Re: [Iscsitarget-devel] Proposal: New Hybrid IO Type

[Ross S. W. W. <rw...@me...>]

> -----Original Message-----
> From: Ming Zhang [mailto:bla...@gm...] 
> Sent: Monday, June 25, 2007 9:44 AM
> To: Ross S. W. Walker
> Cc: Isc...@li...
> Subject: Re: [Iscsitarget-devel] Proposal: New Hybrid IO Type
> 
> On Sun, 2007-06-24 at 13:53 -0400, Ross S. W. Walker wrote:
> > 
> > List Developers,
> > 
> > I am putting forward an idea for a new hybrid io type that takes
> > the best parts of fileio and blockio and implements it using VFS.
> > 
> > The main advantage to blockio is that it can handle arbitrarily
> > large io requests in their entirety which for large io operations
> > provides a significant boost in throughput. It's disadvantage
> > though is that it bypasses page cache for buffering which means
> > each io has the added to-disk latency.
> > 
> > The main advantage of fileio is that it uses page-cache to buffer
> > each requests (mostly for reading) and can leverage the insanely
> > great read-ahead cache algorithm in VFS. It's major disadvantage
> > though is that all io operations are performed in 4k page sized
> > chunks which limits the throughput on all non-cached operations
> > (mainly write). 
> 
> i wonder if you did any profiling like blktrace to validate 
> this claim?

I didn't do a blktrace, I am basing this on 1) that is what the code
does, and 2) the only way to allow requests to merge is to let them
run asynchronously and have the queue unplug and plug itself, or by
overlapping ios, but by unplugging the queue at each read/write it
prevents this from happening and the overlapping ios only work well
if the application on the initiator supports it.

> could we find out what is the real issue of fileio?

I believe it is all in the max size of the request. All VFS operations
work in 4k pages, they expect page-cache to be there and to allow io
scheduler to handle the plugging and unplugging.

The performance difference is only visible with requests >= 64K as it
is unnoticeable with smaller request sizes.

> ps, some people in LKML are now proposing large size page 
> cache, maybe u
> have interest on that.

Large page cache? Or large requests from page-cache?

-Ross

______________________________________________________________________
This e-mail, and any attachments thereto, is intended only for use by
the addressee(s) named herein and may contain legally privileged
and/or confidential information. If you are not the intended recipient
of this e-mail, you are hereby notified that any dissemination,
distribution or copying of this e-mail, and any attachments thereto,
is strictly prohibited. If you have received this e-mail in error,
please immediately notify the sender and permanently delete the
original and any copy or printout thereof.