[OpenAFS] Overview? Linux filesystem choices

Douglas E. Engert deengert@anl.gov
Fri, 01 Oct 2010 09:20:54 -0500

On 10/1/2010 4:35 AM, Harald Barth wrote:
>> ZFS has some really nice features, but Oracle just priced themselves
>> out of the market for scientific computing.
> Looks like their strategy to me. Time will tell if it will be successful.
>> It's hard enough to sell buying enterprise disks and servers, when
>> consumer stuff is much cheaper[1], but add a doubling of the head
>> node price to have a safe filesystem and it just won't fly.
> As you say, I just can't "sell" it to the researchers. They rather
> have double the capacity and double the performance instead. Btw, the
> only real troublesome advanced HW failures with single bit rot that we
> have encountered where ZFS would have saved the day was with an
> advanced "enterprise" SAN system (RIO). The simple "consumer" stuff
> in my experience just fails in a more simple manner.

ZFS did save the day for us. An AFS filserver was reporting I/O errors,
which ZFS was reporting as errors. A closer look a the SAN showed a few
bad disks, but no other users of the SAN were reporting errors. After
replacing the disks, AFS was working again.

> With daily backups, our HD failues are rare enough and not too much of
> a pain so I bet that any researcher rather would have double the
> storage and performance than double the price (or more) for Oracle-FS.
> Another way to tackle the data corruption issue in the AFS case would
> be to add checksum functionality to the fileserver backend. In
> contrast to NFS, we have the advantage that noone reads the data
> directly from the file system but always through the client.

ZFS stores the checksum of a block, not in the block, but in a higher level
block. This can then detect if a block failed to be written when its read back.
Other file systems could not detect this, as they would read old data,
with an old checksum. Keep this in mind if checksums are added to AFS.

> Harald.
> _______________________________________________
> OpenAFS-info mailing list
> OpenAFS-info@openafs.org
> https://lists.openafs.org/mailman/listinfo/openafs-info


  Douglas E. Engert  <DEEngert@anl.gov>
  Argonne National Laboratory
  9700 South Cass Avenue
  Argonne, Illinois  60439
  (630) 252-5444