Cloud architectures: fabric and networking
Geoff Uyleman
services?
Will we see an adoption of FCoE in cloud architectures? Possibly a way
to leverage some of the monolithic storage subsystems so many of us have
on the floor today? Presenting LUNs to, say, hadoop or perhaps a Lustre
file system implementation on an intermediary x86 cluster (hello
complexity--can HDFS be locally mounted on each x86 workload node?) that
we can use to present to the x86 workload cloud?
What about 8Gb FCP for the backend block I/O? And, for that matter, what
about IP or SDP over DDR infiniband (20Gb should be more than sufficient
for storage I/O)? This works well for clustered x86 HPC systems today,
and I have seen it working for storage I/O as well (SDP/ IB, not IP).
Is there a need to keep TCP/IP in the stack or are we better off without
and switch everything L2 (as we have been doing with FCP anyway)and
leverage our inter-datacenter Fibre loops without any TCP/ IP
implementation? We seem to be able to switch all of our traffic just
fine without needing to resort to routing at all. The FCoE standard
seems to confirm this point.
Therefore, is there a future for 10Gb iSCSI or is this standard doomed
to die in the mid storage sub system market stuck at 1Gb because all the
value is gone once the cheap 1Gb switches are obsolete, and it will take
years for the non-FCoE 10GbE switches to come close to a similar cost/
benefit advantage? (one of the cloud infrastructures that gets mentioned
a lot in this group runs their block I/O back-end over iSCSI for these
cost/ volume reasons).
So to come back to my first question, what is working for you guys and
what do you think is going to stick?
Thanks!
Geoff
ge...@uyleman.com
Ezra Zygmuntowicz
We're using AoE for our storage fabric along with RHCS and CLVM to
manage it all. This is all used to present LUNs to our virtualized
clusters in single filesystems as well as GFS clustered filesystems
shared between sets of VM's.
Cheers-
-Ezra
http://engineyard.com