Subject: Re: Interactive responsiveness under heavy I/O load
To: Steven M. Bellovin <smb@research.att.com>
From: Thor Lancelot Simon <tls@rek.tjls.com>
List: tech-kern
Date: 01/26/2004 19:54:12
On Mon, Jan 26, 2004 at 12:28:57PM -0500, Steven M. Bellovin wrote:
> 
> Thor, I find -current to be almost unusable for interactive work when 
> there's something I/O-intensive running at the same time.  I have
> NEW_BUFQ_STRATEGY enabled, as best I can tell.  Here are my vm. sysctl 

What kind of disk do you have?  The option for NEW_BUFQ_STRATEGY is
stupidly implemented -- it's turned on or off down in the disk driver.
If you happen to be using an 'ld', for instance, there's no #ifdef in
ld.c, so the option does nothing.  ("Don't get me started...")

Are you using softdep?  When Paul removed the fixed limit on the number
of entries in the metadata cache, he accidentally upset the entire house
of cards on which the stability of the I/O system rested, in the softdep
case at least.  The _only_ thing keeping softdep from flooding the queues
with more I/O per second than they could actually handle was that our old
buffer cache implementation imposed an extremely low limit on the number
of pending directory operations at any given time, causing writers (well,
directory-operators, really) to block and preventing them from pouring
as much I/O into the pipe as softdep would let them in one second.

I have a patch, as I said, to spread the I/O out by playing with the
use of "dirdelay".  I can send it to you to try out if you like.

Thor