Only if you multiply the estimates for each task by the DCF would it match dividing the work request by the DCF. Doesn't matter whether it is client side or server side.
jm7 |------------> | From: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |Richard Haselgrove <[email protected]> | >--------------------------------------------------------------------------------------------------------------------------------------------------| |------------> | To: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |<[email protected]> | >--------------------------------------------------------------------------------------------------------------------------------------------------| |------------> | Cc: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |BOINC Developers Mailing List <[email protected]> | >--------------------------------------------------------------------------------------------------------------------------------------------------| |------------> | Date: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |11/09/2011 11:54 AM | >--------------------------------------------------------------------------------------------------------------------------------------------------| |------------> | Subject: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |Re: [boinc_dev] APR, DCF and non-deterministic projects | >--------------------------------------------------------------------------------------------------------------------------------------------------| |------------> | Sent by: | |------------> >--------------------------------------------------------------------------------------------------------------------------------------------------| |<[email protected]> | >--------------------------------------------------------------------------------------------------------------------------------------------------| It depends where you apply the correction. I believe that, in the old code which was removed by changeset 21153, the work *request* was left unmodified, but the estimated duration of each candidate task added to the scheduler reply was scaled to suit the host - that requires a muliplication. You can see the arithmetic in the Einstein server logs, accessible via the web. I'm seeing 2011-11-09 11:57:55.6499 [PID=2713 ] [send] CUDA: req 55.93 sec, 0.00 instances; est delay 0.00 2011-11-09 11:57:55.6499 [PID=2713 ] [send] active_frac 0.999854 on_frac 0.986984 DCF 1.156106 2011-11-09 11:57:55.7491 [PID=2713 ] [version] Best version of app einsteinbinary_BRP4 is ID 289 (18.33 GFLOPS) 2011-11-09 11:57:55.7533 [PID=2713 ] [send] est. duration for WU 109135289: unscaled 7639.12 scaled 8949.40 2011-11-09 11:57:55.7533 [PID=2713 ] [HOST#831490] Sending [RESULT#256378135 p2030.20100425.G44.31+01.48.C.b1s0g0.00000_3752_0] (est. dur. 8949.40 seconds) ----- Original Message ----- From: <[email protected]> To: "Richard Haselgrove" <[email protected]> Cc: "BOINC Developers Mailing List" <[email protected]>; <[email protected]>; "David Anderson" <[email protected]> Sent: Wednesday, November 09, 2011 4:15 PM Subject: Re: [boinc_dev] APR, DCF and non-deterministic projects > Shouldn't the work request be divided by the DCF, not multiplied? A DCF of > 35 means that the work takes 35 times as long as the server estimates... > > jm7 > > > |------------> > | From: | > |------------> > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |Richard Haselgrove <[email protected]> | > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |------------> > | To: | > |------------> > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |David Anderson <[email protected]>, BOINC Developers Mailing List <[email protected]> | > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |------------> > | Date: | > |------------> > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |11/09/2011 11:03 AM | > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |------------> > | Subject: | > |------------> > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |Re: [boinc_dev] APR, DCF and non-deterministic projects | > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |------------> > | Sent by: | > |------------> > >--------------------------------------------------------------------------------------------------------------------------------------------------| > |<[email protected]> | > >--------------------------------------------------------------------------------------------------------------------------------------------------| > > > > > >> That's true. The discussion should not drown in talk about credit. >> As you point out, credit just happens to be one very visible indicator >> that the estimates are sometimes way off. >> >> -- Janus > > Waaaaay off. > > While we're thinking about the wider principles, could we apply a band-aid > to staunch these massive server over-allocations, please? > > Another example: > > Simap requested work first, but with a DCF of 0.70, the client calculation > of the amount allocated was below the amount requested. > > Even after the 4-second completion of the NumberFields, DCF was still 35.52 > > So the request for 69736 seconds of work (under 5 hours, for this > quad-core) resulted in an allocation estimated at 2510769 seconds: that's > enough to keep all four cores occupied for over a week, according to client > calculations. Since project deadlines are exactly 1 week, NumberFields went > into immediate High Priority on all four cores, locking out all other > projects until things settle down a bit. That doesn't count as "play nicely > with other projects", in my book - but it's the removal of the DCF > adjustment from the server code which is causing it, not any fault of the > project. > > > 09-Nov-2011 14:18:35 [NumberFields@home] Starting task > wu_12E10_SF157-1_Idx6_Grp14124of41476_0 using GetBoundedDecics version 202 > 09-Nov-2011 14:18:39 [NumberFields@home] Computation for task > wu_12E10_SF157-1_Idx6_Grp14124of41476_0 finished > 09-Nov-2011 14:18:39 [NumberFields@home] Starting task > wu_12E10_SF157-1_Idx6_Grp13897of41476_0 using GetBoundedDecics version 202 > 09-Nov-2011 14:18:42 [NumberFields@home] Started upload of > wu_12E10_SF157-1_Idx6_Grp14124of41476_0_0 > 09-Nov-2011 14:18:46 [NumberFields@home] Finished upload of > wu_12E10_SF157-1_Idx6_Grp14124of41476_0_0 > 09-Nov-2011 14:18:54 [boincsimap] Finished upload of 20111101.952163_1_0 > 09-Nov-2011 14:18:57 [boincsimap] [sched_op] Starting scheduler request > 09-Nov-2011 14:18:57 [boincsimap] Sending scheduler request: To fetch work. > 09-Nov-2011 14:18:57 [boincsimap] Reporting 1 completed tasks, requesting > new tasks for CPU > 09-Nov-2011 14:18:57 [boincsimap] [sched_op] CPU work request: 34402.22 > seconds; 1.00 CPUs > 09-Nov-2011 14:18:57 [boincsimap] [sched_op] NVIDIA GPU work request: 0.00 > seconds; 0.00 GPUs > 09-Nov-2011 14:19:01 [boincsimap] Scheduler request completed: got 5 new > tasks > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] Server version 611 > 09-Nov-2011 14:19:01 [boincsimap] Project requested delay of 7 seconds > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] estimated total CPU task > duration: 27270 seconds > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] estimated total NVIDIA GPU > task duration: 0 seconds > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] handle_scheduler_reply(): got > ack for task 20111101.952163_1 > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] Deferring communication for 7 > sec > 09-Nov-2011 14:19:01 [boincsimap] [sched_op] Reason: requested by project > 09-Nov-2011 14:19:03 [boincsimap] Started download of 20111101.955716 > 09-Nov-2011 14:19:03 [boincsimap] Started download of 20111101.955743 > 09-Nov-2011 14:19:07 [NumberFields@home] [sched_op] Starting scheduler > request > 09-Nov-2011 14:19:07 [NumberFields@home] Sending scheduler request: To > fetch work. > 09-Nov-2011 14:19:07 [NumberFields@home] Reporting 2 completed tasks, > requesting new tasks for CPU > 09-Nov-2011 14:19:07 [NumberFields@home] [sched_op] CPU work request: > 69736.27 seconds; 2.00 CPUs > 09-Nov-2011 14:19:07 [NumberFields@home] [sched_op] NVIDIA GPU work > request: 0.00 seconds; 0.00 GPUs > 09-Nov-2011 14:19:09 [NumberFields@home] Scheduler request completed: got > 19 new tasks > 09-Nov-2011 14:19:09 [NumberFields@home] [sched_op] Server version 613 > 09-Nov-2011 14:19:09 [NumberFields@home] Project requested delay of 21 > seconds > 09-Nov-2011 14:19:09 [NumberFields@home] [sched_op] estimated total CPU > task duration: 2510769 seconds > 09-Nov-2011 14:19:09 [NumberFields@home] [sched_op] estimated total NVIDIA > GPU task duration: 0 seconds > _______________________________________________ > boinc_dev mailing list > [email protected] > http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev > To unsubscribe, visit the above URL and > (near bottom of page) enter your email address. > > > > > _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address. _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
