Hello fellow number crunchers. Been away from Seti@Home for some time now but have recently been trying to get my crunching going again. Please forgive the newb questions, I have been reading through the forums as much as possible to try and find the answers, with no real luck. Probably just not searching the right things.

Recently upgraded to a EVGA 2070 Super and added it into my rig with my former EVGA 660ti.

updated the cc_config.xml file to use both GPU's and have also been playing with the app_config.xml file to maximize their usage.

Couple of things:

1. My app_config.xml file reads as such:
<app_config>
<app>
<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
<app>
<name>astropulse_v7</name>
<gpu_versions>
<gpu_usage>.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>

However, the event logs is telling me that 'astropulse_v7' is an invalid app. What is the correct app to use?

2. My 660ti and the 2070 Super both crunch WU's are roughly the same speed. ~30 mins for the 2070 and ~30-40 mins for the 660ti. I HAVE to think that something is not configured correctly here. I cannot, in my wildest dreams, imagine that my 2070 super should not just smoke the 660ti. What sort of numbers should I be able to obtain with these two cards? Dev 0 is the 2070, Dev 1 is the 660ti in the example image.

With using the old CUDA50 app both cards will perform about the same as you are seeing. You need to dump the CUDA50 app as it was made for cards of the 660Ti vintage and is entirely inappropriate for a 2070 Super. You need to be running the SoG application to fully use the 2070 Super. Unfortunately, if you are using the stock client and applications provided by the project, it may be many months before the project figures out you can run the SoG application to best advantage. I would recommend you install the Lunatics applications and in the installers configuration pages, select the SoG application in the Nvidia card configuration page. The default will be the CUDA50 app which you don't want, you need to select the SoG application. You can download the Lunatics installer from Mike's World

http://mikesworld.eu/download.html

---

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

A proud member of the OFA (Old Farts Association)

Thank you Keith. I very much appreciate the guidance. I noticed that I only have 9 cuda 50 and 6 cuda42 tasks remaining in my cache. Everything I seem to have gotten since then is all Opencl. I am probably just going to let them run their course rather than abort them and then reap the fruits of those tasks under the new application.

I am looking at a way to get that 660ti into another machine. It runs through the OpenCL tasks decently but the 2070 is still quicker than it (although not as much as I would think). The bigger thing about the 2070 is that I can ratchet it up to run more than one task without much notice of degradation in processing time. The 660ti can't. So, with the inability to tell the 2070 to run 2, 3 tasks and the 660 only one that I can find, I am looking for a way to put the 660ti in another machine and let it just crunch away.

Normally it takes 11 validated tasks of each type to develop the APR for each app. Once all the apps, CUDA50, CUDA42 and OpenCL SoG apps have an APR, the SoG app will have the highest APR and the highest performance and the scheduler will send you only SoG tasks. There may be a few stragglers here and there but it will stabilize on the SoG.

Yes, there really isn't any good way to run singles on the 660Ti and doubles or triples on the RTX 2070 Super. But you may find that running singles on the 2070 is just as fast as running doubles and triples once you add up the total times for multiple tasks and get the average. It is normally a wash in compute times for the high end cards and I would state the majority just run singles. This was not the case with the older hardware and the old CUDA50 app. But for SoG, it is in fact. Also SoG really demands an entire cpu thread to support each SoG task.

---

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

A proud member of the OFA (Old Farts Association)

Hey Keith, thanks for the recommendations on the Lunatics installer. I had read about that in several places but was actually having a difficult time in finding a good place to download it from.

In regards to installing that, it seemed pretty straight-forward. Let most of my tasks run through and then installed it this morning. I dl'd the main installer.exe and all the other applications as well into a single folder. I then ran the .exe and chose the options for my Nvidia card to run SOG and CPU tasks.

Question I have, the applications - do I need to unzip them also into the project folder or since I had them in the download folder to install from does it automatically put them in the right places?

Sorry for the newb questions here, just trying to make sure I can get myself running as optimally as possible.

I see I am getting advice from someone who knows, considering you have 4 machines in the top 20. :)

Side question, do you know what the [64] means in the GPU field? Cannot seem to make an educated guess based on comparisons between various machines. ( [64] NVIDIA GeForce RTX 2080 (4095MB) driver: 430.40 OpenCL: 1.2 )

No you can just put the archive away in a safe place in case you need to rerun that app to change configuration. The installer put everything that is needed in the correct places. You are processing correctly now with the anonymous platform and using the optimized apps.

You still have some optimizing to do with the 2070 Super card which is running at the stock settings the Lunatics Installer provides. That is just the minimal, general case for all types of cards. You can more aggressively tune your 2070 Super to process tasks faster with command line entry.

Copy this code:

-sbs 1024 -period_iterations_num 1 -tt 1500 -high_perf -high_prec_timer -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64

into the mb_cmdline-8.22_windows_intel__opencl_nvidia_SoG.txt file
or the file may be named
mb_cmdline_win_x86_SSE3_OpenCL_NV_SoG.txt located in the setiathome.berkeley.edu directory.


Just open the file with Notepad and copy and paste the tunings into it and then save it. The very next task that starts will pick up these more aggressive tunings and process the task faster.

My computers are running what we call a "spoofed client" They are reporting more gpus to the schedulers than physically exist in order to get a bigger cache.

---

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

A proud member of the OFA (Old Farts Association)

Done. Thank you again for all your assistance. That 2070 is going at about 6x the 660ti right now. More as I would have expected and with this latest tune, should hopefully go even faster but will watch it over the next few days.

Been toying with new build for awhile, think it might be time to pull the trigger. The i7-870 is getting a little long in the tooth.

Was looking at Intel i7-9700k but have recently been researching and more and more looking at the Ryzen 9 3900x with 32 GB of Quad channel RAM, mobo, M.2's etc, and dropping the 2070 Super in there and seeing how that rig will fair. :)

Prolly close to 20 years since I have had AMD in a machine. They have come a long way and finally look to have outdone Intel in both desktop and server chips.

Haven't seen the new tunings show up yet in your latest pendings. Did you save the command line file in Notepad? You want to always use Notepad to edit BOINC files. Not Word. You need to make sure you choose save type as All. It should have preserved the filename as is. If your latest tasks don't show up the new tunings shortly, something must have gone wrong and I will have to see what the filename is in your app_info file_ref statement.

[Edit] Now seeing the tasks with the new tunings. Looks good, shaved a minute off crunch times it looks.

---

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

A proud member of the OFA (Old Farts Association)

The 3900X is a dual channel platform cpu on AM4 socket. So you will not be able to run quad channel. You can populate all memory channels and 4 X 8GB sticks is a popular choice.

If you desire a quad channel AMD cpu you will have to move to the Threadripper X399 platform and either the previous generation cpus or wait till the end of the year when we expect the Threadripper 3000 series cpus to come out with Zen 2 technology.

---

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

A proud member of the OFA (Old Farts Association)

Gotya. Sorry for the misunderstanding. Prob going to wait a little bit before making that build, but looking hard at AMD. Threadripper may not be out of the question, but have to look at cost vs. return...

Yeah, the 2070 is now crunching anywhere from ~ 1.25 mins to ~5 mins a WU. The 660 at ~30. :) Probably should pull that out of the machine but hate to lose any of that processing really. Every little bit helps and it would just be sitting in a box if it is not in this machine right now.

I have noticed two things since installing Lunatics:

1. The estimated times are WAY off - I expect that in some time these will settle out and come more in line. e.g. says 2h20m for tasks on my GPU's when they take 1-30 mins. The CPU processes say over a day and are running well ahead of that.

2. The client seems to not want to auto-communicate with Seti and transfer processed WU's and ask for more work. They are just sitting the queue saying Ready to Report. Also, in the project list, it used to have a timer of about 5 mins between comm attempts. That is no longer there?

Since you are running new apps, they haven't established any APR yet for the client and scheduler to give accurate estimates of time remaining. As soon as you have returned and validated 11 tasks for each application, the APR won't be established. Also the 11 can't be overflowed tasks or ones with 100% radar blanking.

Essentially you are starting with a brand new host. The only errored tasks are the stock application SoG ones that were abandoned by you manually or the Lunatics installer was not able to preserve them by changing their type to the new application for some reason. All tasks normally are preserved in a Lunatics installation. I don't think those should be held against you and prevent you from asking for work. Could be wrong.

Check the Manager and make sure that network communications is enabled. Do a manual Update to communicate with the project to report tasks and ask for replacements. Set sched_op_debug in the Event Log options permanently. It will show you how many seconds of cpu and gpu work you are requesting at every scheduler connect. Set work_fetch_debug for one cycle to see what the client is asking for work and what commitments it shows for all work and for all projects. You should be communicating with the project every 305 seconds normally. Unless you have disabled comms or have recently errored out a lot of work.

---

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

A proud member of the OFA (Old Farts Association)

Will do. Not sure what happened in the installer. Didn't abort any tasks before installing Lunatics. There were only 2 in process before installing. After coming back up, there were only 8 tasks in the task list when there were 3 days worth prior (not sure how many tasks exactly).

I'll let 'er run and check those settings.

Yes, Richard designed the Lunatics installer to preserve all work during the conversion. All it was supposed to do was change the gpu tasks from SETI@home v8 v8.22 (opencl_nvidia_SoG)windows_intelx86 type to SETI@home v8 Anonymous platform (NVIDIA GPU) type in the client_state file. Normally it works perfectly as designed. It seems it did all of the cpu tasks correctly but only 74 of the gpu tasks correctly. 26 gpu tasks didn't get converted correctly for some reason. It appears you "ghosted" most of your cache as your tasks in progress now shows over 540 when you should only have 200 tasks. Looks like you need to run my Lost "Ghost" task recovery protocol to get those lost tasks back into your cache. You can now recover 80 tasks at a time so only about 5 cycles of the recovery protocol should get them all back. Set NNT on the host and let your cache fall down so that you have room for 80 tasks to be recovered and then run the protocol. Let the cache fall back down again and repeat. The procedure is here:
https://setiathome.berkeley.edu/forum_thread.php?id=84176&postid=2003388

Also the reason the client isn't asking for work since you have exceeded the normal amount of cache for your host of 200 tasks. Run the ghost recovery protocol and get the task in progress showing 200 tasks and you should be back to normal.

---

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

A proud member of the OFA (Old Farts Association)

Been slowly recovering some of the tasks. Thanks for that procedure!

2 things I have noticed:
- The 2070 does not seem to show any progress on a task in the first 30-45 secs and then starts to crank away and finishes the task in ~5 mins. Is this a limitation of the CPU not able to feed the GPU fast enough? I am running a i7-870 so yes, I know that is pretty long in the tooth these days and might be more of the kicker I need to upgrade machine...

- I am still a little surprised to see the performance of the 2070. I was looking at some other machines running a plane jane 2070 card (The top machine in the states is 8589036 - Ian&Steve C.) and that is finishing GPU tasks in 4x-5x faster than me. Granted, much newer/faster CPU's (see point 1) and probably some even more optimized apps than I am running.

No, you have basically equal run_times and cpu_times for your tasks. So you cpu isn't overcommitted. On the initial face of it.

BUT, you do how some worrying outliers in your tasks on every page. Some tasks are taking over 1000 seconds to crunch while the majority only take less than 300 seconds. So for those tasks, I would say that the cpu is busy doing other things like desktop maintenance or another cpu intensive background task. I would say that it is likely your old cpu is having to struggle to keep all your gpu tasks fed.

Yes, you can't compare your 2070 to Ian's because he is running Linux and the special app which is 5X times faster than the app you are running.

---

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

A proud member of the OFA (Old Farts Association)

Hey Keith, I think that those "outliers" might be the tasks being run on the 660ti. They definitely take about 1000-1200 secs to run on that card compared to about 300 secs for the 2070.

What still have me slightly perplexed and just does not feel to be normal to me is the fact that even when doing nothing else on this machine, I can still watch tasks on either card that take anywhere from 30-50 secs before they start making any progress on them. I watched one today that did not move to 0.001% progress until 52 secs in. Once progress does start, they crunch though them pretty quick.

Any thoughts on that at all? for a 5 mins task on the 2070, 20% of the time spent looks to be doing nothing at all.

---

https://boincstats.com/signature/-1/user/2360850/sig.png

Hey Keith, I think that those "outliers" might be the tasks being run on the 660ti. They definitely take about 1000-1200 secs to run on that card compared to about 300 secs for the 2070.

What still have me slightly perplexed and just does not feel to be normal to me is the fact that even when doing nothing else on this machine, I can still watch tasks on either card that take anywhere from 30-50 secs before they start making any progress on them. I watched one today that did not move to 0.001% progress until 52 secs in. Once progress does start, they crunch though them pretty quick.

Any thoughts on that at all? for a 5 mins task on the 2070, 20% of the time spent looks to be doing nothing at all.

I have noticed that too on some tasks. Depends on the task name. Some species, don't move the progress bar for a half a minute or so and then rapidly increases to completion. Has to do with how the task is loaded into the FFT array for binning. Part of Petri's optimizations. I think he pre-parses the task to see how to set it up for crunching. Nothing to worry about as long as the task takes its normal amount of time to crunch.

Also DA put in a "quasi" progress mechanism in the client code to pacify the complaints from other projects that don't ever see ANY progress on the progress bar until it hits 100%

Again nothing to worry about.

---

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

A proud member of the OFA (Old Farts Association)

Thanks! Have a great day/night ahead!

---

https://boincstats.com/signature/-1/user/2360850/sig.png

Hello Keith/all,

Built out the new Ryzen machine. Installed Boinc and then the Lunatics installer.

Completed the following as previously done on old machine:

Copy this code:

-sbs 1024 -period_iterations_num 1 -tt 1500 -high_perf -high_prec_timer -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64

into the mb_cmdline-8.22_windows_intel__opencl_nvidia_SoG.txt file
or the file may be named
mb_cmdline_win_x86_SSE3_OpenCL_NV_SoG.txt located in the setiathome.berkeley.edu directory.

Created app_config.xml file with the following and have my pref set on that Comp location to 50% CPU's.

<app_config>
<app_version>
<app_name>setiathome_v8</app_name>
<plan_class>opencl_nvidia_SoG</plan_class>
<avg_ncpus>1</avg_ncpus>
<ngpus>1</ngpus>
</app_version>

<app_version>
<app_name>astropulse_v7</app_name>
<plan_class>opencl_nvidia_100</plan_class>
<avg_ncpus>1</avg_ncpus>
<ngpus>1</ngpus>
</app_version>
</app_config>

However, no matter what I do, I cannot get the 2070 to use more than 0.04% of the CPU for it's tasks. Am I missing something here? It does not seem to be affecting the run times on the tasks as they are completing between 3-4 mins, but something just does not seem right here. Another thing that I have noticed on this machine is that it no longer takes 30-45 secs for a GPU task to show ANY progress. So, that seems to be a good thing.

"Running (0.04 CPUs + 1 NVIDIA GPU)"

---

https://boincstats.com/signature/-1/user/2360850/sig.png