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 : General discussion : Things better than what is going on (Message 155511)
Posted 5 days ago by Profile GrebulonerProject donor
What did one math book say to the other?
I 've got so many problems...

I feel sorry for this book.
Being made of paper with printed maths AND having problems!
I imagine getting up to the wrong side of the bed - regardless of which side - can become tedious.


That's why I sleep on a Möbius strip bed. It's always the right side!

That and waking up on the wrong side of a Klein bottle got to be too disorienting.
2) Message boards : Number crunching : Multithreading? (Message 155448)
Posted 11 days ago by Profile GrebulonerProject donor
really getting into the cores x frequency as a measure of general overall performance when I'm looking at different chips for roughly the same price.

Not forgetting architecture makes a significant contribution I hope. For example, Skylake family is about 14% ahead of Broadwell per clock. Zen 2 was about 4% up on Skylake, and Zen 3 another 10% or so on top of that. This applies to LLR like workloads when NOT limited by ram performance.


I'm looking to get used Xeons to upgrade the X99 boards I already have. Broadwell is the latest architecture to hit the datacenter cheap CPU sell-off, but in some cases, moving up 2 cores or adding a few hundred MHz adds $100 to the price, so going to a weird number only to use fewer cores is a waste of money, especially as once you get to about half the cores in use, the all-core lowest turbo bin is in play, so there is no frequency advantage. Cheap Skylake-SP is coming, but the disadvantages lie in needing an expensive server/workstation board and a 6000 series part is the minimum for AVX-512 to be useful.

I'm currently working on the Icelake-SP table and a 26-core has choices of 1, 2, 13 and 26 threads, but the 24-core has 1, 2, 3, 4, 6, 8, 12, and 24 which much more nicely fits the FFT distribution.

At the cost of more work for you to do, there are two more scenarios you can consider:
1, not using all the cores e.g. only 24 cores out of 26. Active cores get more power budget than if all were in use so may clock slightly higher depending on CPU.
2, using more threads than cores e.g. 7 tasks of 4 threads each for 28 threads on 26 core CPU. Guarantees all cores in use.


Unfortunately, working out different possibilities for CPUs by using fewer cores than available is well beyond the scope of the table. When it comes to the very high core count server CPUs, the table eschews model numbers and goes right to "24-core," "26-core", etc. with their common cache quantity. So if one were so inclined, while it might not be a perfect fit, could move up or down to a corresponding row to get a basic idea. It's really a starting point more than anything else, and someone who wants to tweak will work from there and explore the many other posts with real-world results.

Speaking of real-world results, I do want to thank you for all your Ryzen work. The CCX/CCD testing you've done has been a great help on the AMD side of the chart.

3) Message boards : Number crunching : Multithreading? (Message 155436)
Posted 11 days ago by Profile GrebulonerProject donor
I'm making surprisingly good progress so far, at least with Intel desktop CPUs and servers starting with Sandy Bridge. Mobile is a mess to sort out later (as is Alder Lake) and I'm purposely leaving out the low end stuff since it doesn't show up on the CPU performance chart. Peter, when I'm ready to share, I'll start a new thread that might get stickied.

That still won't reach enough people, it needs to be available when you're choosing projects.


I'm not an admin; just a nerd with a math degree, outdated programming skills, no life, and lots of free time. I'm at 400 rows of spreadsheet fun, and it still requires a user to be able to look at their CPU model and match it to a line that is generic across a specific generation like "i5 6000/7000." Things get more involved if you have a Xeon.
4) Message boards : Number crunching : Multithreading? (Message 155433)
Posted 11 days ago by Profile GrebulonerProject donor
I've been looking at getting now-cheap used Broadwell-EP CPUs for my X99s and really getting into the cores x frequency as a measure of general overall performance when I'm looking at different chips for roughly the same price. One of the fun things I'm learning as I make this chart is that for many-cored CPUs, the more factors that the number of cores has, the better result you're likely going to have with crunching if you want to fully utilize your CPU (and for my part, if I wasn't going to use all the cores, I'd just buy a cheaper chip with fewer cores).

For example, I'm currently working on the Icelake-SP table and a 26-core has choices of 1, 2, 13 and 26 threads, but the 24-core has 1, 2, 3, 4, 6, 8, 12, and 24 which much more nicely fits the FFT distribution. On 321, the 26-core would have to run 2x13 threads using just 18/71.5MB of cache vs. 6x4 threads using 54/66MB of cache for the 24-core. In my own experience with the variety of core-count systems I have, keeping the number of threads as low as possible is the most efficient way to go, so that 26-core would likely end up having much lower overall throughput on the non-tiny subprojects.
5) Message boards : Number crunching : Multithreading? (Message 155432)
Posted 11 days ago by Profile GrebulonerProject donor
I'm making surprisingly good progress so far, at least with Intel desktop CPUs and servers starting with Sandy Bridge. Mobile is a mess to sort out later (as is Alder Lake) and I'm purposely leaving out the low end stuff since it doesn't show up on the CPU performance chart. Peter, when I'm ready to share, I'll start a new thread that might get stickied.

Question for those in the know (programming team?): How is cache allocation/task size calculated or viewed for GFN, AP27 and WW? I didn't see anything useful in stderr's. Or is it that these apps are really suited for GPUs and as such are so massively parallel that you just want to assign all cores+logical threads, anyway?
6) Message boards : Number crunching : Multithreading? (Message 155426)
Posted 12 days ago by Profile GrebulonerProject donor
I think there are already several threads on how to determine the optimal settings for one's CPU(s). I can spend a few hours working through task size vs. cache/core just on a single CPU of mine, then come back the next day after running work to check my results and the resulting steady state power/temp/utilization. While I'd have plenty of fun, personally, spending a couple months working up what would be an enormous FAQ, there are additional challenges in that most subproject tasks are increasing in FFT size (i.e. cache amount used), and so the recommendations will change possibly frequently for some CPUs but not others, which requires lots of tracking, and anyone using their integrated graphics for crunching would have to make different choices of threading.

It must be possible to collate this data and provide a rough guide on the website for all to see. In fact.... surely given the cache size and the FFT size, it should be easy to calculate what would be best without having to actually try it?


Yes, a rough guide is absolutely possible, but you've partially solved your own problem in the process: FFT sizes vs. cache. Wikipedia has an essentially complete specs list of all the major CPUs (ES/QS oddities notwithstanding, nor whatever allocation you get from cloud hosts). Look up your CPU's cache, the FFT size of what subproject you want to do (x8 to get MB/task) and do some simple division. Further optimization is then up to the individual's own testing.

I've started working up a guide as a spreadsheet, (goodness knows I'm not going to type it out in a post window!), which would allow for easy(er) updating as FFTs increase (and because I don't have admin privileges). There are literally thousands of different CPU models across vendors and generations that need to be collated into generic categories when possible. Understatement: This will take some time.

If we did it every day, you wouldn't be eating more cheese on any day. ;-)

It would be more than other people, or more than last year. Damn it you just made me eat some.


Excellent, my work here is done. Cheese is so tasty!
/It's just a silly quasisensical catchphrase I made up 25 years ago in high school.
7) Message boards : Number crunching : Multithreading? (Message 155417)
Posted 12 days ago by Profile GrebulonerProject donor
In that case I suggest some kind of FAQ for which processors do best on which tasks at which number of threads. I would think most people don't think about it or have the time to test everything.



I think there are already several threads on how to determine the optimal settings for one's CPU(s). I can spend a few hours working through task size vs. cache/core just on a single CPU of mine, then come back the next day after running work to check my results and the resulting steady state power/temp/utilization. While I'd have plenty of fun, personally, spending a couple months working up what would be an enormous FAQ, there are additional challenges in that most subproject tasks are increasing in FFT size (i.e. cache amount used), and so the recommendations will change possibly frequently for some CPUs but not others, which requires lots of tracking, and anyone using their integrated graphics for crunching would have to make different choices of threading.

Actually thinking about here before my bedtime, when all truly great ideas are had, I might undertake this as a personal side project just for the fun of it, though for several classes of CPU such work would be theoretical until someone with one could provide authentic data to help out (like Alder Lake P+E or newly released architectures that have a different cache paradigm).

And why only Thursdays?


If we did it every day, you wouldn't be eating more cheese on any day. ;-)
8) Message boards : Number crunching : Multithreading? (Message 155415)
Posted 12 days ago by Profile GrebulonerProject donor
The easy answer is user freedom. Let the individual decide what they want to do with their rigs.

The longer answer is that on those projects, the tasks are very small (will almost always fit into the CPU cache with 1 thread/task regardless of core count, though MEGA is more iffy) and for the great majority of PG-connected computers, multithreading is actually less efficient.

Some people still run ancient low-end hardware and multithreading may be the only way they can get tasks done in a certain amount of time, and when being first mattered, it was a chance for them to increase their odds.

Also, it's all the exact same executable for all LLR2 subs. Why spend time on maintaining two very slightly different versions?
9) Message boards : General discussion : Sub-Project Bonuses (Message 155237)
Posted 29 days ago by Profile GrebulonerProject donor
The GFN projects are vastly faster on GPU than on CPU, so while the CPU times are long, the GPU times are quite short and not really worthy of a long job bonus until you get up to the tasks that take a long time on GPU.
10) Message boards : Number crunching : LLR2 installed on all big LLR projects (Message 155140)
Posted 41 days ago by Profile GrebulonerProject donor
This is a kind of official announce that PPSE is running on LLR2 now. Well, it's running LLR2 since April 1st.

Technically, PPSE is still "unknown" state and can be reverted back if servers will be unable to handle the load. To avoid dangerous server overload, we have to apply some restrictions:

- There will be no more PPSE challenges;
- PPSE primes will be not eligible for future Tour de Primes.


I think getting PPSE to LLr2 is a very good thing, Thank You for doing that!! But I have a question for you...are those last two points permanent or just 'for now'. The reason I ask is that one of my teammates crunches PPSE almost exclusively and does love the Challenges, with a single pc though he's never in the list of stats.


To the second point, with T5k average digits getting bigger and bigger, GFN16's days are numbered leaving PPS as the sole "attainable" subproject for most crunchers in TdP. Are there discussions going on behind the scenes about what the future of TdP eligible primes will be? It would be sad that in the near future we'd all be fighting for an elusive megaprime or two for an entire month, instead of what we have now, where even modest searchers can get a few small primes and their participation badges.


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2022 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 0.41, 0.47, 0.56
Generated 26 May 2022 | 21:44:27 UTC