netdev
[Top] [All Lists]

Re: Xeon smp performance

To: satya srikanth <satyasrikanth2001@xxxxxxxxx>
Subject: Re: Xeon smp performance
From: P@xxxxxxxxxxxxxx
Date: Mon, 08 Mar 2004 15:15:30 +0000
Cc: netdev@xxxxxxxxxxx
In-reply-to: <20040308132354.77437.qmail@web21101.mail.yahoo.com>
References: <20040308132354.77437.qmail@web21101.mail.yahoo.com>
Sender: netdev-bounce@xxxxxxxxxxx
User-agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.6) Gecko/20040124
satya srikanth wrote:
Hi,
I am writing a simple sniffer program that will
sniff packets in a gigabit network. I am using a dual
processor 2 GHz Xeon processor with one Intel Pro-1000
desktop adapter, running 2.4.20-8smp linux. I found that for all packet sizes and network
bandwidth, 2.4.20-8 uniprocessor version that uses
only one processor is performing better than
2.4.20-8smp using two processors in terms of packet
drops and CPU utilization. Each processor in smp is
utilizing more CPU than one processor in uni-processor
case. What is the reason for this peculiar behaviour?
Is it possible for me to use the power of second
processor without adding more NICs. Will I face
similar problems with 2.6 also? Anybody familiar with
these please help me out.

That matches my experience. My chipset was Intel E7501 with 2 intel pro-1000 adapters. I used the base 2.4.20 kernel with latest intel drivers.

I had 2 processes capturing packets and noticed that
2 CPUs is 7% slower than one CPU. Note I was using a non SMP
kernel for the one CPU case.

I'm going to have to look into improving this soon.
First I'm going to look at IRQ and process affinity.

Pádraig.


<Prev in Thread] Current Thread [Next in Thread>