task asssignment logic wrt average turn-around time
17 Sep 2007 14:12:08 UTC
Topic 193148
(moderation:
In assigning tasks does the system take into consideration the average turn-around time of the machines being assigned a task? i.e., is a particular task assigned to machines with comparable (or at least not significantly disparate) turn-around times?
task asssignment logic wrt average turn-around time
)
Not currently. There was a two-tier system during a past run, which preferentially assigned short Work Units to slow machines.
That scheme was associated with a nasty problem late in the run. The supply of short WUs ran short weeks before the end of the run, and the implementation burned a lot of the available server capacity fruitlessly searching for them.
On the other hand work assignment during the current run seems to have steadily moved to longer and longer WUs, with most of the ones currently being handed out pretty near the maximum. For those of us with some slower hosts, it is good news that the new run promises to have shorter maximum compute times.
RE: Not
)
As witnessed by my 400MHz Celeron, which got itself matched up against an 8-way Xeon E5365 3.0GHz server on this monster WU.
Don't worry, partner, I'll deliver on time - 7 days 6 hours CPU so far, 63% done.
That reminds me - I'd better set 'no new work' on that box: otherwise I'll be holding up the finish of the whole run, if I let it pick up another monster to run after this one.
RE: In assigning tasks does
)
In the next phase S5R3, to begin in just a few days, the length of the individual tasks should be shorter and distributed more evenly. Also, turn-around-time doesn't tell the whole story, it depends on the time configured between network accesses, result cache size, and performance. Some of the most powerful machines on EaH have rather long turn-around times, but there's no reason not to give them monsters, for example.
CU
H-BE
Thanks for prompt replies. I
)
Thanks for prompt replies. I asked because my "pending" bucket is over 20% of my "credited bucket" (granted my credits are a measly 16400) and when I pull the thread on the pending tasks the machine of the task-mate has an average turn-around time 3 to 10 times that of my machine. Makes me wonder why those long ta machines weren't paired up on the tasks.
Since I can't hold a candle to the multi mega machine mega credit guys, I take pride in my "recent average credit" score and don't like to see the big up and (especially) down spikes that result while waiting for the "penders" to complete and when they finally do.
Call me instant gratification ;>
I wouldn't worry about
)
I wouldn't worry about holding up the run. the holdup's always been the people who DL WUs but never complete them. In the past a few weeks after the next run started all the outstanding incomplete WUs were pushed out to a half dozenish machines each to quickly get them completed. If you've got a slow box, or a long queue that could be minor an issue in that you'd be doing work needlessly but you'd still get credit as long as you completed it within your deadline.