@Keith: your situation puzzles me as well.

Your data may suggest a correlation: high cpu usage - lower computation time. I checked with my WUs: 3 of them have 62.5 ms/step and CPU usages of 14, 14 and 291s. One has CPU 851 and "only" 60.1 ms/step. That's not really a clear-cut picture either.

MrS

Previous tests did not have this high a margin.

6.44 and earlier all had similar results. Work 1 and Work 2 ran pretty even or close. There would be some difference but always within a small range of the average. My usage has not changed much between versions. What I did then I still do now. I do not remember changing any other software between versions. So all things being equal, work 1 and work 2 should still have similar run times between the two, I can accept that 6.45 is a little slower, but the margin between the two makes no sense when there was not one before.

Average CPU sec of 6.44 and earlier versions of all work completed.
Home = 55,900s
Work 1 = 59,254s
Work 2 = 59,124s
There was a difference of 2 minutes 10.42 seconds for work 1 vs work 2
There was a difference of 53-55 minutes for home to work.

Now with 6.45
Home = 59,327s
Work 1 = 68,403s
Work 2 = 61,482s

OK so Home went up by 3,426s (57m)
I would expect work 1 and 2 to do about the same margin, allowing a little for the speed difference. In past versions when there was a change in version, this was observed, an equal increase or decrease.

Now with 6.45
Work 1 went up 9148s (152m)
Work 2 went up 2350s (39m)

152m and 39m are not the same or even close. There is now a 133m differnce in these two's averages vs the previous 2m difference.

The increase is not even close to the 57m increase of my home system, one is much more and 1 less.

[SCRATCHES HEAD]

[EDIT]
Another thing, if I add up all the run times for the first 36 tasks (30 day period) of work 1 and work 2, the totals 2,133,162 and 2,128,182 only have a difference of 4,979s (82m) total.

Now the gap between the two per task is larger than that 30 day gap.

keith,
on a 8800GT you should get approximately 70 ms/step.
This is consistent with the Linux application and it is what we had before when the application was using 100% of cpu.
I don't quite understand why your other 8800GT takes 80 ms /step still having the same shader frequency:

# Clock rate: 1512000 kilohertz
# Time per step: 74.176 ms

#

# Device 0: "GeForce 8800 GT"
# Clock rate: 1512000 kilohertz
# Time per step: 80.404 ms
#

Note that previous CPU time from older application is not a reliable measure of elapsed time, as the cpu was just polling the gpu and so, probably using less than 100%.

gdf

keith,
on a 8800GT you should get approximately 70 ms/step.
This is consistent with the Linux application and it is what we had before when the application was using 100% of cpu.
I don't quite understand why your other 8800GT takes 80 ms /step still having the same shader frequency:

# Clock rate: 1512000 kilohertz
# Time per step: 74.176 ms

#

# Device 0: "GeForce 8800 GT"
# Clock rate: 1512000 kilohertz
# Time per step: 80.404 ms
#

Note that previous CPU time from older application is not a reliable measure of elapsed time, as the cpu was just polling the gpu and so, probably using less than 100%.

gdf

I may be on to something now. Yesterday I switched off BOINCview both at home and on work #2. I did not have time to check results at home this morning.

But this morning after I rebooted both work computers, and then checked times. work 1 is about 60% done with 3:46 cpu time, about same as it was the first few. Work #2 is now 70% done with 00:55 cpu time. Work #2 never had one that quick yet.

My thinking was memory useage or network traffic, specifically communication with BOINC. BOINCview will acculate a lot of messages in memory, just as BOINC does, except my BOINCview is monitoring 16 computers. I'm going to leave it off the next couple of days and over the weekend to see what happens. The only other thing reduced by this is network traffic and communication with the BOINC client. Although this has nothing to do with GPU useage.

It is still too early to tell if that is a cause. I'm still PUZZELED.

Here are the last 5 results, all from 6.45 from work #2.

Something I did, I guess turning off BOINCview has made a difference, but in the wrong direction. GPU time is now more. CPU time is less though. CPU time is about the same as work #1, and GPU time is close but only about half way. This was the only change I made yesterday. You can see a drop in CPU and increase in GPU on the last two (first to on list). This I do not understand these results.

work #2
CPU---- Claimed Granted GPU time ms/step
478.53 3232.06 3232.06 64824.85 76.27
2951.58 3232.06 3232.06 63049.30 74.18
3513.09 3232.06 3232.06 61384.59 72.22
3540.48 3232.06 3232.06 61432.72 72.27
3456.95 3232.06 3232.06 61631.57 72.51

work #1
CPU---- Claimed Granted GPU time ms/step
522.89 3232.06 3232.06 69912.97 82.25
623.58 3232.06 3232.06 68343.72 80.4
285.72 3232.06 3232.06 68463.05 80.55

home since turning off BOINCview has shown no real change, CPU time is about the same as is GPU time.

home
CPU---- Claimed Granted GPU time ms/step
3439.88 3232.06 3232.06 6.45 58286.17 68.57
3567.89 3232.06 3232.06 6.45 58595.14 68.94
3703.58 3232.06 3232.06 6.45 60286.58 70.93
3545.36 3232.06 3232.06 6.45 59464.68 69.96
3436.73 3232.06 3232.06 6.45 59270.56 69.73
3723.67 3232.06 3232.06 6.45 58289.86 68.58

I'll keep looking, but I want to let things run a few days, especially over the weekend where I will not be at work so they will have two days undisturbed.

Computation error.

19.09.2008 1:31:47|PS3GRID|Restarting task RN18493-GPUTEST3-3-10-acemd_0 using acemd version 645
19.09.2008 1:31:50|PS3GRID|Computation for task RN18493-GPUTEST3-3-10-acemd_0 finished
19.09.2008 1:31:50|PS3GRID|Output file RN18493-GPUTEST3-3-10-acemd_0_1 for task RN18493-GPUTEST3-3-10-acemd_0 absent
19.09.2008 1:31:50|PS3GRID|Output file RN18493-GPUTEST3-3-10-acemd_0_2 for task RN18493-GPUTEST3-3-10-acemd_0 absent
19.09.2008 1:31:50|PS3GRID|Output file RN18493-GPUTEST3-3-10-acemd_0_3 for task RN18493-GPUTEST3-3-10-acemd_0 absent

I still have varied times between my one hosts at work.

Mostly this seems to be to BOINCview runnning on the host, but do not know why it interferes. I turned it off last weekend, and over the weekend all takes finished with less CPUtime. On Monday I restarted BOINCview and CPU times went back up and have reamined higher all week long.

If I try same test at home (shutting it down), it has no effect, times have remained the same.

Times are averages or approximate.

If running BOINCview the host has CPU time of about 57m and GPU time of 60,841s=1014m=16.90h

If not running BOINCview, the host CPU time drops to 4m but the GPU time goes up to 66,478s=1107m=18.47h, an increase of 1.57 hours.

It seems to me though, it would be wiser to give up about one hour of CPU time to save 2 hours of GPU time. Net gain is 1 wall clock hour less processing.

From a credit point of view
On this host, it earns about 2.5-2.75cs per minute for GPU and only 0.23cs per minute for CPU (from another project). This is just an estimate. So I would be giving up 14cs from the CPU to gain 300cs from the GPU as each task would finish 2 hours quicker, thus giving 2 extra hours of use to the GPU.

OK, these are not exact numbers, but the difference is big enough to justify the CPU usage. I think you can see the point.

The only question is how I make my other work computer which as always had a low CPU time use more CPU time. I guess I can install BOINCview on that one too.

Strange problem:

04.10.2008 6:54:33|PS3GRID|Started upload of yr20704-GPUTEST3-7-10-acemd_3_3
04.10.2008 6:54:35|PS3GRID|[error] Error reported by file upload server: Server is out of disk space
04.10.2008 6:54:35|PS3GRID|Temporarily failed upload of yr20704-GPUTEST3-7-10-acemd_3_3: transient upload error
04.10.2008 6:54:35|PS3GRID|Backing off 5 min 30 sec on upload of yr20704-GPUTEST3-7-10-acemd_3_3

What's up with disk space?!

fixed.

gdf

My GPU performance for this box has been on a decline...

http://www.ps3grid.net/result.php?resultid=83851

This is the last wu finished. The previous wu was not as bad. And before this, my result times for this GPU were:

# Time per step: 77.024 ms
# Approximate elapsed time for entire WU: 65470.047 s

I re-booted prior to starting the "83851" wu.

app 6.45
WinXP64 sp2
177.84 driver
6.3.14
8800GT

---

Consciousness: That annoying time between naps......

Experience is a wonderful thing: it enables you to recognize a mistake every time you repeat it.

I think I fixed it. It might have been another application (not related to Boinc) that was started by mistake, and hogged the cpu cycles the GPU needed.

I'll know in a day or so...

---

Consciousness: That annoying time between naps......

Experience is a wonderful thing: it enables you to recognize a mistake every time you repeat it.