I just started running CUDA on a quad core with two 8800 GTS's using 6.3.19. I get two CUDA units running but only two units on the quad core where I used to have four running.

I notice that in the first few seconds I get four quad core units running but when the two CUDA units start, two of the other units stop.

How do I get six units running at the same time?

More good news, A change has been checked in, next client may be perfect.

David 29 Oct 2008 client: the preemptability order was wrong

Now: // and we want the best replacement at the front,
New: // and we want the most preemptable task at the front

See changeset 16355

This too might explain why not everyone sees this, or it is intermittent and random, as each client has different proejcts, number of tasks with different deadlines etc.

[cross fingers]

If not, well try again......

That's good news and I think good feedback from everyone here :)

MrS

---

Scanning for our furry friends since Jan 2002

Have you noticed a performance decrease between leaving one CPU to manage the GPU task, or using it for other workunits (provided that you are using 6.3.19!!!)?

gdf

Have you noticed a performance decrease between leaving one CPU to manage the GPU task, or using it for other workunits (provided that you are using 6.3.19!!!)?

gdf

Not specifically.

Becuase before I was running without ncpus+1 so my sytems wer running less than 100% CPU usage. They run mostly fine now. I do see a little gain in GPU elapsed time, but I've had to ignore most results because testing can throw off elapsed time. I check the numbers later and see, but it is small, like 1,000 to 2,000 seconds.

I did notice the applciation checkpoint about every minute. I think that consumes a lot of resources. Is it necessary to checkpoint that frequently ?

10/29/2008 4:17:55 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed
10/29/2008 4:18:27 PM|Docking@Home|[checkpoint_debug] result 1t7k_mod0013sc_10741_42324_0 checkpointed
10/29/2008 4:19:06 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed
10/29/2008 4:20:18 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed
10/29/2008 4:21:29 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed
10/29/2008 4:22:43 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed
10/29/2008 4:23:56 PM|PS3GRID|[checkpoint_debug] result EX30463-GPUTEST4-0-10-acemd_0 checkpointed

I have seen my times go up on a Q6600 with a gtx260 from 28000 sec to over 30000 sec. Slight increase to free up that other core. FYI

---

GTX280


6.3.14 + 6.48
GPU=1

Time per step: 26.120 ms
Time per step: 27.412 ms

6.3.14 + 6.48
GPU<1

Time per step: 31.970 ms
Time per step: 49.019 ms
Time per step: 34.924 ms
Time per step: 35.736 ms

6.3.19 + 6.48
GPU<1
Time per step: 31.853 ms just one WU until now)


It`s just if some1 wants to know.

---

Aaaahhhh!

My 6.3.19 lost connection to its BOINC client, or more precisely: the BOINC client quit. The last log entry was a file transfer.. reminds me of the problem I had with 6.3.14.

MrS

---

Scanning for our furry friends since Jan 2002

I found something in stderr.out on one WU:

No heartbeat from core client for 30 sec - exiting

Should I be worried about it or "have a break, have a kitcat"?

same problems for me on vista x64

gpu is still starved of cpu time running 4 cpu tasks and 1 gpu. Changing to real time priority for the gpu didn't help

with 4 cpu and 1 gpu
http://www.gpugrid.net/result.php?resultid=98085

with just 1 gpu running and no other tasks
http://www.gpugrid.net/result.php?resultid=96758

Problems with vista32 - have to suspend other project to get new tasks.

Request for next version:
Have the GPU reserve/use 16% of CPU to have better performance.
eg.
quad core = 4 cpu task + 1 GPU (but GPU uses 16% of CPU)

the remaining 9% will be for other cpu task.

As of now, I set Boinc to use 3 cpu to have a better performance for GPU but losing 9% CPU doing nothing.

I don't think a setting like you suggest x-files 27 will work for everyone even though it might be perfect for you.For example one of my hosts uses approx 8% of CPU and another uses 0-2%.

Until they can develop a client that has a sliding scale on cpu usage applied to the gpu a set % will not work well. I don't think the developers are even close to this option.

More performance tests for Windows and Linux would be useful. Keep posting.

gdf

It happened again! Half a day of idle time :/

Now I rebootet my machine, but I seriously doubt that will help. I found a file "stderrdae.txt" in the user data directory where BOINC logs unhandled exceptions. There are several entries from trying the 6.3.14 client and 2 entries from today. There's really a lot of information and from briefly looking over it the errors seem to be pretty similar.

@Rebirther: that seems to be the same problem I am having.. at least I also get this message in the output when the error ocurred.

Edit@GDF: should I send you the file with the debug information or to the developers?

MrS

---

Scanning for our furry friends since Jan 2002

3 CPU + 1 GPU
25.859 ms
21980.556 s

4 CPU + 1 GPU
46.794 ms
39774.961 s

Having the GPU use only like 1% CPU clearly loses some performance...about 55%

It's difficult to judge times, but I'd say that with 4+1 i'm not loosing more than 4% compared to 3+1. In the case of 4+1 the GPU task gets between 6 and 20% of my quad, whereas 3+1 leads to 16 - 20% utilization.

MrS

---

Scanning for our furry friends since Jan 2002

It's difficult to judge times, but I'd say that with 4+1 i'm not loosing more than 4% compared to 3+1. In the case of 4+1 the GPU task gets between 6 and 20% of my quad, whereas 3+1 leads to 16 - 20% utilization.

MrS

There's really no way to Judge Times because the Client does what it feels like when it feels like it. I set mine to 5 ncpu's and it will run 4 @ times, I set it to 4 ncpu's and it will run 5 @ times, so how you gonna say it's better one way or the other when the Client doesn't obey what your settings are ???

Unless you can sit there & watch the Wu run the entire time to see if it run's the whole time @ a 4 or 5 setting then you don't really know what it did the entire time ...

I did set my Client to use only 1 CPU on my DualCore.

That way only one Seti WU and one GPUGrid WU are running.

Times are exactly as before (no wonder - the GPU WU gets one whole Core to mess with)
I did try one GPU-WU running together with 2 CPU WUs -- the running time seemed to be >10% longer than without - so I didn´t want to lose to much speed and decided to not test that any longer.


Cu KyleFL

it's the same over here - VISTAx64 and XP32 - boinc currently is not able to assign resources to projects and fools around. it's everything:

1st: running fewer CPU-tasks than possible
2nd: suspending GPU-tasks to run CPU and vice versa
3rd: not getting CPU-WUs (and running dry) because GPU-task filled the queue
4th: requesting GPU-tasks according to number of CPUs (who the hack got that idea?)

currently it's FUBAR!

frank.

I just did the Linux64 upgrade from Boinc..16 to ..19. And re-attached to the GPUgrid project site. Seems to be fine.

I did notice my (quad) Linux64 box running (only cpu tasks) for a few minutes prior to the upgrade. But it changed to 4+1 just before I performed the upgrade, and it is still running 4+1. If it has been switching back and forth, I have not seen an increase in my wu completion times (8800GS). The Linux64 GPU has been running a steady 1-2% cpu usage for Boinc..16 and now ..19.

I have not run out of werk for either GPU or CPU (yeti...).

---

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

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