When are diskless compute nodes inappropriate?
bill at math.ucdavis.edu
Thu Jul 17 16:48:38 EDT 2003
On Tue, Jul 15, 2003 at 11:48:05AM -0400, Tod Hagan wrote:
> Okay, I'm convinced by the arguments in favor of diskless compute
> nodes, including cost savings applicable elsewhere, reduced power
>, and increased reliability through the elimination of
> moving parts.
Indeed. Although similar reliability can be had if you can survive
a disk failure.
> With all the arguments against disks, what are the arguments in favor
> of diskful compute nodes? In particular, what are the situations or
Swap, and high speed disk I/O. 35 MB/sec of sequential I/O to a local disk
is very hard to centralize. If you can make do with much less then it's not
to much of a big deal.
For our 32 node cluster on boot we:
netboot a kernel
kernel loads a ramdisk
disk is partitioned
disk is mkswaped
/scratch and /swap are mounted.
So this leave ZERO state on the hard disk, so if a disk dies just reboot
and the node works (but doesn't have /swap and /scratch), if you pull
a disk off a shelf and stick it in a node you just reboot.
Very nice to minimize the administrative costs of managing, patching,
backing up, troubleshooting etc of N nodes, with possibly different images,
and of course any state.
My central fileserver is a dual-p4, dual PC1600 memory bus, 133 Mhz/64 bit
PCI, and several U160 channels full of 5 disks each. I see 200-300 MB/sec
sustained for large sequential file reads/writes. Granted the central
fileserver can not keep up with 32 nodes wanting to read/write at 35 MB/sec,
but it's enough to usually not be a bottlneck.
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf