large filesystem & fileserver architecture issues.

[Mark Hahn]

> I read the abstract last evening and got a taste for it.  That is really a 
> remarkable idea to use the ethernet checksum for data integrity of stored 
> data.  Thanks for the heads-up.

for me, the crux of the idea is:

	- if you want big storage, $/GB drives you to IDE.

	- IDE is not amazingly fast, reliable or scalable.

	- building storage bricks out of IDE makes a lot of sense,
	since they can now be quite dense, low-overhead, etc.

	- ethernet is a wonderfully hot-pluggable interconnect for this 
	kind of thing.

	- doing raid over a multicast-capable network is pretty cool.

	- using eth's checksumming is pretty cool.

	- doing it this way (all open-source, including software raid)
	means the system is much more transparent - you are not dependent
	on some closed-source vendor tools to control/monitor/upgrade
	your storage.

Ben's approach (along with Lustre, for instance) seems very sweet for HPC
type storage needs.

one thing I do ponder, though, is whether it really makes sense to hide 
raid so firmly under the block layer.  it's conceptually tidy, to be sure,
and works well in practice.  but suppose:

	- to create a filesystem, you hand some arbitrary collection of 
	block-device extents to the mkfs tool.  you also let it know 
	which extents happen to reside on the same disk, bus, host, UPS,
	geographic location, etc.

	- you can tell the FS that your default policy should be for 
	reliability - that raid5 across separate disks is OK, for instance.
	or maybe you can tell it that a particular file should be raid10
	instead.  or that a file should be raid1 across each geographic site.
	or that updates to a file should be logged.  or that it should 
	transparently compress older files.

	- the FS might do other HSM-like things, such as incorporating
	knowlege of what's on your tape/DVD/cdrom's.

_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf