There is the website (or local) preference "On multiprocessors, use at most xxx% of the processors", but this controls how many CPU cores BOINC can use as a whole, across all projects

I was under the impression that that's per project. (Otherwise, ¿how does that setting work when a client's attached to multiple projects?). In fact, I'm pretty sure changing that setting that will do the job, as I'm seeing jobs go out with odd numbers of cores assigned.

Matt

Well, for a start it's in the global preferences section: http://www.gpugrid.net/prefs.php?subset=global

and that section is headed (I copy from the GPUGrid website)

"These apply to all BOINC projects in which you participate.
On computers participating in multiple projects, the most recently modified preferences will be used."

Since the projects don't actually communicate directly between each other, this works by each user's host computer(s) acting as a message relay between all the projects that computer is attached to.

Each scheduler contact (both request and reply) contains the user's current global preference settings. That includes:

Current values (for all four venues)
Timestamp of last change
Project where the most recent update is believed to have been made

If any RPC contains different values from those held by the receiver, and if the timestamp is newer, then the receiving computer/project updates its internal records.

Provided there is a continuous chain connecting every project, and every computer, on the user's account, eventually all computers and all projects end up with an identical set of preference values for that global set. It only breaks down if a user has computers 1,2,3 attached to projects A,B,C - and computers 4,5,6 attached to projects D,E,F with no overlap. Then, there's no migration route between the first cluster and the second cluster.

Sorry MJH, but you are wrong on how that setting works.

So how does the client resolve conflicting settings from different projects?

The computer has resources, and the client keeps the resources occupied.

Unless a task is in jeopardy of missing deadline, the client feeds coprocessor resources first, and then feeds the rest of the CPUs. If a coprocessor job has been included in the "run-list", the client schedules [ncpus + 1] work of work. Otherwise, if only CPU jobs, it schedules [ncpus] worth of work. If a multithreading job enters the run-list, we are allowed to overcommit the CPU.

As an example, let's say we have a quad-core hyperthreaded machine, such that Windows sees 8 virtual CPUs. Let's say I have 2 ASIC coprocessor devices, 2 AMD GPU coprocessors that are the exact same model, and 3 NVIDIA GPU coprocessors that are all different models. Let's say we're attached to 30 different projects.

So, a possible schedule for this might be:
ASIC + 0.05 CPU
ASIC + 0.05 CPU
AMD + 0.25 CPU
AMD + 0.25 CPU
NVIDIA + 0.3 CPU
NVIDIA + 0.4 CPU
NVIDIA + 0.5 CPU
7 single-threaded CPU tasks
... all of differing projects. Note: The coprocessor CPU portions add up to: 1.8, and ncpus is 8, and so BOINC schedules 7 CPU tasks, such that the sum of CPU is 8.8, which is less than 9 [ncpus + 1].

Let's say we have the "Use at most x% CPUs" setting set to 75%, such that BOINC knows we only want to schedule for 6 virtual CPUs. Now a schedule might be:
ASIC + 0.05 CPU
ASIC + 0.05 CPU
AMD + 0.7 CPU
AMD + 0.7 CPU
NVIDIA + 0.3 CPU
NVIDIA + 0.4 CPU
NVIDIA + 0.5 CPU
1 6-threaded MT CPU task
... all of differing projects. In this example, for the coprocessor jobs, the sum of CPU portions is 2.7. Even though 2.7 + 6 = 8.7 is larger than [ncpus + 1 = 7], we are allowed to overcommit the CPU to get that MT task to process.

Hope this helps explain the scheduler a bit better. TLDR; The "Use at most X% CPUs" just tells BOINC how many CPUs it should consider available for it to allocate across the list of projects and tasks.

Your app here is the first I've ever seen that is both MT and GPU Coprocessor. I expect you to have some scheduling problems with it :)

One final note: The server only knows the make/model of the highest-powered GPU, for each GPU type (NVIDIA or AMD or Intel), I believe. You probably already knew this. Not sure if that will affect your task assignment logic.

If you allow multi-gpu setups to get work we can limit GPUGRID.net to one GPU via the <exclude_gpu> tag in cc_config.xml.

---

about the AMDTEST wu:s
-on linux they did not download at all; same driver version as on w7; boinc asked only for cpu wu:s, even if it detected the amd card in the logs
-on w7 they finish without error, but do not follow the amount of dedicated cpu:s, they run on 4 cores regardless of local specification

about the AMDTEST wu:s
-on linux they did not download at all; same driver version as on w7; boinc asked only for cpu wu:s, even if it detected the amd card in the logs
-on w7 they finish without error, but do not follow the amount of dedicated cpu:s, they run on 4 cores regardless of local specification

If you read through this thread you can read that Matt says, it is Windows x64 only. Hope this helps.

---

Greetings from TJ

No more WUs?

No more WUs?

Filled up now.

The fix credit 1500 per WU set. or?
Some tasks errored out after only-CPU stage. This and that.
part of stderr

clCreateProgramWithSource
starting mdrun 'Protein in water'
200000 steps,    400.0 ps.
Inner product between old and new vector <= 0.0!
constraint #0 atoms 1730 and 1731
Wrote pdb files with previous and current coordinates

WARNING: Listed nonbonded interaction between particles 1713 and 1731
at distance 2.985 which is larger than the table limit 1.995 nm.

This is likely either a 1,4 interaction, or a listed interaction inside
a smaller molecule you are decoupling during a free energy calculation.
Since interactions at distances beyond the table cannot be computed,
they are skipped until they are inside the table limit again. You will
only see this message once, even if it occurs for several interactions.

IMPORTANT: This should not happen in a stable simulation, so there is
probably something wrong with your system. Only change the table-extension
distance in the mdp file if you are really sure that is the reason.


Inner product between old and new vector <= 0.0!
constraint #0 atoms 73 and 74
Wrote pdb files with previous and current coordinates


Unhandled Exception Detected...

- Unhandled Exception Record -
Reason: Access Violation (0xc0000005) at address 0x000000013F573B37 read attempt to address 0x123DE8A0

Engaging BOINC Windows Runtime Debugger...

If you allow multi-gpu setups to get work we can limit GPUGRID.net to one GPU via the <exclude_gpu> tag in cc_config.xml.

Is this a possibility?

Would groveling help? :D





lol

Cheers!

---

Not just yet. It will require a change in the application too

Matt

How about credits? How are they calculated per WU? For now it's 1500 per WU but on XEON 1230V2 + 7950@1100 it gives ~ 160000 daily? Sounds little disappointing cause it takes 8 threads of CPU. Do You plan any adjustment in future or is it final?

How about credits? How are they calculated per WU? For now it's 1500 per WU but on XEON 1230V2 + 7950@1100 it gives ~ 160000 daily? Sounds little disappointing cause it takes 8 threads of CPU. Do You plan any adjustment in future or is it final?

A fixed amount of credits are now allocated per WU - like acemdlong , it's 1000 + 50% bonus for prompt return.

The actual amount of credits are still a placeholder value, while we are in beta.

At the moment this is a beta app which does not make good use of the gpu and it does not make much differece how many cpu's are used.
Unfortunately the project does not allow to download and run 2 or even mor wu's. So I looked for a way to increase the gpu usage to values from 93% to 100%.
https://dl.dropboxusercontent.com/u/50246791/gpugrid%20setup%201.JPG
Works on three pc's, all wu finish and validate. But may be different with a new version.

Alexander

I just did a new build that has onboard AMD Radeon(TM) R7 Graphics. The driver version is Catalyst 14.12. I am not getting any work for this task on it.

MSI Afterburner is seeing 5 different lines/instances for this GPU. I also put 2 GTX780 cards in this build. I have the device management (for the profile called home on the GPUGRID preferences page of the website) set to use 100% of everything and to run everything except the CPU specific items. (As follows)
--------------------------------------------------------
Resource share 100
Use CPU yes
Use NVIDIA GPU yes
Run test applications? yes
Is it OK for GPUGRID and your team (if any) to email you? yes
Should GPUGRID show your computers on its web site? yes
Default computer location home
Maximum CPU % for graphics 100
Run only the selected applications
ACEMD short runs (2-3 hours on fastest card): yes
ACEMD beta: yes
ACEMD long runs (8-12 hours on fastest GPU): yes
Docking CPU only app: no
Molecular Dynamics on CPU: no
Molecular Dynamics on AMD GPUs: yes
If no work for selected applications is available, accept work from other applications? no
Use Graphics Processing Unit (GPU) if available yes
Use Central Processing Unit (CPU) yes
---------------------------------------------------------

Regardless of these settings and the fact the there are WUs showing on the Server Status page, this system is not getting any of these units. It did get one unit as I was building, before I added the GTX cards and it had an error while computing. This was my fault for adding software while still adding hardware. I was adding cards as they finished WUs on their originally installed computer before transferring to this new one. The new system is http://www.gpugrid.net/results.php?hostid=197994 This ADM GPU is showing as AMD Spectre (762MB) on the PC reference page here http://www.gpugrid.net/show_host_detail.php?hostid=197994

Do I have a setting or misconfiguration or do you need more info to determine? Or is this GPU just m=not able to do the work? Thx

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org

Tasks aren't assigned to machines with both Nvidia and AMD cards.

Matt

Tasks aren't assigned to machines with both Nvidia and AMD cards.

Matt

OK, Thanks. I went AMD over Intel on the board so that maybe I could double up the tasks being done with AMD and NVidia. Hopefully I can get use out of the onboard GPU somewhere.

916 is up.
Should report progress and run on hosts with multiple gpus

That explains why my 2x AMD GPU rig got 6 tasks. Says they are 9.15 (mt) though.


Never underestimate the power of proper groveling. ;-)


Cheers!

---

Vista 6 core 2x 7870 GPUs machine got 9.15 (mt). Using <exclude_gpu> it was on GPU 0. Seldom touched the GPU CPU usage was 0-100 on all cores bouncing around. While Time Remaining stoped after about 10 minutes WU continued for, well it is at 97% and has run 44:55.

Windows 8.1 6 core 2 x 7870 9.15 (mt) did hit the gpus for 30% for a few minutes then to 0% for a few. CPU was all cores at about 20-30% and computation error after 119-122 seconds.

---

Vista 6 core 2x 7870 GPUs machine got 9.15 (mt). Using <exclude_gpu> it was on GPU 0. Seldom touched the GPU CPU usage was 0-100 on all cores bouncing around. While Time Remaining stoped after about 10 minutes WU continued for, well it is at 97% and has run 44:55.

After running for 2:46:41 and 14 minutes of that at 100% complete I aborted.

---