[LSF/MM TOPIC] Performance of seekdir(SEEK_HOLE/DATA) and FIEMAP

[LSF/MM TOPIC] Performance of seekdir(SEEK_HOLE/DATA) and FIEMAP

[Trond Myklebust]

Hi all,

If there are any free slots on the filesystem track, then it would be
nice to have a conversation about use of FIEMAP and SEEK_HOLE/DATA
performance.

The issue is this: as large sparse files become more prevalent, there
is a potential performance advantage to be made if applications can
quickly and easily detect areas of no data.
Full disclosure: my interest here is seeing Anna's work on implementing
the NFSv4.2 sparse file support come to fruition; right now she is
finding that the performance of the filesystem tools (both seekdir()
and FIEMAP) are insufficient to allow us to implement the READPLUS (htt
ps://tools.ietf.org/html/rfc7862#section-6.2.1) operation and realise
the promised performance goals.

If there is time left on the schedule, then it might be nice to have a
full discussion of this topic with all the filesystem folks present,
but if not, we could do it as a lightning talk.

Thanks
  Trond

-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@primarydata.com

Re: [Lsf-pc] [LSF/MM TOPIC] Performance of seekdir(SEEK_HOLE/DATA) and FIEMAP

[Jeff Layton]

On Thu, 2017-03-16 at 16:09 +0000, Trond Myklebust wrote:
> Hi all,
> 
> If there are any free slots on the filesystem track, then it would be
> nice to have a conversation about use of FIEMAP and SEEK_HOLE/DATA
> performance.
> 
> The issue is this: as large sparse files become more prevalent, there
> is a potential performance advantage to be made if applications can
> quickly and easily detect areas of no data.
> Full disclosure: my interest here is seeing Anna's work on implementing
> the NFSv4.2 sparse file support come to fruition; right now she is
> finding that the performance of the filesystem tools (both seekdir()
> and FIEMAP) are insufficient to allow us to implement the READPLUS (htt
> ps://tools.ietf.org/html/rfc7862#section-6.2.1) operation and realise
> the promised performance goals.
> 
> If there is time left on the schedule, then it might be nice to have a
> full discussion of this topic with all the filesystem folks present,
> but if not, we could do it as a lightning talk.
> 
> Thanks
>   Trond
> 

It turns out that we still had one open slot later in the afternoon on
Tuesday so I added this there.

Cheers,
-- 
Jeff Layton <jlayton@poochiereds.net>