Re: benchmarking squid on solaris/x86

From: Henrik Nordstrom <hno@dont-contact.us>
Date: Mon, 1 Apr 2002 23:09:51 +0200

On Thursday 21 March 2002 22:15, Lincoln Dale wrote:

> feel free to convert squid to use them and you'll become a
> non-believer in SIGIO like i am.

The Linux RT Signal API is not sutiable as a general I/O API for an
application due to it's rather complex nature and odd effects. But my
opinion after playing quite extensively with it is that it can be
embedded in a library suitable for application development.

The good thing of the API is that is is relatively simple for the
kernel to deal with in a reasonably efficient manner.

But it do not attempt to address data copying or socket passing, only
event signalling.

> my experience comes from taking a few weeks to convert a commercial
> proxy to use SIGIO and being underwhelmed by the new bottlenecks
> and races inherent in SIGIO.

I do agree SIGIO is not the perfect interface, but I don't regard it
as worse than /dev/poll.. neither better for that matter, only
different.

What I did learn when I last played with Linux RT signal I/O (which
is inherently differnet in nature from plain SIGIO) is that the Linux
kernel had a lot to go before making full use of the RT Signal I/O
potentials. There are many not too uncommon I/O patterns that can
easily bring out the worst aspects of the model in the current
implementation but do not really need to. This was during
2.4.TEST-something. Have not studied the details lately.

Cheers
Henrik
Received on Mon Apr 01 2002 - 14:22:28 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:14:55 MST