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 : Number crunching : Year of the Rat Challenge (Message 138989)
Posted 17 days ago by Pavel Atnashev
Setting affinity of threads is a part of the proper configuration. If you don't do it, you're betting that OS's default scheduler plan coincides with the layout of a particular chip. Things get even more complicated when you have NUMA architecture.
2) Message boards : Number crunching : Year of the Rat Challenge (Message 138985)
Posted 17 days ago by Pavel Atnashev
With proper configuration there is zero difference between Windows and Linux. But it could be rather challenging to achieve the proper configuration with BOINC.
3) Message boards : Number crunching : Year of the Rat Challenge (Message 138834)
Posted 21 days ago by Pavel Atnashev
If my math is right, we should find one SR5 prime per 40k tasks. With 10k tasks/day we have a good chance to find one during the challenge.
4) Message boards : Number crunching : Transit of Mercury Across the Sun Challenge (Message 134672)
Posted 146 days ago by Pavel Atnashev
You are giving up a lot of chances to return tasks first. You should download tasks just-in-time, unless you are purposely giving the wingman a chance.


I respect my wingmen. Sometimes it takes 3 months for them to validate a task that I returned in 16 hours.

Since there are diminishing returns on applying more threads is there an optimal number of threads less than the number of cores?


No. Performance is limited by access to data, not by the speed of cores.

How much of the executable needs to be cache-resident to complete a single iteration of LLR?


The code is tiny and not worth the mention. What is really huge is FFT data. 2304K FFT requires 20MB of L3 cache. If you have less cache, or if you run more than one LLR, you get such a huge drop in performance that can't be compensated by less multithreading overhead.

Some of my 2660v2's have just one very slow memory module. It doesn't affect performance a bit because 2304K FFT fits into 25MB L3 cache.
5) Message boards : Number crunching : Transit of Mercury Across the Sun Challenge (Message 134670)
Posted 146 days ago by Pavel Atnashev

I still have to figure out how to convince Boinc on Windows to put tasks on numa nodes. And then stopping it from trowing tasks around from one node to the other.


When I failed at this I started writing my own software. It's the only known way for me.

The bulk of my power is Xeon 2660v2. I run one task per CPU, so it's 144 CPUs at the moment. L3 cache is the king.
6) Message boards : Number crunching : Transit of Mercury Across the Sun Challenge (Message 134668)
Posted 146 days ago by Pavel Atnashev
I have 100% task cache, so you need to cut that in half. 56+88 tasks running.
7) Message boards : Number crunching : Happy New Year Challenge (Message 113081)
Posted 817 days ago by Pavel Atnashev
Is there a program to bind BOINC tasks to NUMA nodes?
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2020 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 5.09, 5.79, 6.15
Generated 3 Apr 2020 | 1:31:01 UTC