lkml.org 
[lkml]   [2010]   [Feb]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 00/11] [RFC] 512K readahead size with thrashing safe readahead
On Wed, Feb 03, 2010 at 10:24:54AM -0500, Vivek Goyal wrote:
> On Wed, Feb 03, 2010 at 02:27:56PM +0800, Wu Fengguang wrote:
> > Vivek,
> >
> > On Wed, Feb 03, 2010 at 06:38:03AM +0800, Vivek Goyal wrote:
> > > On Tue, Feb 02, 2010 at 11:28:35PM +0800, Wu Fengguang wrote:
> > > > Andrew,
> > > >
> > > > This is to lift default readahead size to 512KB, which I believe yields
> > > > more I/O throughput without noticeably increasing I/O latency for today's HDD.
> > > >
> > >
> > > Hi Fengguang,
> > >
> > > I was doing a quick test with the patches. I was using fio to run some
> > > sequential reader threads. I have got one access to one Lun from an HP
> > > EVA. In my case it looks like with the patches throughput has come down.
> >
> > Thank you for the quick testing!
> >
> > This patchset does 3 things:
> >
> > 1) 512K readahead size
> > 2) new readahead algorithms
> > 3) new readahead tracing/stats interfaces
> >
> > (1) will impact performance, while (2) _might_ impact performance in
> > case of bugs.
> >
> > Would you kindly retest the patchset with readahead size manually set
> > to 128KB? That would help identify the root cause of the performance
> > drop:
> >
> > DEV=sda
> > echo 128 > /sys/block/$DEV/queue/read_ahead_kb
> >
>
> I have got two paths to the HP EVA and got multipath device setup(dm-3). I
> noticed with vanilla kernel read_ahead_kb=128 after boot but with your patches
> applied it is set to 4. So looks like something went wrong with device
> size/capacity detection hence wrong defaults. Manually setting
> read_ahead_kb=512, got me better performance as compare to vanilla kernel.
>

I put a printk in add_disk and noticed that for multipath device get_capacity() is returning 0 and that's why ra_pages is being set to 1.

Thanks
Vivek


\
 
 \ /
  Last update: 2010-02-03 17:01    [W:0.061 / U:0.584 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site