Re: bounce buffers and i/o schedulers with aacraid

[markw@osdl.org]

On 13 Aug, Andrew Morton wrote:
> markw@osdl.org wrote:
>>
>> We're still trying to avoid bounce buffers with the aacraid driver and
>> noticed something interesting in some profiles (which I'll copy farther
>> down) with the deadline scheduler and AS.  Using our DBT-2 workload, we
>> see with the deadline scheduler our patch to avoid bounce buffers
>> doesn't change the profile much.  But with AS, we don't see
>> bounce_copy_vec or __blk_queue_bounce near the top of the profile.  Any
>> ideas why?
> 
> It shouldn't make any difference.
> 
> One thing to be careful about is to make sure that the pages which are
> being put under I/O are in the same place across different tests.
> 
> Suppose your machine already had 3G of pagecache and you then run the test.
> You would magically find that newly allocated pages come out of
> ZONE_NORMAL and no bouncing is needed for them.  So the moral is to make
> sure that the starting conditions are the same for each test: almost all
> memory free.

I think we've tracked the problem to old firmware on the raid
controllers we have.  They'll be upgraded shortly.

Mark