BOINC Pentathlon 2019

[leadeater]

3 minutes ago, Windows7ge said:

Lot of good 192 threads does you when the same number of jobs can be done with 1/4 the threads working 4x as fast. My 32 thread server would be on par with that behemoth if each job of yours took 8 hours.

Yea it's just due to me trying to run dual instance on the single OS, I'd need to do per core allocation for each project and task of it but that's basically impossible. The job times not doing that for Zika were around that hour and half, I'm actually amazed just how bad it is when you clash cores with multiple tasks.

 

Edit:

Also the reason I was trying is 192 is just too much for boinc, single project only uses like 70%.

[Windows7ge]

10 minutes ago, leadeater said:

Yea it's just due to me trying to run dual instance on the single OS, I'd need to do per core allocation for each project and task of it but that's basically impossible. The job times not doing that for Zika were around that hour and half, I'm actually amazed just how bad it is when you clash cores with multiple tasks.

 

Edit:

Also the reason I was trying is 192 is just too much for boinc, single project only uses like 70%.

Get on that VM train then that crazy machines gotta have Hyper-V. Or is Hyper-V on Windows Server only capable of handing out 50% of the cores?

 

I got the Emerald Project Badge. Next stop, Sapphire.

 

Oh, cool we're tied for 13th place with BOINC.Italy

[leadeater]

2 minutes ago, Windows7ge said:

Get on that VM train then that crazy machines gotta have Hyper-V. Or is Hyper-V on Windows Server only capable of handing out 50% of the cores?

Think I'm just going to put Linux on it, kind of not much point now but I think it's pretty much required for this task and that many sockets. VMs would work well though, just make a bunch of 24 or 48 core VMs and pin cores to each VM just in case.

 

Not used a quad socket system for this sort of thing before, typically just database servers.

[Windows7ge]

5 minutes ago, leadeater said:

Think I'm just going to put Linux on it, kind of not much point now but I think it's pretty much required for this task and that many sockets. VMs would work well though, just make a bunch of 24 or 48 core VMs and pin cores to each VM just in case.

 

Not used a quad socket system for this sort of thing before, typically just database servers.

Like you were telling me pin the VM cores to NUMA nodes so work doesn't have to cross physical processors to access memory (at least I think that's the issue).

 

From what I've discovered so far you can do a minimal install, it only needs about ~16GB of disk space, 4GB of RAM should suffice, and as many cores as you want to throw at each.

 

I went with Ubuntu 18.04 LTS just out of convenience but it has a lot of extra stuff running in the background that doesn't need to be there so I'd try to find something more light weight/barebone. I'm going to give Mint Mate 18.3 a shot like Ithanul has good luck with.

[Windows7ge]

You know what. A while back I was experimenting with finding the lightest weight Debian Linux distro I could to perform a particular task (the plan being to run about 10 instances of it in a VM server). I wouldn't mind going around again testing a series of random distros to see which has the least overhead when crunching BOINC tasks (best turn around time). It may prove useful to those who many want to virtualize the BOINC manager. This was part of my idea as an alternative form of advanced bunkering.

[porina]

1 hour ago, leadeater said:

Could be worse, just checked the last few Zika jobs I let run to the end. Everything else is yoyo, these were just close enough to the end to not suspend/abort them.

Be careful with suspending tasks. My understanding is if any task is user suspended, then boinc will not get any new work even for other projects.

[jiyeon]

Ubuntu keeps locking up my PC so I've decided to do only 5 cores out of 6 at 80% computing power.

 

 

I noticed now that the time it takes to complete some projects has shot right up, is this as a result of getting deeper into the BOINC economy or could it be due to my constant rebooting and suspending due to my PC locking up?

[Ithanul]

53 minutes ago, Windows7ge said:

Like you were telling me pin the VM cores to NUMA nodes so work doesn't have to cross physical processors to access memory (at least I think that's the issue).

 

From what I've discovered so far you can do a minimal install, it only needs about ~16GB of disk space, 4GB of RAM should suffice, and as many cores as you want to throw at each.

 

I went with Ubuntu 18.04 LTS just out of convenience but it has a lot of extra stuff running in the background that doesn't need to be there so I'd try to find something more light weight/barebone. I'm going to give Mint Mate 18.3 a shot like Ithanul has good luck with.

I need to probably look into the NUMA setup on my 1950X.  Probably could be why the times on it are not the best those WUs could be.

 

Linux distros can really run light when setup.  I ran for a time a FX8350 with a very small SSD and barely 4GB of RAM (I could probably shrunk it down to 2GB of RAM depending on project).  Note about my Mint Mate 18.3, it is not running the kernel it comes with.  I switched it out to a newer kernel to get the schedular benefits for my 1900X.

 

26 minutes ago, Windows7ge said:

You know what. A while back I was experimenting with finding the lightest weight Debian Linux distro I could to perform a particular task (the plan being to run about 10 instances of it in a VM server). I wouldn't mind going around again testing a series of random distros to see which has the least overhead when crunching BOINC tasks (best turn around time). It may prove useful to those who many want to virtualize the BOINC manager. This was part of my idea as an alternative form of advanced bunkering.

I be interested in helping with the testing as well.

Hopefully to soon have 5960X up.  That will give me two x99 plateforms and two x399 plateforms to do tests on.  I maybe can get the 4690K up as well to test too.

[Windows7ge]

11 minutes ago, seoz said:

Ubuntu keeps locking up my PC so I've decided to do only 5 cores out of 6 at 80% computing power.

 

I noticed now that the time it takes to complete some projects has shot right up, is this as a result of getting deeper into the BOINC economy or could it be due to my constant rebooting and suspending due to my PC locking up?

Sounds like overclock instability. I've heard the workload BOINC puts on your CPU isn't quite like that of synthetic stress testers or gaming benchmarks used for validating stable overclocks. You may have to take it down a notch to keep all cores active. For this type of workload generally more cores is better than faster cores.

[porina]

12 minutes ago, seoz said:

I noticed now that the time it takes to complete some projects has shot right up, is this as a result of getting deeper into the BOINC economy or could it be due to my constant rebooting and suspending due to my PC locking up?

Units may vary in length, see if it is consistent over time.

 

It also depends on how often units checkpoint. You may lose some progress if you stop-start often but the clock doesn't reset. I don't know how the projects we're doing currently behave.

 

1 minute ago, Windows7ge said:

I've heard the workload BOINC puts on your CPU isn't quite like that of synthetic stress testers or gaming benchmarks used for validating stable overclocks.

Best not to look at boinc as a single thing, but to consider each individual project. Workloads can and do vary a lot.

[jiyeon]

4 minutes ago, Windows7ge said:

Sounds like overclock instability. I've heard the workload BOINC puts on your CPU isn't quite like that of synthetic stress testers or gaming benchmarks used for validating stable overclocks. You may have to take it down a notch to keep all cores active. For this type of workload generally more cores is better than faster cores.

The thing is, on Windows, I would easily do 4.9GHz 1.2V and would never in a million years crash even on CPU-intensive games partnered with my also-overclocked GTX 1060.

 

Now on Ubuntu, I'm doing 4.7GHz 1.25V and I'm getting locks every couple hours.

 

I wouldn't want to go down as far as 4.5GHz because then I feel I'd take a big hit on how much points my BOINC produces.

[GOTSpectrum]

2 minutes ago, seoz said:

The thing is, on Windows, I would easily do 4.9GHz 1.2V and would never in a million years crash even on CPU-intensive games partnered with my also-overclocked GTX 1060.

 

Now on Ubuntu, I'm doing 4.7GHz 1.25V and I'm getting locks every couple hours.

 

I wouldn't want to go down as far as 4.5GHz because then I feel I'd take a big hit on how much points my BOINC produces.

 

7 minutes ago, Windows7ge said:

Sounds like overclock instability. I've heard the workload BOINC puts on your CPU isn't quite like that of synthetic stress testers or gaming benchmarks used for validating stable overclocks. You may have to take it down a notch to keep all cores active. For this type of workload generally more cores is better than faster cores.

Yep you learn very quickly if your overclock is stable with BOINC.

 

My CPUs are clocked right to the limit, 96c on both my main CPUs.

[Windows7ge]

14 minutes ago, seoz said:

I wouldn't want to go down as far as 4.5GHz because then I feel I'd take a big hit on how much points my BOINC produces.

6 cores @ 4.5GHz > 5 cores @ 4.9GHz for BOINC. My servers only go up to around about the low 3GHz if not lower but that's across 48 cores and 96 threads. The productivity is insane because of how parallel the workload is it doesn't just utilize 4 or 6 faster cores.

 

15 minutes ago, porina said:

Best not to look at boinc as a single thing, but to consider each individual project. Workloads can and do vary a lot.

Noted.

 

@Ithanul

When the marathon is over I'll re-lookup a list of random Debian distros and test them all out (assuming BOINC exists in their repositories - I haven't had much luck running it by downloading it from the BOINC website)

[Ithanul]

10 minutes ago, Windows7ge said:

 

@Ithanul

When the marathon is over I'll re-lookup a list of random Debian distros and test them all out (assuming BOINC exists in their repositories - I haven't had much luck running it by downloading it from the BOINC website)

Yeah, I stick to pulling BOINC from repos or repo managers.  Never had much luck with installing from the official website when dealing with Linux distros.

[porina]

Looks like I'm last one in office at work tonight, one more meeting to go... and I can't buy help think about two PCs I have under my control in the lab, on a test network that isn't monitored. They're bought to use up a previous quarter's budget, for something in the future I personally don't think will ever happen. So in effect, two PCs "free" for any use. In the context of this thread, I don't think I need to spell out where this is going...

 

My problem is they're Dell mini-towers and the cooler isn't much better than the Intel stock cooler. I just tried WCG on one of them. A single task running, it increases the noise over idle but not objectionably so. Two tasks, it is starting to get noisy. All out, 12 tasks! Did someone just start a jet engine? Noise will draw attention so back to 1 task at a time. Not the best use but due to the layout, I can't even put a better cooler in there as it is too limited in height.

 

I'm wondering if running XTU to limit power would give better throughput. It is interesting, taking around 20W with 1 task, 30W with 2 tasks so scaling is weird. It hits 95W for a short time with 12 tasks, before dropping back to the 65W TDP limit. It seems Dell do follow that. Shall have to have a little think about this...

[Windows7ge]

1 hour ago, Ithanul said:

Yeah, I stick to pulling BOINC from repos or repo managers.  Never had much luck with installing from the official website when dealing with Linux distros.

You mentioned you're using a non-standard kernel for your mint mate install. How does one upgrade or change the kernel?

[GOTSpectrum]

Can't wait to drop these bunckered einstein WUs.

 

Going to be interesting to see what we can do on the GPU front.

[Egg-Roll]

My server has 44 pending but only one verified, is that normal?

[GOTSpectrum]

Just now, Egg-Roll said:

My server has 44 pending but only one verified, is that normal?

Yeah, when WUs are given it they are given out in pairs(or maybe more, not sure(, you have to wait for the twin to be submitted to verify the results.

[Ithanul]

2 hours ago, Windows7ge said:

You mentioned you're using a non-standard kernel for your mint mate install. How does one upgrade or change the kernel?

Since it is Mint and a branch off Ubuntu, I use Ukuu to do the kernel switch outs.  I advise, do not have GPU drivers installed when doing the kernel switch out if Nvidia is the GPU.  Install drivers after doing the kernel switch out to minimize headache.

 

2 minutes ago, Ben Quigley said:

Yeah, when WUs are given it they are given out in pairs(or maybe more, not sure(, you have to wait for the twin to be submitted to verify the results.

?  WCG is quorum of 1.  It don't need a partner to complete their WU.  

 

Einstein@Home is a quorum of 2.  Those will require a partner to complete their WU.

[GOTSpectrum]

THIS IS A PUBLIC MOTIVATIONAL MESSAGE!

 

We are putting up a valiant effort guys, currently sitting in 14th, keep it up, push your sillicon to the extreme, harness every core you can find! Your gaming rig, your HTPC, your file servers, hell even your mum's old dell. 

 

LETS SHOW THEM TEAM LTT IS IN THIS FOR REAL! 

 

I'll review production this evening and get back to you all if any changes are needed.

 

Don't forget the GPU project is up in a couple days so you may want to bunker if you can with VMs where possible.

 

Day 4 report is live: 

 

https://www.seti-germany.de/forum/content/1047-BOINC-Pentathlon-2019-Day-4

[GOTSpectrum]

6 minutes ago, Ithanul said:

?  WCG is quorum of 1.  It don't need a partner to complete their WU.  

 

Einstein@Home is a quorum of 2.  Those will require a partner to complete their WU.

This I did not know,

 

Any ideas on how they validate then?

[Windows7ge]

14 minutes ago, Ithanul said:

Since it is Mint and a branch off Ubuntu, I use Ukuu to do the kernel switch outs.  I advise, do not have GPU drivers installed when doing the kernel switch out if Nvidia is the GPU.  Install drivers after doing the kernel switch out to minimize headache.

Do you remember the specific version you're on? Kind of important so I don't just install the latest and hope for the best.

[leadeater]

4 hours ago, porina said:

Be careful with suspending tasks. My understanding is if any task is user suspended, then boinc will not get any new work even for other projects.

Yea I'm not dual project on the same instance but rather on different ones on same server, did that so it wouldn't be a problem.

[porina]

27 minutes ago, Ben Quigley said:

Any ideas on how they validate then?

If it is like other projects, they require two different people to return a unit and the results match. If they don't, another unit is sent out for completion until there is a match. Without this kind of system, it increases the risk from faulty hardware returning bad results, or even worse, intentional cheating for credits.