Looking at this task I had 33 some hours on the CPU clock and I know it was running for well over a day ... yet the internal time reported was 48K seconds (~13 hours) ...

And if I understand the payment schedule this was supposed to be one of the shorter run time tasks ...

Side note, can you add the run time to the output file, I mean, LHC had tasks of 100,000; 1,000,000, and 10,000,000 turns and you could tell on the client which type it was. Here, we can only guess based on the credit paid (as far as I can tell) ...

Anyway, I am real curious about this one ... sure wish I was not spending so much CPU time on the tasks ... any chance we can get the 6.56 back?

Well, that does not explain why the next task which started auto-magically is running with an estimated time of 4 hours to go with 7 hours done and 70% complete ...

I will admit to not trying to find out the clock speed and will look at that next time ... But, the machine is always on, always running BOINC and For the moment I am not doing anything else.

The only RADICAL thing I have done on the machines I am running these last couple days is slowly change the project mix adding some on one machine or the other ... on the i7, the machine in question I added Rosetta ... and then when I had tons of failures I took them back down (well one task left to run ...

Nvida monitor shows GPU core 3D at 601 MHz and GPU memory at 900 MHz CPU 100% memory ~55% For this card there does not seem to be adjustments that can be made by this tool (Assuming it can do it at all) I was waiting for my golden chariot before I started asking silly questions ...

But, that begs the question ... the clock I am seeing seems to me to be lower than what you say it should be and this task is running faster ...

{edit} isn't 1.5 GHz faster than 600 MHz ... it was when I went to school... {/edit}

Well, that does not explain why the next task which started auto-magically is running with an estimated time of 4 hours to go with 7 hours done and 70% complete ...

I will admit to not trying to find out the clock speed and will look at that next time ... But, the machine is always on, always running BOINC and For the moment I am not doing anything else.

The only RADICAL thing I have done on the machines I am running these last couple days is slowly change the project mix adding some on one machine or the other ... on the i7, the machine in question I added Rosetta ... and then when I had tons of failures I took them back down (well one task left to run ...

I am running Rosetta[1 hr run time] along side my gpu without issue.

Cpu is stock as well as Gpu. Not real speedy but it seems[at the moment] to be dependable.

mike

Rosetta is all about experimenting with new stuff, be it algorithms or applications ;)

But back to topic.. the 600 MHz you're seeing is the core clock. This one is probably alright now. Then there's memory clock (also fine) and since G80 nVidia chips have one more, separate clock domain for the shaders. It's around 1.512 GHz for a stock 9800GT. These clocks are what is reported in the task output. So if your 1.5 GHz shaders run at 0.6 GHz you can easily see why the WU was much slower than normal ones. I can't tell you though why the chip was clocked down.

MrS

I think 0.04 CPU tells everything.
Formerly 0.9 and lowered since App6.56 to 0.04. We falling back to App6.55 and no fallback in CPU-Config for the Boinc Scheduler had happend.
But WHY???
My 3232 Credit WU Runtimes are increased from ~39000s to ~42000s.
Maybe the old 0.9 CPU settings should be also restored.

I am not sure if I have re-created the problem on my other machine with this task ...

The other machine has an 8500GT which I have decided is not for this application and the task has 89 hours on it with 5 to go ...

GPU 100%
Memory 44% (1G)
Core 459 MHz
GPU Memory 300 MHz

Temp 71C

Figured out how to get into the control panel to look at internal settings ... :)
Clock is at factory defaults.

I don't think I will "tune" this system ...

Again, not complaining, I am just wondering if there is POSSIBLY something wrong with a couple tasks, or if I just got lucky ... :)

I sent a note to the guy that got issued this task as I was planing to run it to completion ... if nothing else, maybe we can look at it under a microscope?


Oh, the second task on the i7 since Godzilla seems to be running at normal speed ... 10% done after only a hour or so .. on the clock ...

Well, this task ran to completion with over 90 hours of run time.

It validated and I am not sure if the server is going to cancel the re-issue or not. At the moment it says it isn't ...

I did try the two cards in the one computer and ran into another issue and I wrote a bit about it on the Dev mailing list if you are interested. The bottom line is that with the CPU run time bug in 6.55 it is not worth it to run two GPUs in the same machine. You have the appearance of two extra tasks, but the overhead is such that you are going to lose one core ... or so much of one it does not matter.

One task was pulling 5% CPU and the other was at 8% which on an 8 core means you have lost a whole CPU ...

*SO*, I moved the card to the computer with the 8500 and pulled that card out ... and if you look at the task notes you can see the last few minutes were run with the 9800 GT card. I was a little worried that making the switch would cause problems ... but ... SOMETHING went right...

Now I am trying to get the display back on the other computer. I was getting weird aliases on the screen so I suspect I got a bad driver install ... windows sure does not pick up that the currently avaialable drivers software is Ok to use on the new card.

The bottom line is that with the CPU run time bug in 6.55 it is not worth it to run two GPUs in the same machine. You have the appearance of two extra tasks, but the overhead is such that you are going to lose one core ... or so much of one it does not matter.

One task was pulling 5% CPU and the other was at 8% which on an 8 core means you have lost a whole CPU ...

That's a new topic, but I'll comment anyway.

Until we can get 6.56 back we have to accept that a GPU needs CPU power. About a half core per GPU (in earlier apps it was an entire core). And it doesn't matter if 2 GPUs use 2 half cores in one machine, i.e. an entire core, or if you put the 2 GPUs in separate machines and they need a half core each in both machines.

The sum is the same. So I really disagree with "it is not worth it to run two GPUs in the same machine". However, you could conclude that you'd rather loose 2 slow half-cores than 2 fast ones. But that should have been pretty obvious anyway ;)

And I suspect it's better to sacrifice one logical core of the i7 than to sacrifice e.g. one physical core of a C2Q. Some proper measurements would be needed, though.

MrS

Those clocks sound like they could be the 2d clocks... Perhaps try ATI tool to modify the 2d clocks to equal 3d clocks...

Its worth a try...

(Yes ATI tool works with nVidia cards :) )

Well, you lose roughly the same computer power, probably a little less I think ... One expense you missed and maybe I was not clear is that you have the appearance of one other task running when in fact it is starved for CPU time. I did not wait to run tasks to completion but if BOINC does not catch on you could conceivably have one task in de facto suspension while BOINC thinks it is running ... then comes the deadline and you blow the deadline.

I still don't understand. Let's say you have a quad and 2 GPUs and you run 4+2. Then you get 4 tasks at low priority (win scheduler) and 2 tasks at normal priority. The basic load if the machine are the 4 low tasks and the 2 normal tasks "steal" time slices from 2 different cores, at least most of the time (again, the windows scheduler handles this).

Either I'm totally mistaken or I can't find the idle / starved task.

MrS

Ah, now I get it!

As far as I know BOINC doesn't set CPU affinity, which is good. In windows any default tasks can use any CPU and the scheduler shuffles them around in a strange manner (Linux also does it this way). A task is only pinned to a specific core if the CPU affinity is set accordingly. So in your 8+2 case the CPU time for both GPUs would be taken from all 8 cores on average.

MrS

Ah, now I get it!

As far as I know BOINC doesn't set CPU affinity, which is good. In windows any default tasks can use any CPU and the scheduler shuffles them around in a strange manner (Linux also does it this way). A task is only pinned to a specific core if the CPU affinity is set accordingly. So in your 8+2 case the CPU time for both GPUs would be taken from all 8 cores on average.

MrS

Ah, now I get it!

As far as I know BOINC doesn't set CPU affinity, which is good. In windows any default tasks can use any CPU and the scheduler shuffles them around in a strange manner (Linux also does it this way). A task is only pinned to a specific core if the CPU affinity is set accordingly. So in your 8+2 case the CPU time for both GPUs would be taken from all 8 cores on average.

MrS

But not here. 3 tasks are running with 25%, on task with 21% and the acem... - task takes 4%.

Seems like in theory reality matches theory much better than in reality..

Why would a scheduler run the only 2 tasks with normal priority on the same core, whereas all other cores run tasks with low priority? That seems stupid! In my WinXP I installed BOINc in protected mode, which means that I can't access the tasks in task manager. Would anyone mind to check the affinity settings? In task manger rightclick the acemd.exe and choose the last option, don't know how it's called in english. All vailable cores should be selected, otherwise BOINC (or the user :p ) did set CPU affinity.

MrS