LWN: Comments on "Lustre 1.0 released"

Lustre/Intermezzo/P2P

giraffedata — Fri, 19 Dec 2003 17:12:09 +0000

It is not uncommon for a home with a few machines to have a few tens to hundreds of gigabytes of disk space scattered across several machines. It would be nice to unify that under one namespace.

Agreed, but that has little to do with Lustre. You're alluding to a distributed filesystem, which is fundamentally different from what Lustre is: a shared filesystem. The Lustre approach to unifying the storage is to remove the storage from all those systems and stick it into a central filesystem to which all the systems have access. It's specially designed to have that logically centralized system be physically spread out enough that it isn't a bottleneck even for huge numbers of storage users.

The separate storage approach is what makes Lustre impractical for small applications -- who wants to pay for a bunch of dedicated metadata and object storage servers just to serve a few home computers?

I don't doubt that features could theoretically be added to Lustre to allow it be a distributed filesystem and/or be practical in homes and mobile situations, but those features would have little to do with Lustre technology as it is known today.

Fault tolerance

snitm — Thu, 18 Dec 2003 15:19:45 +0000

Client level Raid1 or Raid5 isn't possible; however CFS has been working on the infrastructure for adding such support to Lustre, from a Lustre cvs commit message:

Add new LOV EA format which gives us fields for features needed in the post
1.0 stage (e.g. RAID, OST migration/replacement) as well as more efficient
storage than the current layout when there are few stripes on lots of OSTs.

Lustre/Intermezzo/P2P

brugolsky — Thu, 18 Dec 2003 14:23:07 +0000

Lustre is perhaps a bit too tightly coupled for home, and especially mobile wireless use. Additionally, it is designed for configurations where there is a clear distinction between clients and servers.

The typical home network consists of a bunch of machines with an OS install that is an every-shrinking fraction of the available disk-space. It is not uncommon for a home with a few machines to have a few tens to hundreds of gigabytes of disk space scattered across several machines. It would be nice to unify that under one namespace.

Laptop/PDA users want the aggressive caching and disconnection/reintegration of Intermezzo, but in a more distributed P2P fashion, with replication. Something like Ivy (http://www.pdos.lcs.mit.edu/ivy/) is a bit closer to meeting those needs.

Happily, Peter Braam and Andreas Dilger have suggested that once Lustre has met all of its requirements, they are inclined to revisit Intermezzo and apply some of the lessons learned in constructing Lustre.

Fault tolerance

rmacaulay — Thu, 18 Dec 2003 07:20:13 +0000

Can a piece of data that would be stored on a Object storage target be stored on multiple targets to enable fault tolerance on the storage nodes w/out the need for shared storage?

Only one active MDS is possible (for now)

snitm — Thu, 18 Dec 2003 05:48:58 +0000

Lustre 1.0 only supports one active Metadata server; multiple active MDSs is on the roadmap.