I found myself wondering what is going on with gpugrid because i get more and more units who take more calculations on the cpu.

Is this because of the errors of the cuda environment ?

but what is worse is they get rewarded much less even though they take much more power.

Old units :
764655 509901 2 Jun 2009 22:05:51 UTC 5 Jun 2009 1:05:13 UTC Over Success Done 1,636.94 4,531.91 5,664.89

Newer
796169 524891 9 Jun 2009 10:24:46 UTC 10 Jun 2009 6:30:33 UTC Over Success Done 4,772.05 3,809.64 4,762.04

Newest

807975 531950 11 Jun 2009 22:11:37 UTC 13 Jun 2009 4:27:20 UTC Over Success Done 8,047.47 3,809.64 4,762.04

Hi, looks to me you 'upgraded' BOINC from a 6.4 to a 6.6 version. I had this happen and with me it was because of a bug in BOINC. If you go back to 6.4.x it will be twice as fast again.

---

Join team Bletchley Park, the innovators.

I correct myself, you are using 6.5.0. I notice that the estimated time beloweach successful workunit is more or less the same at 88000. It is the CPU score that varies strongly. It may be that the involvement of the CPU varies strongly based on results, but does the result itself take significantly longer on a real clock ?

---

Join team Bletchley Park, the innovators.

Yeah, it's just you ;)

No, seriously, I checked my results regarding their cpu usage and found the following:

1700 cpu seconds: time per step 90 ms -> 500k steps
3100 cpu seconds: time per step 78 ms -> 650k steps
6000 cpu seconds: time per step 67 ms -> 625k steps

Ups.. before I did the math I would have expected the last one to come out at an even higher number. The problem with my numbers probably is that the simulated systems feature different numbers of atoms / complexity. This would normally be proportional to the time per step. But since they're using different calculation methods now (the new amber force fields etc.) there are probably further differences in the WUs, which can not be explained easily without further knowledge.

However, I don't think the effect is as severe as you say:

but what is worse is they get rewarded much less even though they take much more power.

Let's do the number for my rather typical WU 1 and WU 2 with low and high cpu usage, respectively.

WU 1: 45297s -> 1.90 WUs/day -> 9628 credits/day for the GPU. The CPU would yield ~4000 credits/day without GPU-Grid. GPU-Grid needs 1.9 * 1775s of these, that's 0.94h or a reduction of cpu production by 39 credits to 3960. Sum: 13588 credits/day.

WU2: 41744s -> 2.07 WUs/day -> 9856 credits/day. CPU: 2.07 * 5928s -> 3.4h -> a reduction of cpu production by 142 credits/day to 3856. Sum: 13714 credits/day.

So in this case the WU with higher cpu utilization actually earns more. The precise numbers will vary from system to system and from WU to WU, but I think this should be proof enough.

MrS

---

Scanning for our furry friends since Jan 2002