[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-nfs
Subject:    Re: [NFS] how long is wdelay?
From:       Neil Brown <neilb () cse ! unsw ! edu ! au>
Date:       2000-12-28 0:21:53
[Download RAW message or body]

On Wednesday December 27, ezk@cs.columbia.edu wrote:
> How long does an NFS server wait when the wdelay option is used (on by
> default).  It seems possible to turn this option on/off, but not to
> configure how long it waits to cluster adjacent writes.

It waits 10 milliseconds, but only if the previous write request
was to the same file.

> 
> I also heard that this option isn't that useful in practice?  Does anyone
> have more info about why that might be?

I haven't looked into it deeply, and haven't measured the effect at
all, but it strikes me as a very coarse heursitic that is implemented
at the wrong level, so I don't hold out a lot of hope for it.

- if there are write requests for two concurrent files happening, 
  then there will be many fewer delays inserted, thus substantially 
  reducing the possible effectiveness.
- it assumes that temporally adjacent writes to a given file are
  likely to be adjacent on the target device, which may often be true,
  but is definately a simplification.
- it assumes that fsync will mark the inode not I_DIRTY, which doesn't
  seem to be the case for ext2 (though I'm not sure how significant
  this is).

- and as I said, it is really at the wrong level.  You need support in
  the filesystem and the device driver and all the interfaces to get
  the most out of gathering writes.

But the comment say "it seems to work", so maybe it does for some
workloads. 

NeilBrown


  

> 
> Thanks,
> Erez.
> 
> _______________________________________________
> NFS maillist  -  NFS@lists.sourceforge.net
> http://lists.sourceforge.net/mailman/listinfo/nfs

_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
http://lists.sourceforge.net/mailman/listinfo/nfs

[prev in list] [next in list] [prev in thread] [next in thread]