62 CPUs Hash rate = 40 CPUs Hash rate

[hhajian]

Hi Folks,
I am running xmr-stak-cpu miner on AMD Opteron 6272 which support AES. Other than OS (CentOS 7.3) nothing else is running on this server. What I observe is that, when I run the miner with 62 CPUs it gives me the same hash rate as if I run it with 40 CPUs. In best case I get around 1.6 -1.8 kH/s. Here is the part of the config:

"cpu_threads_conf" :
[
{ "low_power_mode" : false, "no_prefetch" : true, "affine_to_cpu" : 0 },
{ "low_power_mode" : false, "no_prefetch" : true, "affine_to_cpu" : 1 },
....
....
{ "low_power_mode" : false, "no_prefetch" : true, "affine_to_cpu" : 61 },
],

Can any one explain why I dont get higher hash rate?

Thank you.

[flashfie1]

simple you ran out of cpu cache. you need 2MB of cache per instance. You only have 1MB per core, so you can efficiently only use half. allocate 8 threads to each cpu.

[setuidroot]

I run these same type of CPUs and have found the best way to configure them for optimal performance and lower power usage. Send me an email [email protected] and I'll help you specifically. Or you can hit me up on Telegram (@LoginRoot.)

In short, you want to run one instance of stak for each CPU socket, assuring that the CPU affinities are confined to the lowest latency NUMA nodes for each CPU. You will have best results utilizing some "low power mode" double thread affinities per CPU sticking to physical cores and first group (or lowest latency) NUMA node per CPU socket. If done correctly, these CPUs are pretty sweet... they kick major ass with Aeon mining, given the 3-4x hashrate boost for CPUs on cryptonight-lite as well.

To see the NUMA nodes just run "lscpu"

Anyone with AMD Opteron 6XXX series CPUs looking for help with this, feel free to email me as listed above.

I have been mining with and working with a brand new XMR pool (luxpool) with 4x 8 core Opterons and have achieved 1.6 kH/s peak at times when running an optimal configuration. I have found best performance in my servers has been with 4x 2GB DDR3 RAM DIMMs (2 per socket, one in group A and one in group B.) But YMMV given server brand/board type. I am running Ubuntu 16 LTS server with pre-spectre kernel version (post spectre patches does NOT hurt AMD performance, but does cause slight increases in power draw.) Do NOT run outdated kernels on any server unless it is a dedicated mining server running safely behind firewall/NAT that allows no inbound connections. Don't even think of doing this with Intel CPU servers on a public IP... that's too much risk for little (if any) gain.