Subject: Re: MAXPHYS
To: Thor Lancelot Simon <tls@rek.tjls.com>
From: Chuck Silvers <chuq@chuq.com>
List: tech-kern
Date: 02/05/2001 01:24:56
On Sun, Feb 04, 2001 at 05:40:10PM -0500, Thor Lancelot Simon wrote:
> Note that it was essentially pointless to do so before the advent of UBC
> because increasing MAXBSIZE beyond 64K would have been horrifically
> inefficient.  At the moment, there seem to be some performance problems
> with UBC in which it doesn't cluster up to MAXPHYS anyway, even with a
> 64k MAXPHYS, but those will surely be fixed soon; so it would seem to me
> that replacing the constant MAXPHYS with a scheme that propagates this
> kind of information through the device tree would be a good idea.

actually UBC does cluster up to MAXPHYS.  I just did some experiments
with dd on today's -current and the average transfer size according to
iostat was around 63k.  this has been true since long before UBC was
integrated into -current.

UVM/UBC currently has a number of other assumptions limiting the i/o size
to 64k, but those will be fixed eventually.

-Chuck