Re: Just A Few Questions...
> penalty of diskless nodes mostly happen at program startup times,
> and it is mostly unnoticable.
>
> However, for programs which use diskspace as scratch working place,
> the network penalty is rather big, but they are probably better off
> using ramdisk than local disk.
I confirm that. If there is no user I/O traffic, you are OK. Even with
user some such traffic, it only really bothers you if it is so high that
the user's program is not getting much CPU time anyway, so this not an
appropriate use of a number any crunching machine in any case.
If you have a large (>10, say) number of nodes, a Gbit connection for the
server is a good idea, helps a good deal. There are 3COM switches with 24
100-ports plus 1 1000-port that you can use for that. It also helps having
a system disk server separate from the front-end, or a separate home disk
server, but that is only justifiable on really big clusters, I guess.
If you have the memory needed, a RAM disk is the best thing for scratch,
even it is really only a way to correct bad programming, since the program
should be written to use memory internally. A raid-0 of 4 to 6 fast SCSI
disks on a local 160 Mbps SCSI card on the node can also be quite fast,
but is a bad idea, quite expensive and looses badly to the RAM disk.
Cheers,
----------------------------------------------------------------
Jorge L. deLyra, Associate Professor of Physics
The University of Sao Paulo, IFUSP-DFMA
For more information: finger delyra@latt.if.usp.br
----------------------------------------------------------------
Reply to: