GPU Upgrade Shows No Improvement in Work Unit Completion

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

RE: I don't understand how

Quote:

I don't understand how to do the second point he makes about P0.

Phil

Here's the relevant thread. The download you want is called nVidia Inspector:

https://einsteinathome.org/node/197852

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109405164623
RAC: 35451132

RE: ... my last PCIe 2.0

Quote:
... my last PCIe 2.0 x16 slot will only accept a single slot card.


Please realise the second slot may be x16 'mechanically' (you can plug in a GPU card) but it won't be x16 'electrically'. There is an overall limit on PCIe 'lanes' and it's likely that the slot in question is really only x4 which would probably have a negative effect on performance even if you did have enough room to plug in a higher end card. x4 probably doesn't impact adversely on a low end card which doesn't have such high bandwidth requirements.

With regard to your questions, I'll answer them fully, even though subsequent messages do indicate that you have figured out certain things. I'm always mindful of the many people 'lurking' without asking questions who may derive a benefit.

Quote:
1. Under Einstein preferences what does "Run CPU versions of applications for which GPU versions are available" refer to?


The current GPU apps originally started out as CPU only before a GPU version was developed. If I remember correctly, there was a transition period where both a CPU version and a GPU version existed. This may even happen in the future with the GW app. That setting simply allows you more control over which devices run which types of tasks. There is no current CPU app for BRP6 so the setting is not relevant to you.

Quote:
2. How do I get into the CUDA beta testing program?


By clicking the link to Einstein@Home preferences on your account page and then looking for the setting that says, "Run test applications?" and ticking the box.

Quote:
3. How is it that I have 8 CPU tasks and 4 GPU tasks running .2 CPUs when I only have 8 cores?


Because 0.2 is an estimate of what might be required in CPU cycles in order to support the GPU. With fully loaded CPU cores (8 CPU tasks running) a CPU task has to stop crunching and its CPU has to be 'released' in order to service the GPU request. The latency of this will slow down the GPU task and increase the elapsed time of a CPU task without necessarily showing as an increase in the CPU time of the CPU task. I did try to cover this previously.

Quote:
4. How to I insert a screenshot in a message?


Find yourself a site that will host images for you. There are lots of free ones around but you should look at the terms and conditions to make sure you're happy with them. Upload your image to that site - you will be given a link to your image when you upload it. Insert that link, surrounded by [ img ] and [ /img ] tags into a new message. Please make sure you also write something about the contents of the image and you do explain what it's supposed to show. Also, please don't post 'big' images that will 'overflow' the width of a screen. It's quite annoying to have a thread 'distorted' in this way.

When you are composing messages, you will see a link, "Use BBCode tags to format your text" just to the top left of all the shortcut buttons you see above the message composition window. Click that link and experiment with the examples shown until you figure it all out :-). You don't get to 'PASS GO' and collect your $200 until you master this stuff :-).

Quote:
5. Would you mind taking a look at my settings?


Of course not. I've only commented on the ones you should think about. Others may give you different advice, in which case you'll need to make your own decision. With these, there is no 'correct' or 'incorrect' answer, necessarily.

Quote:

Einstein Preferences:
...

GPU utilization factor of FGRP apps .5


There is no GPU app for any current Fermi Gamma Ray Pulsar search so you should leave this at 1 in case such an app appears suddenly in the future (probably rather unlikely at the moment). As a precaution, if any new GPU app suddenly appeared, it would be prudent to start out running 1x until you see what happens.

In addition to the ones you mentioned, you should note that you do have the ability to select on CPU apps as well as GPU apps. The searches for gravity waves and for gamma ray pulsars in binary systems are quite different and you may choose to support both types, which is the default (I think). However you can select just one if you wish.

Quote:


hcc1
4

.500
1.250



Until you really understand what you are doing, you don't need an app_config.xml file. It has a rather specialised purpose and is not normally required. If such a special purpose should arise, the file has to be tailored for that purpose. Any such file is project specific and is inserted in the particular project directory. The above example is NOT for Einstein because the app name given (hcc1) is NOT an Einstein app.

The app_config.xml file is documented here.

Quote:

....
....


cc_config stands for Core Client Configuration. The core client is the program boinc.exe. There is a GUI front end called BOINC Manager (boincmgr.exe). The manager just provides a nice interface which allows you to communicate your wishes directly to the client, if you want to. The manager defaults to a 'simple' view which I never use. I always set the 'advanced' view because that allows me to see the details I want to see. If you want to tweak things and improve performance, you really need to become familiar with the advanced view. For 'production', the client is designed not to need much interaction and a lot of people will just set and forget.

When you launch BOINC, it uses a default, built-in set of configuration options which are quite appropriate for most people and pretty much always appropriate for people starting out. To change any of these defaults you need to create a brand new configuration file and call it precisely cc_config.xml and place it where BOINC will find it. It only needs to contain options you wish to change. It's used to override the in-built defaults. Before deciding to override any default you should RTFM. Once again, most people will happily use BOINC without really needing to override defaults.

I hope the above answers a few more questions. You need to hasten slowly. You're likely to get in a real mess by trying to change too many things all at once and before you really understand the basics. The ONLY things you should be worrying about right now are:-

  • * Run GPU tasks 2x on the 970 - already accomplished.
    * Use the CUDA55 app - already accomplished (even if you don't see it yet).
    * Make sure you have 'free' CPU cores to support your GPUs - set "Use at most 75% of the CPUs" - not correct yet.
    * Make sure you don't cripple your CPU time - set "Use at most 100% of CPU time" - not correct yet.
    *

While you are tweaking set your work cache to 0.5 days - unfortunately this is probably just shutting the stable door after the horse has bolted - but do it just in case ....
* Worry about fixing the memory clock later - when you have taken the time to work out how to properly use NVIDIA Inspector. Keep reading the thread about this.

Cheers,
Gary.

cliff
cliff
Joined: 15 Feb 12
Posts: 176
Credit: 283452444
RAC: 0

Hi Phil, RE: I clicked

Hi Phil,

Quote:


I clicked apply but it defaults back to 3004 MHz. What's up with that?

Regards,
Phil


Arrgh:-) DO NOT under ANY CIRCUMSTANCES hit APPLY!!!!!!!!!!
What you need to do is setup again [remember to click on unlock MAX, top rhs]
The when the p2 mem is set to 3505, use the 'make link' button [lower left].
Once you have the link on your desktop, simply close NVInspector using the top right x for close.
Then before you fire up BOINC, click on the link to NVI, then run BOINC.

When crunching highlight a working task, then click properties [advanced view] and near the bottom of the properties box there should be a line showing throughput in %per min/hour

You can check throughput with and without using the NVI link to see what diff in throughput you get.

Currently running single instance of a Parkes BRP6 cuda 55 task I'm seeing 2.316 %per min-2.340 %per min..= aprox 42Min.44sec to 44min.40sec per task.

You're running double tasks, so how you determine accrual throughput is up to you, I don't think just adding % for each running task will provide the answer, manually looking at the total time to complete both task's should be most accurate.

YMMV:-) Also I'm running a higher spec card..so there should be a diff in %pmin

Regards,

Cliff,

Been there, Done that, Still no damm T Shirt.

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

Cliff: I did exactly as

Cliff:

I did exactly as you asked.

Quote:
What you need to do is setup again [remember to click on unlock MAX, top rhs]
The when the p2 mem is set to 3505, use the 'make link' button [lower left].
Once you have the link on your desktop, simply close NVInspector using the top right x for close.
Then before you fire up BOINC, click on the link to NVI, then run BOINC.

I created the link, rebooted without Boinc starting, clicked the link but the P2 reverts to 3005.

Thanks,
Phil

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

Gary I really appreciate the

Gary I really appreciate the time you took to write clear, detailed answers to my questions. I made and corrected the changes you suggested. I'm with you that I just need to sit back and let things percolate a bit. Understanding comes with patience.

In addition, I want to thank everyone else in this thread who has been so generous with their time. Thank you.

Phil

cliff
cliff
Joined: 15 Feb 12
Posts: 176
Credit: 283452444
RAC: 0

Hi Phil, RE: Cliff: I

Hi Phil,

Quote:

Cliff:

I did exactly as you asked.

Quote:
What you need to do is setup again [remember to click on unlock MAX, top rhs]
The when the p2 mem is set to 3505, use the 'make link' button [lower left].
Once you have the link on your desktop, simply close NVInspector using the top right x for close.
Then before you fire up BOINC, click on the link to NVI, then run BOINC.

I created the link, rebooted without Boinc starting, clicked the link but the P2 reverts to 3005.

Thanks,
Phil

I dunno whats up Phil, it works for me and others.. Sorry no other ideas right now:-(
Regards,

Cliff,

Been there, Done that, Still no damm T Shirt.

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7024344931
RAC: 1807376

Florida Rancher, your machine

Florida Rancher, your machine has now actually downloaded some work designated as CUDA55. This is the much-discussed "Beta" or "test" work which we have advised you will run to completion in significantly shorter elapsed time than the previous CUDA32 flavor you have been running. The support CPU task will also consume less resource, which will shorten the elapsed times of the CPU jobs you are running at the same time.

By taking a look at the in progress task list for BRP6 work for your host, you can see the transition.

No new work was downloaded for over 90 hours as the system adjusted to your reduced queue length request. When new work again was fetched, first at 23:53 UTC on May 21, the application column includes the substring cuda55, instead of cuda32 for previous tasks.

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

RE: Cliff: I did exactly

Quote:

Cliff:

I did exactly as you asked.

Quote:
What you need to do is setup again [remember to click on unlock MAX, top rhs]
The when the p2 mem is set to 3505, use the 'make link' button [lower left].
Once you have the link on your desktop, simply close NVInspector using the top right x for close.
Then before you fire up BOINC, click on the link to NVI, then run BOINC.

I created the link, rebooted without Boinc starting, clicked the link but the P2 reverts to 3005.

Thanks,
Phil

I just crunch 24/7 (or sleep the system when not crunching and not using it), and when I do reboot once a week or so, I re-do the nVidia Inspector settings.

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

Cliff: I don't know what's

Cliff:

I don't know what's up with that either but searching around there are many others having the same problem but offering no solutions.

My card is a Gigabyte GeForce GTX 970 OC Mini-ITX.

ExtraTerrestrial addresses this in his "Low Memory" thread but I don't understand his second point:

Quote:

Overclocking the memory:

Thanks to skgiven from GPU-Grid I can now overclock my memory! Here's how:

- the GPU must not be crunching BOINC (either pause your GPU project, or all GPUs, or suspend BOINC completely)
- in the nVidia Inspector OC tab set the overclock for P0 (because you can't go any higher than this in P2)
- now you can set up to this memory clock for P2 as well
- apply & have fun

GTX970/980 are reported to reach about 4 GHz memory clock. It should be interesting to see if this can boost Einstein performance another 10%.

Regards,
Phil

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

Archae: That's good to

Archae:

That's good to hear.

Checked my completed CPU WUs I've noticed it now takes 12.5 hours per core where it once took 10.5 hours per core. What can I attribute to this slowdown?

Regards,
Phil

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.