Re: 6.0 VFS/VM over-runs. Was: Apparent strange disk behaviour in 6.0

From: Julian Elischer <julian_at_elischer.org>
Date: Fri, 29 Jul 2005 16:34:22 -0700
Peter Jeremy wrote:

>[CC list pruned]
>
>On Thu, Jul 28, 2005 at 02:13:44PM -0700, Julian Elischer wrote:
>  
>
>>file data written to cache to await output to disk is allowed to grow to 
>>the extent that running programs are swapped out.
>>(or at least paged out).
>>    
>>
>
>I don't see this on 5.3 either.  In multi-user mode (with lots of
>other processes sitting aroung idle) vmstat shows fr and sr sit at
>around 10,000/sec but there's no paging and a "ps -axl" afterwards
>shows everything mostly still resident.  The active memory (avm)
>hardly changes but free memory drops from 233MB to 20MB over 8 seconds
>than oscillates between about 16MB and 30MB.  Disk I/O remains at
>35-40MB/sec (ie disk limited).
>
>This is an Athlon XP-1800 512MB RAM writing a 10GB file onto an ATA disk.
>PREEMPTION is enabled.
>
>Since Marc Olzheim reports it happens with 5-STABLE and Julian reports
>it doesn't happen on 5.4 and I can't reproduce it on 5.3, one option
>would seem to be that it was MFC'd to 5 at some point since 5.4 was
>branched.  Unfortunately, a quick scan through the RELENG_5 commits
>doesn't reveal anything that might be relevant.
>
>The other possibility is that it's RAM size related.  The vmstat
>output Julian posted demonstrating the problem was on a system with a
>fair bit of RAM, but his 5.4 system is presumably smaller and my
>system definitely is.
>  
>

actually the 5.4 system (booted off a CD in single user mode) had the 
same (1GB) ram.

The difference was an IDE idsk vs a SCSI attached RAID but I doubt that 
would make a difference to
this problem.. free memory went down to 8MB or so and just stayed there 
and old pages from
the file were ejected insted by the clock algorythm. The dd was slowed 
down and idle 
time on the system increased.  i.e. backpreasure was applied correctly.
Received on Fri Jul 29 2005 - 21:34:24 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:40 UTC