netdev
[Top] [All Lists]

Re: More measurements

To: andrewm@xxxxxxxxxx (Andrew Morton)
Subject: Re: More measurements
From: kuznet@xxxxxxxxxxxxx
Date: Tue, 30 Jan 2001 22:55:15 +0300 (MSK)
Cc: netdev@xxxxxxxxxxx
In-reply-to: <3A75785A.42B9E7CE@uow.edu.au> from "Andrew Morton" at Jan 29, 1 05:15:07 pm
Sender: owner-netdev@xxxxxxxxxxx
Hello!

> - Changing eepro100 to use IO operations instead of MMIO slows
>   down this dual 500MHz machine by less than one percent at
>   100 mbps.  At 12,000 interrupts per second. Why all the fuss
>   about MMIO?

Interesting. 0.9% is ridiculously small, but 2.5% is some big number. 8)8)

Actually, I am not sure that mmio improves something in your case.
Even these 0.9% can be random or systematic error.
It works if it caches access to io (not a fact in your case) and
if bus load is high. You have no bus load, so that you do not see
any bus congestion effects.


> - Bonding the 905's interrupt to CPU0 slows things down slightly.

I am afraid, you do mistake here. Bounding IRQ without bounding
sending/receiving threads results either in improvement or in disaster
depending on cpu, where thread is scheduled.

Generally, setting irq affinity is valid only if host is router
or if services and sockets are bound too.

Alexey

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