ComputerID 57387
WorkUnit 1250892

A couple of hours ago I downloaded 2 ACEMD beta version v6.18 (cuda) workunits & began crunching the first immediately. I noticed right away that the CPU:GPU ratio is extremely high, well above 90%. And yet the application identifies its resources as "0.27CPUs + 1.00 NVIDIA GPUs". BoincTasks, (which displays "CPU %" dynamically), shows CPU usage on the task at ~95.75%. BOINC Manager (which does not) confirms this in the task properties sheet with a ratio of CPUTime:ElapsedTime consistently in the same range, 95~96%.

Indications with this beta are dramatically different from ACEMD - GPU molecular dynamics v6.03 (cuda), which in my task history shows a CPU:GPU load of ~10%. So the question becomes, where are all those CPU teracycles going?

---

http://www.gpugrid.net/forum_thread.php?id=2059

---

Thanks - Steve

http://www.gpugrid.net/forum_thread.php?id=2059

Thanks. I especially like the statement(s) by GDF: [quote}The 100% CPU core is a not intended and will be removed. Don't know why is there.[/quote]

In my experience (35 years +) that sounds more like alpha (or pre-alpha) than beta SW.

"But that's just my opinion, I could be wrong."
-D. Miller

B/W

The GPUGRID team is small with limited resources and therefore depends more heavily on cruncher support for testing than other projects.

I know where you are coming from when you say the 6.18s were more like an Alpha, but they are not Alphas because most worked! Betas typically have some unwanted side effects, such as high CPU usage, and some even fail on different systems (a high failure would make them more like Alphas).

I consider Alphas to be more along the lines of the first full program and
these are tested/debugged on inhouse test machines with high failures and lots of unwanted side effects.

These Beta applications and work units are all based on existing working models, so that in itself qualifies them to be called Betas, and that makes them more like service pack Betas than application betas.

I would say you could better characterise tasks into Beta one and Beta two groups; with the 6.18 being Beta 2 (not up to pre-release standards). The 6.20s and 6.21s on the other hand would be first Betas, at best, as they all failed due to identical problems.
Fortunately all these tasks failed immediately, so they had little impact on user contribution; just a bit of bandwidth overhead.

I am seeing the same behavior under linux. Ubuntu 8.10/ NVIDIA UNIX x86_64 Kernel Module 195.30/Intel2140@1.60GHz/ 2G ram.

top - 10:20:09 up 10 days, 21:13,  1 user,  load average: 3.00, 3.00, 3.00
Tasks:  71 total,   4 running,  67 sleeping,   0 stopped,   0 zombie
Cpu(s):  0.0%us, 31.7%sy, 54.0%ni, 14.3%id,  0.0%wa,  0.0%hi,  0.0%si,  0.0%st
Mem:   2046860k total,  1784540k used,   262320k free,   202524k buffers
Swap:   658624k total,     2748k used,   655876k free,   322692k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
18432 boinc     30  10  202m  97m  26m R  100  4.9 499:41.54 acemd2_6.04_x86
20405 boinc     39  19  263m 255m  288 R   50 12.8 150:00.46 minirosetta_2.0
22890 boinc     39  19  303m 296m  256 R   50 14.8  23:11.69 minirosetta_2.0
 4597 boinc     20   0 95860  10m 3292 S    0  0.5  45:41.11 boinc
23296 boinc     20   0 18968 1272  988 R    0  0.1   0:00.41 top
15917 boinc     39  19  432m 426m  688 S    0 21.3   0:00.40 minirosetta_2.0
20406 boinc     39  19  263m 255m  288 S    0 12.8   0:00.14 minirosetta_2.0
20407 boinc     39  19  263m 255m  288 S    0 12.8   0:03.03 minirosetta_2.0
20408 boinc     39  19  263m 255m  288 S    0 12.8   0:00.16 minirosetta_2.0
22891 boinc     39  19  303m 296m  256 S    0 14.8   0:00.02 minirosetta_2.0
22892 boinc     39  19  303m 296m  256 S    0 14.8   0:00.44 minirosetta_2.0
22893 boinc     39  19  303m 296m  256 S    0 14.8   0:00.02 minirosetta_2.0
23277 boinc     20   0 74584 1732 1012 S    0  0.1   0:00.02 sshd
23278 boinc     20   0 20512 3388 1536 S    0  0.2   0:00.11 bash

I see you are running 4 tasks at once, on a dual core CPU with one GT240.
At least all the GPUGrid tasks are successful. I'm sure they will fix the high CPU usage in due course.

I am seeing the same behavior under linux. Ubuntu 8.10/ NVIDIA UNIX x86_64 Kernel Module 195.30/Intel2140@1.60GHz/ 2G ram.

I moved the gt240 to an xp32 box and the cpu utilization dropped to normal.

taskman shows acemd2_6.03_windows_intelx86__cuda at 0% cpu most of the time, occasionally peaking to 1 or 2% on a q6600.

It was taking 100% of one core on the above mentioned ubuntu machine up until I did the switch last night.

That was a mistake!
Linux uses a full CPU core, but the GPUGrid task is about 30% faster!!!
OK, you lose a CPU core, but you gain 30% GPU speed - it's a no brainer. Even an everage GPU does about 10 times the work of a CPU, so sackrafice the CPU for the GPU.

The reason why this version uses a full core is that some people experienced problems with the low-cpu version, like application hanging and so on.
Probably the driver is not so stable under Linux for low-cpu.

gdf.