PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Extended Sierpinski Problem : k = 202705 (Message 152454)
Posted 49 days ago by Profile Rafael
And I just switched to ESP 2 days ago!

Dayum, should have done that last week....
2) Message boards : Prime Sierpinski Problem : Optimizing PSP processing with LLR2 .... (Message 152276)
Posted 66 days ago by Profile Rafael
A question for the group: how should I allocate the six physical cores of my i7 8700k CPU when processing PSP units? Do I go after 3x2 cores, 2x3 cores or perhaps 1x4 / 1x5 / 1x6 cores?

I am running a task right now with 5 cores allocated to it on my i7 8700k and it seems that it will take about 27 hrs to complete. Similarly, another system (i3 6300) will require about 32 hrs with only two cores to complete, thus suggesting that the five cores with the i7 8700k are an overkill and I am wasting potential resources.

With LLR2 it is not vital that we rush to return a task and, as a result, I can easily afford a slightly slower return if it will mean greater throughput per unit of time.

Thanks!


Rule of Thumb #1: always have the sum of cache usage for all your tasks a bit below your CPU cache.

Looking at single one of your tasks, it's using 2880k * 8 = ~23MB of cache, which is above your CPU's 12MB limit. So you should run fewer tasks to reduce usage... but since you can't run fewer than one, just run that single task with as many cores as you have.
3) Message boards : Number crunching : "Game Ready" vs. "Studio" driver [Nvidia] (Message 151670)
Posted 112 days ago by Profile Rafael
I recently got a new machine and it came with software called GEFORCE (R) EXPERIENCE (TM) lol

There's an option to switch from the current default "Game Ready" driver to the "Studio" driver. Does anyone know whether however they differ matters for crunching? If it matters, I'm planning on mostly doing GFN >19 & DYFL. Thanks!

They really don't, that switch is just a way of saying "most recent release" and "has been around for a while now so probably there aren't any major bugs, but no new features either".
4) Message boards : Extended Sierpinski Problem : How to Conf. (Message 150108)
Posted 280 days ago by Profile Rafael
This is what I am referring to.
What do you think? should I change something?
Thx.

Ah, I see.

For starters, leave it at 100%, no need to limit the CPU you've paid for.

As for the temperatures, they are actually not high. If that's what you get while running default ESP, you are not doing too bad.

With that in mind, it should be fine to try and boost performance. First order of business, chose between: 1- disable SMT in your BIOS; 2- Go into BOINC -> Options -> Computing Preferences -> Use at most xx% of CPUs (the first field) and set it to 50%. The 1st option will generally offer better Primegrid preformance, at the expense of lowering computing power for everything else due to disabling half of the threads on your computer. Choose one of them, don't do both.

Then, you should set multithreading (different than SMT) in Primegrid's preference page and set "Multi-threading: Max # of threads for each task". You could run tests to see what works best for your system, but a rough bet would be to set it to 4: that means that each task will use 4 cores, and you will be running 3 of them at the same time. Setting to 3 might be better, but it would have some funky interactions with CCX and CCD inside your CPU and it might actually hurt performance; same for 5. So set it to 4 and forget about it if you don't want to spend time learning what's behind that number and how to better find out.

And always remember: keep a look out for your temperatures. If they get to 90, stop, for you need to mess with your system to lower it.
5) Message boards : Extended Sierpinski Problem : How to Conf. (Message 150104)
Posted 280 days ago by Profile Rafael
Hello everyone.

I would like to know how to configure my pc to get the maximum performance in ESP tasks.

I have a Ryzen 9 3900X (limited to 97% by temperature) and Windows 10.

Any suggestions or ideas are welcome.
(Sorry for my english, I use an automatic translator).
Thanks ^^

"Limited to 97%"...? Can you be a little more specific? Because if you're actually having temperature problems, fix that first before bothering optimizing PG performance, as doing so will only increase heat generation.
6) Message boards : Number crunching : Too much work when joining (Message 149954)
Posted 286 days ago by Profile Rafael
That would mean I should be fine using some of my threads for Primegrid and some for a different type of program. Then again, would the OS know what to pair with what? Or does the CPU sort that stuff out?

That's up to the OS to figure out scheduling. As for knowing what to do with it... linux tends to do better in that regard.
7) Message boards : Number crunching : Too much work when joining (Message 149948)
Posted 286 days ago by Profile Rafael
This does beg the question though, does this mean other projects have sloppy programming compared to Primegrid? Could their programs be fine tuned? Or is it just that Primegrid by it's nature does more RAM access and floods the cache if HT is used? Climate prediction forums discussed a shortage of cache in a similar way. They're large data sets, so it could be the same thing as here.

Not necessarily. A proccessor has a lot of circuits designed to do different things, but it turns out the kinds of maths being done only use a few of them, so there's no room to fit in multiple instances. Other apps might use a whole bunch of stuff, and so while one circuit is busy doing one thing, you could make use of a another for a different instance that's calculating a different thing.

Obviously this is an oversimplification, but I hope it illustrates the point.
8) Message boards : Problems and Help : BOINC CPU Settings and GPU tasks (Message 149508)
Posted 309 days ago by Profile Rafael
Hi.
Currently BOINC is allowed to use 6 out of 16 threads on my Ryzen 7 5800X, which means i run SOB with 6 cores. I also run a GPU task and it also needs a cpu thread?

The question is, should i let BOINC use 7 out of 16 threads so the GPU can get a core as well for better performance? Or does the GPU app do this automatically?

The thing i want to avoid is the GPU app sharing 1 of the 6 threads with SOB slowing it down.

GPU core sharing is hardly a problem, especially since you have an SMT capable CPU. Use all 8 cores for SoB, it will hardly make a difference for the GPU, but will surely help on the SoB end.
9) Message boards : Number crunching : a variety of CPU generations, overwelmed by all the subprojects in a good way (Message 149265)
Posted 318 days ago by Profile Rafael
Does the 3600x suffer from this same issue regarding CCX units?

Has 2 CCX as well. AFAIK the 3300x is the only non apu Ryzen 3xxx that has a single CCX.


How does Linux fair in this regard, majority of my hosts are running Ubuntu 20.04.

Apparently Linux does a lot better than Windows in this regard and there's little need to deal with affinity.

I do have a 1800x running Windows. Does this still apply on older Ryzens?

Yes.

I am assuming re: LLR and LLR2 instruction set support remains the same e.g. avx?

I think LLR2 can't run on 32bit? Not sure about that.

So to be clear - is hyperthreading (Boinc at 100% CPU on hyperthreaded processors like Ryzen 3700x) useful for non LLR projects, or should I leave it disabled?

It's generally useful for non LLR projects and thus should be enabled most of the time.
10) Message boards : Number crunching : a variety of CPU generations, overwelmed by all the subprojects in a good way (Message 149252)
Posted 318 days ago by Profile Rafael
As far as threads are concerned, what is more benifitial?
Let's give a Ryzen 7 3700x, for instance. Would running 4x 2 threads give more throughput than 2x 4 CPUs? Ultimately - what should my end goal be, or does that also depend heavily on subproject?

It depends heavily on the subproject, or rather the size of the FFT used. The rule is to try and use as few threads per task as possible while still keeping everything within CPU cache. In your case, you have 32MB of cache, so something like Sob that's going at 3200k*8 =~25MB of cache is probably better left as 1x8threads rather than 2x4t, whereas something smaller like ESP at 1600k*8 = 12.8MB would be better at 2x4t since it all fits.

With that said, this is just an educated guess and the best option is for you to test yourself. You might also be fine with suboptimal settings in the name of doing individual tasks faster, or perhaps leave a core free for the GPU or your daily driver needs. It's all up to you.


If I remember correctly, SGS and are all single threaded anyway and only the LLR projects are multithreaded

SGS and PPSE use LLR, and while not very efficient, you can multithread it if you want to. You might be mistaking it for LLR2, which is a similar algorithm used for most other CPU projects and is generally a bit slower on it's own, but significantly cuts down on double checking work (along with some other goodies).

so will lack of hyperthreading in the single threaded projects be hurtful?

HyperThreading (or SMT in AMD's case) is generally not recommended for LLR/2 tasks. It is recommended that you either have BOINC use only 50% of the CPUs or turn it off in BIOS.


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2022 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 2.96, 2.30, 2.39
Generated 21 Jan 2022 | 4:15:52 UTC