I do seti about once a year for the WOW event, and this year is no different, but this year I have noticed a difference. The GPU times (run times) seem to be all over the board with many being in the 5-8,000 second range, and my RAC seems very low. I have searched for and found a couple of threads with info but so far I am not quite happy with the outcome.

I have installed Lunatics v0.44 and I have put in an app_config.xml as follows:

<app_config>

<app>
<name>astropulse_v7</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

<app>
<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>0.3</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

</app_config>

When I look at other similar rigs, in the top hosts list, their RAC is approximately double what I am getting. (I7 intel cpus with two GTX970 gpus for example)

SIV64x and GPUID, show 99-100 % gpu utilization, gpu temps of 70C or less, cpu temps around 65C; in other words nothing jumps out at me.

If anyone might have insight on the v8 work, and how I might get more out of my rigs, I would appreciate the help.

---

http://lunatics.kwsn.info/index.php?action=downloads;cat=9

Read the reame-file on MB ATI opencl.txt

There is a description in settings to make on some txt files, to get more performance auto of ur (NV 660).

It written with:

-sbs 192 -spike_fft_thresh 2048 -tune 1 64 1 4 u have to enter in the TXT-file in the C:\Program Data\Boinc\projects\"seti" folder.

BEWARE there is a sightly higher power comsumption on your system.

Without testing, i would use the following settings on the 970 rigs:

High end cards settings

---

For the nVidia GPUs there are two major applications in use - the SoG, which does some tasks very well, but at the cost of high CPU usage, and the older CUDA tasks which are more consistent in their timings, not as fast, but use less CPU effort. I think most folks are heading the to SoG applications.

Again, for nVidia, running SoG you must make sure you have enough "spare" cpu cores for each GPU task to have one core each otherwise the performance can be seriously affected.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Until recently I was running a very similar system to your example system, with two GTX970 installed - I found that this system turned more work around running two tasks compared to the three you are listing in you example.

(This PC now has two GTX970 and a GTX1080 - only just installed and the RAC is still climbing quite rapidly...)

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

I would suggest downloading and running the beta 0.45 rev3 Lunatics installer. That allows one to select only the OpenCL MB SoG app. Since SoG arrived, I have stopped running tasks with CUDA since with optimized command lines and text files, I found that running 2 concurrent OpenCL tasks per each 970 gives better throughput and credit than running my previous 3 concurrent CUDA tasks per each 970.

You will have to reduce the number of running CPU tasks to allow one full core to service each GPU task.

Look into the ReadMe_MultiBeam_OpenCL.txt file for the parameters you need to put into the mb_cmdline_win_x86_SSE3_OpenCL_NV.txt file. Use the high-end card values as someone suggested.

I also suggest downloading and using Mr. Kevvy's rescheduler app and Stubbles69's command file to run it at least once a day, preferably twice a day for a 24/7 machine. That should set you up well for the contest.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

On machines with two GPUs (a faster 970 and a slower 750Ti), I'm experimenting with running the guppis with SoG on the 970 (just singles at the moment), and the Arecibos with cuda doubled-up on the 750. Once I get them started with the right app on the right card, they run stable for hours - just needs a big search'n'replace twice a day to switch the plan class over.

Ok guys. Looks like I have some work to do.

Thanks

---

Ok guys. Looks like I have some work to do.
Thanks

Hey Pokey,
I just read the whole thread above.
There is much info there for someone to absorb who only crunches once a year during SETI WoW.

You initially enquired about the GPU task times being all over the place.
Since early spring 2016, S@h is now processing data from a second telescope (Green Bank) other than Arecibo. The current GPU apps were made for Arecibo data.
The Green Bank tasks all have: blc...guppi...vlar in the filenames, and are generally referred to in the forum as: guppi

There are roughly 4 types of tasks that have different run times on nVidia high-end and mid-range GPUs.
1. shorties (mostly Arecibo but also some guppi) are usually less than a minute in duration.
2. Arecibo VLARs: these account for less than 10% of tasks and have the worst times on the CPU and nVidia GPUs.
These usually give the perception of having the worst credit/time because of #3&4 below.
3. guppi: somewhere between 35-50% of tasks received in a day (by my experience during the last few months). These will negatively impact the throughput of your GPU when running multiple tasks in parallel with any of the Cuda## apps from Lunatics v0.44.
4. the other Arecibo tasks (that don't have VLAR in their task/filename): these I refer to as nonVLARs (but the gurus use other labels that is more confusing for people who aren't up-to-date on the S@h forum techno-monbo-jumbo).
These currently account for 50%+ of daily tasks received (by my experience).

To optimize overall throughput of rigs with nVidia GPU(s), the ideal scenario is to process all Guppis on CPU cores and all nonVLARs on GPU, and that is why there is Mr. Kevvy's guppiRescheduler.exe available (and I made a Windows7to10 front-end script to automate the other steps before and after his file runs).

With a GTX 750 Ti and only nonVLARS being processed by Cuda50, the best throughput is 4 nonVLARs in parallel ( which gives me on average ~48mins for 4 tasks = ~12mins/nonVLARs ).
With my new GTX 1060, it seems to be about 6-8mins/nonVLAR with the same parameters in app_config.xml:

<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>0.25</gpu_usage>
<cpu_usage>0.25</cpu_usage>
</gpu_versions>

...but if you don't use a taskDeviceSwapper such as Mr Kevvy's, don't use these settings!!!

As far as the Cuda apps, the biggest advantage is that you can run 4 tasks in parallel with just 1 CPU core support.
Therefore your other CPU core are able to process the guppis. Obviously, for rigs with a powerful GPU and a Core2dua the above strategy is not ideal since guppis would eventually have to be processed by the GPU once the CPU queue has reached its max of 100 guppis.

This should keep you busy for a while! ;-)
Looking forward to your next Qs/comments.
Cheers,
Rob :-D

---

Well, my eyes are now completely glazed over, thank you very much. I can't leave you guys alone for a minute before you start changing stuff. LOL

I will ruminate on this info and make some choices in the next little bit. I am sure I will have questions.

One question would be: Is it recommended setting "no new tasks" and finishing up the work I have before applying lunatics v0.45 beta3. That was the approach I took when I went from stock to v0.44.

And if I don't, What happens to the CUDA units if I'm going to Open CL.

---

Well, my eyes are now completely glazed over, thank you very much. I can't leave you guys alone for a minute before you start changing stuff. LOL
I will ruminate on this info and make some choices in the next little bit. I am sure I will have questions.

hahaha sorry for that! :-p
Fyi, as you can probably tell, there are currently 2 thought-camps: pro-NV_SoG optimizers and pro-Cuda50 optimizers.
The only one that wins hand down is: NV_SoG as stock when running (1 task/gpu). That's the set&forget that most people do ("attached Boinc Client to S@h project and forget it's even active" ...until something slows down your pc!)

3 months ago, the pro-NV_SoG camp was making great progress because guppi to nonVLAR ratios were slowly increasing...but then the whole task-to-device reassignment (most notably Mr Kevvy's guppiRescheduler) changed everything.
Some in the NV_SoG camp even think that task-to-device reassignment is "nefarious"! (see my sig for a taste of bitter)

...and now both camps are replying to your initial post!
Do you feel like you just got teleported into the middle of a battlefield?!? lol

One question would be: Is it recommended setting "no new tasks" and finishing up the work I have before applying lunatics v0.45 beta3. That was the approach I took when I went from stock to v0.44.
And if I don't, What happens to the CUDA units if I'm going to Open CL.

With v0.45 beta3 you no longer need to empty your cache. The only minor_bug is that pre-existing tasks assigned to GPU will have the old label in Boinc Manager but the Boinc Client will use the new GPU app that you selected in v0.45 beta3.

Happy WoW! ;-}
RobG

---

The running the installer version 0.45 does not require you to run your cache down - I've just run it on one of my crunchers with no problems.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

I am probably looking right at it but I can't find the "ReadMe_AstroPulse_OpenCL_NV.txt" file. I found one for intel and ati but no Readme for nv.

---

I found one for intel and ati but no Readme for nv.

I the ati-file is the nvidia description...

---

OK Thanks.

---

I have that file. Must have been in one of the previous Lunatics installers. Sorry to lead you astray. Same information is in all the different platform readmes.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I have decided to go the NV_SoG path. I have installed lunatics v45 beta3 and installed the high end card settings in the mb_cmdline_win_x86_SSE3_OpenCL_NV.txt file

And I have modified my app_config file to do two work units:

<app_config>

<app>
<name>astropulse_v7</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

<app>
<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

</app_config>

Nothing blew up or caught fire yet so all is good.

Mr. Kevvy's rescheduler app and Stubbles69's command file will follow.

If you don’t hear back from me it is because everything is working.

Thanks for the input everybody, you have been a big help. And Keith Myers it looks like we are in the same Pisces group…………….. again this year. At least birthdays don’t change.

---

Yes, let us know how you fare with Mr. Kevvy's Rescheduler and Stubbles69 front end script. I have had good luck up until just now. I had the script fail on this machine, my daily driver. Other machines ran it fine.

The error in the script was that it could not determine the cpu version number.

Reading configuration files...
Error: could not determine CPU version_num from client_state. Nothing changed.

First time that has happened. I wonder what might have changed. Maybe something in the way the project interacts with client_state.xml?

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Hey Rob, question for you. My New machine. with 2 1060's in it that I am just trying to put the final touches on and add it to the farm, needs to have the optimized app on it. Which one, I'm not sure, I believe SoG is the most productive, I can install that, feel free to take a look at the specs of the machine, to see if you see any issues with it running that app. I will be overclocking the CPU a bit, Pretty sure it will het 4.1-4.2, as it's on a Corsair H115i water cooler, so I might even be able to hit a little higher if I'm lucky in the silicone lottery.

I've had a headache for the last few days (slowly getting better, I think it was from PCV glue fumes. Long story.), so I dread having to go thru all the reading to get this thing configured and optimized to produce the most science it can. Can you give me some suggestions as to step by step loading the proper software, adding/modifying any .xml or config files, so I can get it to start producing? Normally I would maybe ask to be pointed in the right direction, but I'm just not up to it yet, but I would really like to get it working properly, so any help would be appreciated. Thanks!

---

I've had a headache for the last few days (slowly getting better, I think it was from PCV glue fumes. Long story.)

Reminds me of painting a concrete floor at a previous work place, indoors with no ventilation.
Huge headaches & lightheadedness. At the end of it all I took a look at the safety directions on the paint cans- not to be used in enclosed spaces without significant forced ventilation...

---

Grant
Darwin NT

Yeah, and the sad thing was, I was working outside, though at almost chest height so it was probably closer than it should have been, but it was gluing up 400' ( 4 lengths, 100' each) of PVC pipe that I am laying between the house and the shed/brewery I am building. 2 3", 1 2" and 1 1.5". I'll have room for anything I can think of, even a loooong tapline, though keeping that cool would be a monumental task. I think I'll just use my keggerator. lol

So yeah, lesson learned, keep that nasty stuff as far away and down wind from you as possible! And take breaks, especially when it's 90+ and humid.

---