Limiting program swap

Limiting program swap

[Bill Davidsen]

I have several machine of various memory sizes which suffer from really 
poor performance when doing backups. This appears to be because all the 
programs other than the backup quickly get swapped to make room for i/o 
buffers.

Is there some standard portable way to prevent this, either by reserving 
some memory for programs which will not get swapped regardless of i/o 
pressure, or alternatively limiting the total memory used for i/o 
buffers, dcache, and similar things?

I did a crude hack for 2.4.17, but if I'm missing some obvious trick I'd 
rather not do something which can't go in the mainline kernel. Anyone 
care to show me what I missed, or is this just a characteristic of Linux?

-- 
    -bill davidsen (davidsen@tmr.com)
"The secret to procrastination is to put things off until the
  last possible moment - but no longer"  -me

Re: Limiting program swap

[Marc-Christian Petersen]

On Wednesday 08 December 2004 20:07, Bill Davidsen wrote:

> I did a crude hack for 2.4.17, but if I'm missing some obvious trick I'd
> rather not do something which can't go in the mainline kernel. Anyone
> care to show me what I missed, or is this just a characteristic of Linux?

Maybe you show us your hack?

ciao, Marc

Re: Limiting program swap

[William Lee Irwin III]

On Wed, Dec 08, 2004 at 02:07:36PM -0500, Bill Davidsen wrote:
> I have several machine of various memory sizes which suffer from really 
> poor performance when doing backups. This appears to be because all the 
> programs other than the backup quickly get swapped to make room for i/o 
> buffers.
> Is there some standard portable way to prevent this, either by reserving 
> some memory for programs which will not get swapped regardless of i/o 
> pressure, or alternatively limiting the total memory used for i/o 
> buffers, dcache, and similar things?
> I did a crude hack for 2.4.17, but if I'm missing some obvious trick I'd 
> rather not do something which can't go in the mainline kernel. Anyone 
> care to show me what I missed, or is this just a characteristic of Linux?

This appears at least superficially related to
/proc/sys/vm/vfs_cache_pressure in 2.6.x-mm (possibly also mainline)


-- wli

Re: Limiting program swap

[Jesper Juhl]

Bill Davidsen wrote:
> I have several machine of various memory sizes which suffer from really 
> poor performance when doing backups. This appears to be because all the 
> programs other than the backup quickly get swapped to make room for i/o 
> buffers.
> 
> Is there some standard portable way to prevent this, either by reserving 
> some memory for programs which will not get swapped regardless of i/o 
> pressure, or alternatively limiting the total memory used for i/o 
> buffers, dcache, and similar things?
> 

I'm wondering if turning the /proc/sys/vm/swappiness knob would help, 
but I'll honestly admit that I don't know.

-- 
Jesper Juhl

Re: Limiting program swap

[Con Kolivas]

Jesper Juhl writes:

> Bill Davidsen wrote:
>> I have several machine of various memory sizes which suffer from really 
>> poor performance when doing backups. This appears to be because all the 
>> programs other than the backup quickly get swapped to make room for i/o 
>> buffers.
>> 
>> Is there some standard portable way to prevent this, either by reserving 
>> some memory for programs which will not get swapped regardless of i/o 
>> pressure, or alternatively limiting the total memory used for i/o 
>> buffers, dcache, and similar things?
>> 
> 
> I'm wondering if turning the /proc/sys/vm/swappiness knob would help, 
> but I'll honestly admit that I don't know.

No, there comes a point where the vm pressure overrides the value of 
swappiness and we start swapping. It is better in 2.6.7 and before with 
issues yet to be resolved in the latest mainline trees. My tree (-ck) has a 
slightly different tuning and a hard maplimit option for just such a reason 
but are unlikely to be of interest to the mainline maintainers so I've never 
promoted them.

Cheers,
Con