Multiple clients

[SmoulderingCircuits]

Probably been debated it's fair share on every forum (done some reading up), but which is most efficient: more clients that take longer or less clients that complete quicker? In theory it should be the latter but when you have clock cycle wastage as a common problem it might be better to run the former (multiple clients on each device).

Usual PPD

5870s > 9k each

1090T > 4.5k

430s > 5k each

Athlon 64 X2 6000+ > 1.4k

E-350 > 0.2k

I will be running multiple clients for a while and - assuming I remember this thread - will report back with any changes to PPD since running multiple clients.

5870s > 2 GPU clients each

1090T > 3x 2core SMPs

430s > 2 GPU clients each

Athlon > Usual SMP

E-350 > 7x 2core SMPs (so when it does send, it's a significant amount)

 

[Pickster]

Probably been debated it's fair share on every forum (done some reading up), but which is most efficient: more clients that take longer or less clients that complete quicker? In theory it should be the latter but when you have clock cycle wastage as a common problem it might be better to run the former (multiple clients on each device).

Usual PPD

5870s > 9k each

1090T > 4.5k

430s > 5k each

Athlon 64 X2 6000+ > 1.4k

E-350 > 0.2k

I will be running multiple clients for a while and - assuming I remember this thread - will report back with any changes to PPD since running multiple clients.

5870s > 2 GPU clients each

1090T > 3x 2core SMPs

430s > 2 GPU clients each

Athlon > Usual SMP

E-350 > 7x 2core SMPs (so when it does send, it's a significant amount)

Because of the way bonus points work, it's better to have just a single client with all avalaible threads working on the one work unit.

That said, unless the Unicore points have been changed recently, the way SMP points are being messed with it will getto the point soon where it will be worth considering multiple unicore clients over a single SMP one.

 

[SmoulderingCircuits]

Because of the way bonus points work, it's better to have just a single client with all avalaible threads working on the one work unit.

That said, unless the Unicore points have been changed recently, the way SMP points are being messed with it will getto the point soon where it will be worth considering multiple unicore clients over a single SMP one.

But a single SMP or GPU don't use all available cycles (especially for AMD) so I'm trying to find out which has better advantage, the quick return bonus or using the remaining ~10% of cycles through putting an extra unit on everything (scheduling should in theory mean quick return would be better but doesn't seem to be a problem).

 

[SmoulderingCircuits]

Currently each 5870 client is bringing in 4.7k (9.4k each) and each 1090T client is bringing in 1.3k (3.9k) but sometimes goes to 3.9k per client (11.7k) so seems to be working better at the moment, will go set up the 430s.

Just hit 13k and still climbing for the two 430s and Athlon, so >

Athlon > 1.5k, leaves (same single SMP)

GT 430s > 5.75k each

 

[riotcity76]

But a single SMP or GPU don't use all available cycles (especially for AMD) so I'm trying to find out which has better advantage, the quick return bonus or using the remaining ~10% of cycles through putting an extra unit on everything (scheduling should in theory mean quick return would be better but doesn't seem to be a problem).

Before F@H had the quick return bonus, it was beneficial to run multiple clients. People were running all sorts of configurations on different types of hardware. It was giving more points overall, but the problem is that it slowed down the F@H project as a whole. F@H relies on results to be returned as quickly as possible. When everybody is running 2 SMP clients with 2 cores each instead of 1 SMP client with 4 cores, WUs take twice the time to finish. This was affecting the speed at which new WUs could be introduced, as it prevented Stanford from introducing new WUs until all units in a previous batch were finished.

This is why the quick return bonus was created: as a deterrent for running the project this way (it's more of an incentive to run a single client than it is a deterrent to run multiple clients, but you get the point). Thus not only is it better points-wise to run a single client now, it's also better for the project overall.

I can't speak much for the GPU side of things, since it doesn't have a quick return bonus (yet), but for the reason mentioned above I personally would run one client per card.

 

[SmoulderingCircuits]

Before F@H had the quick return bonus, it was beneficial to run multiple clients. People were running all sorts of configurations on different types of hardware. It was giving more points overall, but the problem is that it slowed down the F@H project as a whole. F@H relies on results to be returned as quickly as possible. When everybody is running 2 SMP clients with 2 cores each instead of 1 SMP client with 4 cores, WUs take twice the time to finish. This was affecting the speed at which new WUs could be introduced, as it prevented Stanford from introducing new WUs until all units in a previous batch were finished.

This is why the quick return bonus was created: as a deterrent for running the project this way (it's more of an incentive to run a single client than it is a deterrent to run multiple clients, but you get the point). Thus not only is it better points-wise to run a single client now, it's also better for the project overall.

I can't speak much for the GPU side of things, since it doesn't have a quick return bonus (yet), but for the reason mentioned above I personally would run one client per card.

I'm finding running multiple clients keeps the PPD at it's highest, whereas single clients don't utilise all the processing power available.