Hi,
On Friday 01 August 2008 00:46:19 Daniel Hain wrote:
> I've been informed that the project team has an Aug 6 deadline for
> providing EOF materials to Marketing and Docs, pending resolution of
> this case (timeout on 8/4). I would like to summarize where I think we
> are so that the project team can address any issues between now and when
> the case times out.
As an administrator of a site where CacheFS has been deployed for quite some
time, I'd like to add my $.02 to this discussion:
We're running a growing compute cluster of various types of machines since the
late 1990s to perform various types of bioinformatics analyses. We have been
(and still are) using CacheFS to provide program binaries as well as (mostly
read-only) data to all of these machines from a single file server (2 in
fact: one for the data, another for the binaries).
While being aware of some deficiencies of its current implementation, CacheFS
is doing quite a good job in our setup and in my opinion there doesn't seem
to be an alternative for us at the moment. Running without it will basically
bring the file server to a halt when the nodes are busy (we learned that by
accidentally disabling CachFS).
We did some experiments with NFSv4 a while back but its caching did not seem
sufficient for our workload (as mentioned elsewhere in this thread). Exludus'
RepliCator software works well for distributing the data to the individual
nodes but IMHO is less ideal for providing binaries to a large number of
nodes.
We had (admittedly a rough) look at other solutions, both software and
hardware, for providing binaries and regularely updated sets of data to a
growing number of compute nodes but they usually require a considerable
amount of human and/or financial resources. At least nothing seemes as easy
as setting up an NFS server with CacheFS clients... ;-) Having to do without
CacheFS (without an appropriate alternative) would make life a lot more
complicated for us...
Thanks for your attention,
Torsten
