Author

Topic: CPU mining - method to keep data in the right RAM? (Read 133 times)

newbie
Activity: 40
Merit: 0
Guess that there's nothing much going on besides crap or beginner talk all over again billion times a day and one in million users actually hase some advanced knowledge here…

From what I found, starting with 7, windblows should have support for NUMA, as well as xmr-stak itself. As it pins memory via hwloc, I presume it does that with NUMA in mind. With xmrig I reach pretty much the same hashrate but that's obvious as these two miners are usually on par as they implemented same function over time. As the HT has been turned on while turbo set to highest, I still have better hashrate than what I found for these CPUs (in a single occasion I actually found actual data for these particular CPUs), about a third, so I think it does work.
newbie
Activity: 40
Merit: 0
Got it, NUMA.

Question is how to optimise for it under windblows? Linux ppl seem to use Docker.
newbie
Activity: 40
Merit: 0
OK ty, topic's been moved.
newbie
Activity: 56
Merit: 0
Hi, you'd probably have better luck in the altcoin mining section here: https://bitcointalk.org/index.php?board=160.0
newbie
Activity: 40
Merit: 0
I am about to start mining some cryptonight (most likely) on couple 12c/24t-per-server Westmere servers. I know about keeping the threads on physical cores, however, I somewhere also read there could also be performance hit if by chance the data actually end in wrong memory. What I mean is, each CPU has its own integrated memory controller and memory banks, but for the OS, the memory is shared in a pool. So it could happen than data worked on by CPU1 end in RAM2 on CPU2. It than has to load them all the way through CPU2 which impacts performance by up to about 10 %.

There is a method for this issue to always keep the data for CPU1 in RAM1 and the same for CPU2/RAM2. I just cannot find how the problem/solution is called right now, anybody would remember?
Jump to: