Prev: problem with pcl duplex printing
Next: grep to hunt for uniqueness of a id number
From: Thomas Womack on 21 Apr 2010 11:53
I have a dual-quad-core Opteron system, with ubuntu-9.10 installed, on
which I run eight CPU-bound tasks at a time. They run for about 24
hours at a time.

Often I notice that some of the tasks are progressing slowly, and find
(with mpstat -P ALL 2) that the scheduler is leaving one core free,
and running two tasks on one of the other cores. I can fix this with
taskset, but I don't quite understand why it happens at all. Is there
a way to fix it?

I have other quad-core systems with ubuntu-9.10, on which I run four
tasks at a time and always find that they have been distributed one
per core in the obvious way.

Tom
From: Robert Heller on 21 Apr 2010 13:31
At 21 Apr 2010 16:53:31 +0100 (BST) Thomas Womack <twomack(a)chiark.greenend.org.uk> wrote:

>
> I have a dual-quad-core Opteron system, with ubuntu-9.10 installed, on
> which I run eight CPU-bound tasks at a time. They run for about 24
> hours at a time.
>
> Often I notice that some of the tasks are progressing slowly, and find
> (with mpstat -P ALL 2) that the scheduler is leaving one core free,
> and running two tasks on one of the other cores. I can fix this with
> taskset, but I don't quite understand why it happens at all. Is there
> a way to fix it?

I *suspect* one or another of these things are happening:

You do have other processes on the system (eg various background
daemons) that are grabbing cycles from time to time. This might 'bump'
one or another of your CPU-bound tasks from time-to-time and might end
up with one sharing a core with another CPU-bound task.

One of your CPU-bound tasks does some sort of I/O request (even if only
to write a one-line message to a log file). This could cause it to be
bumped.

You are running out of physical RAM and somebody had a page fault. Opps...

>
> I have other quad-core systems with ubuntu-9.10, on which I run four
> tasks at a time and always find that they have been distributed one
> per core in the obvious way.

This sounds more like luck than anything else.

>
> Tom
>

--
Robert Heller -- 978-544-6933
Deepwoods Software -- Download the Model Railroad System
http://www.deepsoft.com/ -- Binaries for Linux and MS-Windows
heller(a)deepsoft.com -- http://www.deepsoft.com/ModelRailroadSystem/

 | 
Pages: 1
Prev: problem with pcl duplex printing
Next: grep to hunt for uniqueness of a id number