So,
apparently something has changed during the last server update. The behavior is not normal. I have asked the developer about it, I could move back to previous server release, but I would like to hear from then first.

gdf

Problem found and fixed. Now the behavior is as before: 1GPU+1CPU

When there will be no performance drop in using the 0 CPU, we will do the change in the policy.

gdf

Thank you GDF

Problem found and fixed. Now the behavior is as before: 1GPU+1CPU

When there will be no performance drop in using the 0 CPU, we will do the change in the policy.

gdf

Thanks gdf

Just got a new task and it seems the behavior is back to normal (switched back to use 100% CPU Cores and 1xGPU + 1xCPU Tasks are running)

We know to appreciate your work and your quick finding & fixing the problem.


Cu KyleFL

This should be configurable by the user. Someone like to use all cores for other projects and someone not.

A GTX280 needs for one WU

with ncpu-entry 8h 24m, thats 384,77 Credits/h.
without ncpu-entry 10h 28m, thats 308,99 Credits/h.

The difference is here 75,78 cr/h. The additional usable core with ncpus=5 can't reach such a amount of credits in one hour, also not on PrimeGrid. There I get 53,23 cr/h while crunching PSP Sieve.

So for a GTX280 a ncpu-entry cores+1 is not advisable.

That takes not the same effect on a 8800GT or smaller graphic card.

You have to calculate for your own.

---

....... Still have the issue - Quad still wants to run 5 processes, and the PS3Grid process runs veeeerrrrryyyyyy slowly - how can this be fixed?

....... Still have the issue - Quad still wants to run 5 processes, and the PS3Grid process runs veeeerrrrryyyyyy slowly - how can this be fixed?

It may be look very slowly, but it's only showing the CPU using time and not the real time. The system should be run in normal state after the next connect to the server (if you don't have created a cc_config.xml with a ncpus-entry in the BOINC-data directory).

The new server software has a flag for GPU using software to add a non existent core for crunching with all CPU-cores and a additional virtual core for the GPU. This flag is switsched by GDF, so your client should work as normal after the next connect.

---

It may be look very slowly, but it's only showing the CPU using time and not the real time. The system should be run in normal state after the next connect to the server (if you don't have created a cc_config.xml with a ncpus-entry in the BOINC-data directory).

The new server software has a flag for GPU using software to add a non existent core for crunching with all CPU-cores and a additional virtual core for the GPU. This flag is switsched by GDF, so your client should work as normal after the next connect.
[/quote]


Thanks Kokomiko - I can see from today's WU that the step time has moved from 42ms to 70ms - so something has slowed my GTX260. Unfortunately, the project is mostly offline at the moment... so I'll have to wait for the cure!

Burdett:

You can do a workaround in your BOINC settings:

Just tell your BOINC, that (on Multiprozessor Systems use only xx Cores)

On a Quad: 75%
On a Duo: 50%

That should get you back to normal as long as the "wrong" server flag is active.
After getting the next new work, stop the cpu taks, take the setting back to 100% and restart the CPU Task.

That did the trick for me.


Cu KyleFL

Burdett:

You can do a workaround in your BOINC settings:

Just tell your BOINC, that (on Multiprozessor Systems use only xx Cores)

On a Quad: 75%
On a Duo: 50%

That should get you back to normal as long as the "wrong" server flag is active.
After getting the next new work, stop the cpu taks, take the setting back to 100% and restart the CPU Task.

That did the trick for me.


Cu KyleFL

Done & working - thanks!

I can't manage to make the cc_config.xml trick work anymore... any other with the same issue?

I can't manage to make the cc_config.xml trick work anymore... any other with the same issue?

What have you tried? Cores +1 or -1? What's your goal? Running one virtual core more or reserve one core for feeding the GPU with max speed?

---

I have a dual core and I tried Ncpus 1, 2 and 3 (-1, 0, +1) but none of the options worked.

I want to run two "normal" tasks + the GPUGrid one.

The fact is that, whatever I wrote in the cc_config file, it just ran normal tasks, with the GPUGrid one "waiting for elaboration".

Without the cc_config file, GPUGrid works but it reserves a core for itself (1 normal task + the GPUGrid one).

I have a dual core and I tried Ncpus 1, 2 and 3 (-1, 0, +1) but none of the options worked.

I want to run two "normal" tasks + the GPUGrid one.

The fact is that, whatever I wrote in the cc_config file, it just ran normal tasks, with the GPUGrid one "waiting for elaboration".

Without the cc_config file, GPUGrid works but it reserves a core for itself (1 normal task + the GPUGrid one).

You´ll get ~30% longer computing time on the GPU-WU if you run a CPU-WU on the Core that the GPU-WU is using.
Is it possible that you did put the cc_config file in the wrong directory?
Or that you didn´t restart your Boinc after making your changes?


Cu KyleFL

He is running Linux, there the slowdown is almost not measurable, at least on my machines there seems to be no impact to the "time per step" value when I set ncpus+1

Plus I added the following entry to roots crontab to increase the priority of the acemd process automatically:

* * * * * renice 5 `ps -ef | grep acemd | grep -v grep | awk '{ print $2 }'`

So for a GTX280 a ncpu-entry cores+1 is not advisable.

That takes not the same effect on a 8800GT or smaller graphic card.

You have to calculate for your own.

I take it from reading this I can run the ncpu-entry cores+1 if I have a 8800GT because it doesn't need a CPU Core reserved for it ... ???

So for a GTX280 a ncpu-entry cores+1 is not advisable.

That takes not the same effect on a 8800GT or smaller graphic card.

You have to calculate for your own.

I take it from reading this I can run the ncpu-entry cores+1 if I have a 8800GT because it doesn't need a CPU Core reserved for it ... ???

No, it´s just math:

A 8800GT needs for one WU (just a rough estimaion of me)

with ncpu-entry ~16h , thats ~200 Credits/h.
without ncpu-entry ~20h thats ~160 Credits/h.

The difference here would be ~40cr/h as opposed to the 75,78 cr/h difference if a GTX280 would be involved. The additional usable core with ncpus=5 could propably reach such a amount of credits in one hour, for example on PrimeGrid. There You could get ~50 cr/h while crunching PSP Sieve.
So on a 8800GT the NCPU+1 "could" make sense.

But as already said -- you´d have to do the math yourself for your card, for you´re CPU (propably is your CPU slower and can´t get 50cr/h - or your GPU takes a greater hit on performance).


Cu KyleFL

He is running Linux, there the slowdown is almost not measurable, at least on my machines there seems to be no impact to the "time per step" value when I set ncpus+1

Plus I added the following entry to roots crontab to increase the priority of the acemd process automatically:

* * * * * renice 5 `ps -ef | grep acemd | grep -v grep | awk '{ print $2 }'`

Oh, sorry -- I didn´t get it from his post, that he was running on linux. That - of course - would be a different story.

On Windows I would wait a few additional weeks, until the problem with the scheduler is sorted out and the NCPU+1 can be run without a penatly.
Maybe it´s possible to let Boinc run the GPU-WU with a higher prozess priority than the CPU WUs. That way the GPU-WU always would get CPU time and the CPU-WU would only take the idle-cycles left...


Cu KyleFL

I have a dual core and I tried Ncpus 1, 2 and 3 (-1, 0, +1) but none of the options worked.

I want to run two "normal" tasks + the GPUGrid one.

The fact is that, whatever I wrote in the cc_config file, it just ran normal tasks, with the GPUGrid one "waiting for elaboration".

Without the cc_config file, GPUGrid works but it reserves a core for itself (1 normal task + the GPUGrid one).

You´ll get ~30% longer computing time on the GPU-WU if you run a CPU-WU on the Core that the GPU-WU is using.
Is it possible that you did put the cc_config file in the wrong directory?
Or that you didn´t restart your Boinc after making your changes?


Cu KyleFL

As someone suggested I'm running Linux.

It's not possible that I put the file in the wrong directory because it was there before the server change and it made the trick work. After the server change *puff* it didn't work anymore.

And I'm sure that I restarted BOINC *every* time so that's not the issue.

I really have no idea of what happened but I know that running 1 normal WU and 1 GPUGrid WU pisses me off because it wastes 95% of one of my cores

[quote
On Windows I would wait a few additional weeks, until the problem with the scheduler is sorted out and the NCPU+1 can be run without a penatly.
Maybe it´s possible to let Boinc run the GPU-WU with a higher prozess priority than the CPU WUs. That way the GPU-WU always would get CPU time and the CPU-WU would only take the idle-cycles left...


Cu KyleFL

Tried that with manual assignment of priority with windows-scheduler. Didn´t work, the problem lies elsewhere.