The LLR slowdown across multiple cores is a concern for us, but also it's a mystery. We have yet to determine the cause of it. Therefore, we wish to gather as much information as possible and hopefully the problem will present itself.
We do know that Prime95 (the software used in GIMPS) sees the same slowdowns and it, too, varies from machine to machine. Some suggestions from their observations are:
- The generally accepted reason is too much contention for the memory bus. Although it could be overloading the cache bandwidth or too many "snoop penalties".
- The i7 ought to have little slow down running on the 4 physical cores.
- Many of the Core 2 Quads exhibit slowdowns. And if it the quad is using an nvidia northbridge, you can expect HUGE slowdowns.
If you would like to submit your observations, please provide the following information:
example: Intel, AMD
example: Q9550 @ 3.0GHz
RAM (size, type, speed)
example: PC2-3200, PC2-5300, PC2-6400, PC2-8000, PC2-8500, PC2-9200, PC2-9600
example: ASUS P6T delux
TIMES FOR WU
example: observed run times of actual PG work
A more specific test is one that SETI.Germany was using. They set up a nice test that runs LLR manually in a console. The idea is to run an LLR instance for each core and record iterations. Then decrease one instance and observe iterations in the remaining. Continue this until there's just one instance left. For example:
example: For an X5365 @ 3MHz
8 instances: ~28ms
For a Q9550 @ 3,6MHz
4 instances: ~17ms
Thank you for your time and help.