xfs
[Top] [All Lists]

Re: Realtime Section - XFS

To: alan@xxxxxxxxx
Subject: Re: Realtime Section - XFS
From: Steve Lord <lord@xxxxxxx>
Date: Thu, 07 Jun 2001 17:01:24 -0500
Cc: linux-xfs@xxxxxxxxxxx
In-reply-to: Message from alan@xxxxxxxxx of "Thu, 07 Jun 2001 16:38:01 EDT." <991946281.3b1fe6299bd51@xxxxxxxxxxxxxxxxxxxxxxxx>
Sender: owner-linux-xfs@xxxxxxxxxxx
> 
>       Are there userspace utilities to enable this realtime inode io behavior
> ?
> 
> 

No, just the ioctl call I mentioned to set the state on a file (must be
empty). If you look in cmd/xfstests/src you will find a few utilities
with code which use the ioctl: alloc.c, fsstress.c and randholes.c.

If you wanted to it would not be hard to extend lmdd (part of lmbench)
to set the realtime flag on a file it created, it can already be built
with direct I/O support. You can get the source of this from www.bitkeeper.com
it would then give you a tool to experiment with real time I/O.

Steve

> 
> Quoting Steve Lord <lord@xxxxxxx>:
> 
> > > 
> > >   Can I use and benefit from an rt section without making a specific
> > inod
> > > e
> > > realtime?  Or does it only benefit inodes that have been made
> > 'realtime'.  Do
> > > es
> > > it have any specific overall fs performance effect?  My test is small,
> > so it 
> > > is
> > > fast anyway. also, is there any size guideline to the sizes of rt
> > sections vs
> > > size of the data and log sections?  Or does it only matter to the
> > inodes that
> > > are used in a realtime fashion?
> > > 
> > > Alan Willis
> > > alan@xxxxxxxxx
> > > 
> > 
> > The realtime subvolume is where data for realtime inodes is allocated,
> > so yes
> > it only applies to those. Create a new file, make it realtime, and then
> > do
> > I/O to it. Realtime is a little bit of a misnomer, the differences
> > are:
> > 
> >    o There is only data in the partition, all metadata remains the the
> > data
> >      subvolume. If you put the partitions in the correct place then the
> > only
> >      time headseeks happen is to go and get actual data, no log or
> > metadata
> >      I/O to get in the way.
> > 
> >    o A different space allocator is used, it uses a binary chop type
> > approach
> >      to disk space which is wasteful, but which means it is very hard to
> > 
> >      fragment the files on realtime partitions. This allocator is also
> > faster
> >      and tends to allocate space in a more deterministic amount of
> > time.
> > 
> >    o When you make the filesystem, and when you create a realtime file
> > via the
> >      ioctl call you can specify an extent size, all allocations will be
> > a
> >      multiple of this. I have my doubts that this will work as
> > advertised yet
> >      on linux.
> > 
> > These are all supposed to add up to something which is good for doing
> > streaming
> > I/O, typically of very large files such as video data.
> > 
> > Here are the changes I have to at least attempt to send disk I/O to the
> > 
> > realtime device.
> > 
> > Users of realtime on Irix would probably create a filesystem which was
> > almost
> > all realtime space, you still need the regular partition for the log and
> > the
> > metadata.
> > 
> > Steve
> > 
> > 
> > 
> > 



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