xfs
[Top] [All Lists]

Re: Fragmentation and hole punching

To: Eric Sandeen <sandeen@xxxxxxxxxxx>
Subject: Re: Fragmentation and hole punching
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 7 Jan 2013 06:57:48 +1100
Cc: "Bradley C. Kuszmaul" <kuszmaul@xxxxxxxxx>, xfs@xxxxxxxxxxx
In-reply-to: <50E90A16.8060302@xxxxxxxxxxx>
References: <CAKSyJXchcnnpD6aiCXnaSF3ym1cvGZjYPGKKCgiOHY59vbuntQ@xxxxxxxxxxxxxx> <50E90A16.8060302@xxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Sat, Jan 05, 2013 at 11:22:30PM -0600, Eric Sandeen wrote:
> On 1/5/13 6:52 PM, Bradley C. Kuszmaul wrote:
> > I'm using hole punching now in my code.  Should I have any
> > concerns about fragmentation?
> > 
> > Here's how I use it: I perform pwrite() operations of various
> > sizes. These are "blocks" in my application. These blocks are
> > always 4KB-aligned, but they are of many different sizes. Most
> > writes are near a megabyte in size, but hardly ever exactly a
> > megabyte.  I also punch holes.  The holes always correspond
> > exactly to a previously written pwrite().  A pwrite() always
> > writes into a hole.  (It never overwrites current data).
> 
> Just out of curiosity - what is the end goal here, why all the
> punching-and-rewriting?
> 
> > For example I might write 1MB at offset 1MB then write 3MB at
> > offset 2MB then punch a hole of size 1MB at offset 1MB then do
> > some more activity elsewhere in the file then write 512K at an
> > offset of 1MB (partially filling the hole that I had punched)
> > 
> > I have a lot of experience running this code on xfs without the
> > hole punching. The pattern of the pwrite()'s are the same with
> > or without hole punching. So I'd hope that the hole punching
> > didn't make things any worse, and maybe made things better.
> 
> Once you free up blocks, there is always a chance that some other
> file will end up claiming them, so I can see fragmentation
> downsides; not sure I see any upside there.
> 
> OTOH, fragmentation may not matter at all, depending on your
> access patterns etc.
> 
> My first guess for the hole-punching reason is to maximize
> available space; that makes me think you are really utilizing all
> available space; that makes me think you might really get files
> intermingled on disk if you're running things this way.
> 
> But if you're seeking around in files to read at random offsets
> anyway, it might not matter.

That's the determine factor, really. If read/write is random and is
aligned to the holes being punched, then fragmentation may not
affect performance in any way. There will be increased fragmentation as a
result of punching holes and rewriting like this, and if you then
read or rewrite the data sequentially the fragmentation is likely to
have significant impact on performance as the filesystem ages.

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

<Prev in Thread] Current Thread [Next in Thread>