Re: [PATCH] fix readahead pipeline break caused by block plug

[Wu Fengguang <wfg@linux.intel.com>]

On Tue, Jan 31, 2012 at 10:36:53PM -0500, Vivek Goyal wrote:
> On Tue, Jan 31, 2012 at 05:22:17PM -0500, Vivek Goyal wrote:
> [..]
> 
> > > 
> > > We've never really bothered making the /dev/sda[X] I/O very efficient
> > > for large I/O's under the (probably wrong) assumption that it isn't a
> > > very interesting case.  Regular files will (or should) use the mpage
> > > functions, via address_space_operations.readpages().  fs/blockdev.c
> > > doesn't even implement it.
> > > 
> > > > and by the time all the pages
> > > > are submitted and one big merged request is formed it wates lot of time.
> > > 
> > > But that was the case in eariler kernels too.  Why did it change?
> > 
> > Actually, I assumed that the case of reading /dev/sda[X] worked well in
> > earlier kernels. Sorry about that. Will build a 2.6.38 kernel tonight
> > and run the test case again to make sure we had same overhead and
> > relatively poor performance while reading /dev/sda[X].
> 
> Ok, I tried it with 2.6.38 kernel and results look more or less same.
> Throughput varied between 105MB to 145MB. Many a times it was close to
> 110MB and other times it was 145MB. Don't know what causes that spike
> sometimes.

The block device really has some aged performance bug. Which
interestingly only show up in some test environments...

> I still see that IO is being submitted one page at a time. The only
> real difference seems to be that queue unplug happening at random times
> and many a times we are submitting much smaller requests (40 sectors, 48
> sectors etc).

Would you share the blktrace data?

Thanks,
Fengguang