[WBEL-users] Process CPU Binding

Vincent Raffensberger Vincent.Raffensberger at dtn.com
Fri Aug 5 00:15:54 CDT 2005


I don't believe it's possible to bind a process or thread to a specific physical processor (at least an experimental kernel patch).
Even if you could, I'd be skeptical of any performance benefit.

Is it possible to accomplish what you need by binding an irq to a specific CPU?

I work with many high volume multicast feeds which require me to set the processor affinity of the eth irq to seperate processors.  
The reason for this is that if multiple processors are servicing the interrupts for a device, the packets (udp) could get out of sequence under high loads.

One solution is to disable APIC by booting with the 'noapic' option, but that has negative effects for other aspects of the system.
I choose to bind the ethernet irqs to physical processors instead.

If that would work for you, I have a script to do this.  I'm sure you could modify it to meet your needs:
ftp://dvdrchive.dtniq.com/scripts/eth_affinity


-----Original Message-----
From: whitebox-users-bounces at beau.org on behalf of Adam Dixon
Sent: Thu 8/4/2005 8:15 PM
To: Whitebox-users at beau.org
Subject: [WBEL-users] Process CPU Binding
 
This is probably not a WBEL question, however I am running WBEL on a 
number of servers that run a process of heavy cpu use, where I am told 
that binding it to a particular cpu may provide a performance increase 
due to being no delay in deciding which cpu to process on during it 
running.

So is it possible to bind a process to a particular CPU on an SMP dual 
cpu system (x86). Say lock the process id to cpu0 and then another 
process id to cpu1.

Anyone who can provide information on if this is possible.

Adam
_______________________________________________
Whitebox-users mailing list
Whitebox-users at beau.org
http://beau.org/mailman/listinfo/whitebox-users

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://beau.org/pipermail/whitebox-users/attachments/20050805/d61fb587/attachment.htm


More information about the Whitebox-users mailing list