop the sample down, I'll get more peaks and troughs
and make them happy. How can I do this in Hobbit?
James
From: Hubbard, Greg L [mailto:greg.hubbard at eds.com] Sent: Friday, February 16, 2007 1:30 PM To: hobbit at hswn.dk Subject: RE: [hobbit] CPU Utilization, Less Averaging
Why do users think that the load average graph (which measures the average length of the CPU run queue) should correlate with the CPU utilization (busy time divided by available time)? The values in each graph are calculated very differently.
Just wondering...
GLH
From: James Wade [mailto:jkwade at futurefrontiers.com] Sent: Friday, February 16, 2007 12:31 PM To: hobbit at hswn.dk Subject: [hobbit] CPU Utilization, Less Averaging
Can I set the vmstat collection to average less
and collect more data points? The users here
don't agree with the CPU Utilization Graph,
they don't like the way the Average doesn't
fluctuate much.
Basically, they are comparing the Load graph
with the Utilization Graph and want to see the
graph correlate more with the Load Graph.
My understanding is that the Utilization Graph
is using a 5 minute cpu idle average from vmstat.
Can I change vmstat to take a 15 second average
as an example, and do more counts so more datapoints
are graphed?
Any suggestions would be appreciated.
Thanks.James
On Fri, Feb 16, 2007 at 01:55:41PM -0600, James Wade wrote:
To be honest, it's user perception. They want to see a graph that doesn't average but shows the peaks and troughs.
Have you pointed out that the load average graph is based on a "5 minute average" number calculated by the OS ?
Henrik
On Fri, Feb 16, 2007 at 12:31:16PM -0600, James Wade wrote:
Can I set the vmstat collection to average less and collect more data points?
You can change the collection to run every 15 seconds, but you will have to change the RRD file definition also so it knows that you intend to feed it data that often. Otherwise it will just discard your extra datapoints.
The users here don't agree with the CPU Utilization Graph, they don't like the way the Average doesn't fluctuate much.
Basically, they are comparing the Load graph with the Utilization Graph and want to see the graph correlate more with the Load Graph.
Your users are comparing apples and oranges - those two numbers have very little to do with each other.
Here's the explanation of the "load average" that you see in the "Load" graph:
System load averages is the average number of processes that are
either in a runnable or uninterruptable state. A process in a
runnable state is either using the CPU or waiting to use the CPU.
A process in uninterruptable state is waiting for some I/O access,
eg waiting for disk. The averages are taken over the three time
intervals. Load averages are not normalized for the number of
CPUs in a system, so a load average of 1 means a single CPU system
is loaded all the time while on a 4 CPU system it means it was idle
75% of the time.The CPU utilisation graph IS normalized over the number of CPUs - that in itself makes a difference (unless you have single CPU systems). But more importantly, a process will show up on the "CPU utilisation" graph ONLY when it is using CPU time; and it will show up on the "load average" graph when it is using CPU time, AND when it is waiting to be scheduled for some CPU time, AND while it is waiting for I/O to complete.
Since most server tasks are very I/O bound you will often see little correlation between the two graphs. In fact, if the correlation becomes too close it is usually a sign that your system needs more CPU power.
Regards, Henrik
participants (3)
-
greg.hubbard＠eds.com
-
henrik＠hswn.dk
-
jkwade＠futurefrontiers.com