Re: 2.4.1pre8 slowdown on dbench tests

[Andi Kleen]

On Fri, Jan 19, 2001 at 02:40:23AM +0100, Jens Axboe wrote:
> On Fri, Jan 19 2001, Andrea Arcangeli wrote:
> > On Thu, Jan 18, 2001 at 03:17:13PM -0200, Marcelo Tosatti wrote:
> > > Jens, can be the -blk patch the reason for the slowdown I'm seeing?
> > 
> > This heuristic is way too aggressive:
> > 
> > 	/*
> > 	 * Try to keep 128MB max hysteris. If not possible,
> > 	 * use half of RAM
> > 	 */
> > 	high_queued_sectors = (total_ram * 2) / 3;
> > 	low_queued_sectors = high_queued_sectors - MB(128);
> > 	if (low_queued_sectors < 0)
> > 		low_queued_sectors = total_ram / 2;
> > 
> > 	/*
> > 	 * for big RAM machines (>= 384MB), use more for I/O
> > 	 */
> > 	if (total_ram >= MB(384)) {
> > 		high_queued_sectors = (total_ram * 4) / 5;
> > 		low_queued_sectors = high_queued_sectors - MB(128);
> > 	}
> > 
> > 2/3 of ram locked down in the I/O queue is way too much. 1/3 should be
> > ok. big RAM machines needs way less than 1/3 locked down.
> 
> Yes I agree, that values should probably be tweaked a bit. I'll
> try and squeeze some testing in to generate the best possible
> values.

Please also add a sysctl. I always feel uncomfortable with such hardcoded
heuristics. There tends to be always another workload where the heuristic
fails and manual tuning is useful. 

-Andi