[RC5] Scheduling d.net

Thomas Womack thomas.womack at merton.oxford.ac.uk
Wed Mar 4 01:02:00 EST 1998


>b) Is anyone else concerned about the multi-project model for v3?  d.net
has
>been unstoppable, in the past, because everyone interested in the idea has
>been unified in the direction of a solution of one problem.  If everyone is
>able to pick a different project, then won't the power of d.net be
>splintered into a whole bunch of weaker factions?

Yes - you've just discovered why supercomputers tend to run in a batched
processing mode rather than a multi-tasking one. On the other hand,
supercomputer applications tend to have fairly good estimates for how long
they'll take (computer time is expensive), and you might well partition a
large system so that a big job runs for a long time on 3/4 of it whilst
small test jobs run in quick succession on the rest.

On a third hand, consider three jobs, A, B and C, which would take one, two
and three months respectively.

Do C; do B; do A : C done after 3 months, B done after 5 months, A done
after 6 months.

Split the net into six groups 1..6.

First 3 months : 12 do A, 34 do B, 56 do C. A gets done after 3 months. B
has had a month done; C has had a month done.
Next two months : 123 do B, 456 do C. There's a month left for B; there's
half the net working on it, so it takes two months. C has now had two months
done.
Last month : 123456 do C and finish it off in a month.

So A has been done after 3 months, B after 5 months, C after 6 months. The
factional approach works quite well if there's no consensus on the
importance of the tasks; if A turns out to be important, it'll get scheduled
first and get done in a month.

Tom
--
To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest



More information about the rc5 mailing list