CCL:Largest Linux Cluster?
rbbrigh at valeria.mp.sandia.gov
Thu Jan 24 11:07:28 PST 2002
> > Sandia National Laboratories has C-Plant that runs Linux in addition to several
> > layers of home-grown OS on several thousand nodes. The basic node is a DEC
> > ev6 with myranet (sp). They use no local disk, opting for a huge disk farm.
> do you happen to know how they manage the huge disk farm???
> - resumably raid5 systems...
> - are each raid5 sub-system dual-hosted so that the other cpu
> can getto the data if one of the cpu cant get to it
> - does all nodes access the "disk farm" thru the gigabit ethernet
> or dual-hosted scsi cables ??
> - how does one optimize a disk farm ?? (hdparm seems too clumbsy)
> -- in the old days.... 1980's ... there used to be dual-hosted
> disk controllers where PC-HOST#1 and PC-HOST#2 can both access the same
> physical CDC/DEC/Fujitsu drives
> - wish i could find these dual host scsi controllers for todaysPCs
The Cplant clusters use a parallel I/O system called ENFS, which is a modified
version of the NFS protocol that avoids locking and doesn't provide full
UNIX semantics. Compute nodes send parallel I/O requests via ENFS to a set
of proxy I/O nodes that mount a back-end filesystem through gigE uplinks.
In the case of the 1792-node cluster, the back-end is an SGI O2K running XFS
More information about the Beowulf