Re: [PATCH][2.6-mm] Readahead issues and AIO read speedup

[Ram Pai]

On Thu, 2003-08-07 at 13:58, Andrew Morton wrote:
> Badari Pulavarty <pbadari@xxxxxxxxxx> wrote:
> >
> > On Thursday 07 August 2003 10:39 am, Andrew Morton wrote:
> >  > Badari Pulavarty <pbadari@xxxxxxxxxx> wrote:
> >  > > We should do readahead of actual pages required by the current
> >  > > read would be correct solution. (like Suparna suggested).
> >  >
> >  > I repeat: what will be the effect of this if all those pages are already in
> >  > pagecache?
> > 
> >  Hmm !! Do you think just peeking at pagecache and bailing out if
> >  nothing needed to be done, is too expensive ? Anyway, slow read
> >  code has to do this later. Doing it in readahead one more time causes
> >  significant perf. hit ?
> 
> It has been observed, yes.

we found substantial improvements (around 20% )in Database decision
support work load on Filesystems.

To address your concern regarding possible SDET regression generated by
this patch, we (Steve Pratt) ran a bunch of regression tests on
2.6.0test5-mm2(with and without the patch). I have pasted the results of
SDET and Kernel Bench. We did not find any noticable performance
regression.
 

Here are some results from Steve on test5-mm2
**************************************************************************

			
			    sdet comparison
				of
	2.6.0-test5-mm2 vs 2.6.0-test5mm2-without-READAHEAD-patch


Results:Throughput 

           tolerance = 0.00 + 3.00% of 2.6.0-test5-mm2
           2.6.0-test5-mm2 2.6.0-test5mm2-without-READAHEAD
   Threads      Ops/sec      Ops/sec    %diff         diff    tolerance
---------- ------------ ------------ -------- ------------ ------------
         1         3089         3103     0.45        14.00        92.67 
         4        11181        11294     1.01       113.00       335.43 
        16        18436        18530     0.51        94.00       553.08 
        64        18867        19002     0.72       135.00       566.01 

****************************************************************************
**************************************************************************
			kernbench comparison
				of
		2.6.0-test5-mm2 vs 2.6.0-test5mm2-without-READAHEAD


Results:Elapsed Time  
           tolerance = 0.00 + 3.00% of 2.6.0-test5-mm2
           2.6.0-test5-mm2 2.6.0-test5mm2-without-READAHEAD
                Seconds      Seconds    %diff         diff    tolerance
---------- ------------ ------------ -------- ------------ ------------
         2       96.015       95.035    -1.02        -0.98         2.88 

**************************************************************************

Would you like us to run some other tests?

Thanks,
RP



> 
> > And also, do you think this is the most common case ?
> 
> It is a very common case.  It's one we need to care for.  Especially when
> lots of CPUs are hitting the same file.
> 
> There are things we can do to tweak it up, such as adding a max_index to
> find_get_pages(), then do multipage lookups, etc.  But not doing it at all
> is always the fastest way.
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> 

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/