xmrig/doc/CPU.md
xmrig 2d52118c1b
Merge pull request #2161 from coolhaircut/patch-1
Added Userspace MSR permissions clarification in CPU.md
2021-03-08 04:01:15 +07:00

6.5 KiB

⚠️ Recent version of this page https://xmrig.com/docs/miner/config/cpu.

CPU backend

All CPU related settings contains in one cpu object in config file, CPU backend allow specify multiple profiles and allow switch between them without restrictions by pool request or config change. Default auto-configuration create reasonable minimum of profiles which cover all supported algorithms.

Example

Example below demonstrate all primary ideas of flexible profiles configuration:

  • "rx/wow" Exact match to algorithm rx/wow, defined 4 threads without CPU affinity.
  • "cn" Default failback profile for all cn/* algorithms, defined 2 threads with CPU affinity, another failback profiles is cn-lite, cn-heavy and rx.
  • "cn-lite" Default failback profile for all cn-lite/* algorithms, defined 2 double threads with CPU affinity.
  • "cn-pico" Alternative short object format.
  • "custom-profile" Custom user defined profile.
  • "*" Failback profile for all unhandled by other profiles algorithms.
  • "cn/r" Exact match, alias to profile custom-profile.
  • "cn/0" Exact match, disabled algorithm.
{
    "cpu": {
        "enabled": true,
        "huge-pages": true,
        "hw-aes": null,
        "priority": null,
        "asm": true,
        "rx/wow": [-1, -1, -1, -1],
        "cn": [
            [1, 0],
            [1, 2]
        ],
        "cn-lite": [
            [2, 0],
            [2, 2]
        ],
        "cn-pico": {
            "intensity": 2,
            "threads": 8,
            "affinity": -1
        },
        "custom-profile": [0, 2],
        "*": [-1],
        "cn/r": "custom-profile",
        "cn/0": false
    }
}

Threads definition

Threads can be defined in 3 formats.

Array format

[
    [1, 0],
    [1, 2],
    [1, -1],
    [2, -1]
]

Each line represent one thread, first element is intensity, this option was known as low_power_mode, possible values is range from 1 to 5, second element is CPU affinity, special value -1 means no affinity.

Short array format

[-1, -1, -1, -1]

Each number represent one thread and means CPU affinity, this is default format for algorithm with maximum intensity 1, currently it all RandomX variants and cryptonight-gpu.

Short object format

{
    "intensity": 2,
    "threads": 8,
    "affinity": -1
}

Internal format, but can be user defined.

RandomX options

init

Thread count to initialize RandomX dataset. Auto-detect (-1) or any number greater than 0 to use that many threads.

init-avx2

Use AVX2 for dataset initialization. Faster on some CPUs. Auto-detect (-1), disabled (0), always enabled on CPUs that support AVX2 (1).

mode

RandomX mining mode: auto, fast (2 GB memory), light (256 MB memory).

1gb-pages

Use 1GB hugepages for RandomX dataset (Linux only). Enabled (true) or disabled (false). It gives 1-3% speedup.

wrmsr

MSR mod. Enabled (true) or disabled (false). It gives up to 15% speedup depending on your system. (Note: Userspace MSR writes are no longer enabled by default; the flag msr.allow_writes=on must be set for Linux Kernels 5.9 and after.)

rdmsr

Restore MSR register values to their original values on exit. Used together with wrmsr. Enabled (true) or disabled (false).

cache_qos

Cache QoS. Enabled (true) or disabled (false). It's useful when you can't or don't want to mine on all CPU cores to make mining hashrate more stable.

numa

NUMA support (better hashrate on multi-CPU servers and Ryzen Threadripper 1xxx/2xxx). Enabled (true) or disabled (false).

scratchpad_prefetch_mode

Which instruction to use in RandomX loop to prefetch data from scratchpad. 1 is default and fastest in most cases. Can be off (0), prefetcht0 instruction (1), prefetchnta instruction (2, a bit faster on Coffee Lake and a few other CPUs), mov instruction (3).

Shared options

enabled

Enable (true) or disable (false) CPU backend, by default true.

huge-pages

Enable (true) or disable (false) huge pages support, by default true.

huge-pages-jit

Enable (true) or disable (false) huge pages support for RandomX JIT code, by default false. It gives a very small boost on Ryzen CPUs, but hashrate is unstable between launches. Use with caution.

hw-aes

Force enable (true) or disable (false) hardware AES support. Default value null means miner autodetect this feature. Usually don't need change this option, this option useful for some rare cases when miner can't detect hardware AES, but it available. If you force enable this option, but your hardware not support it, miner will crash.

priority

Mining threads priority, value from 1 (lowest priority) to 5 (highest possible priority). Default value null means miner don't change threads priority at all. Setting priority higher than 2 can make your PC unresponsive.

memory-pool (since v4.3.0)

Use continuous, persistent memory block for mining threads, useful for preserve huge pages allocation while algorithm switching. Possible values false (feature disabled, by default) or true or specific count of 2 MB huge pages. It helps to avoid loosing huge pages for scratchpads when RandomX dataset is updated and mining threads restart after a 2-3 days of mining.

yield (since v5.1.1)

Prefer system better system response/stability true (default value) or maximum hashrate false.

asm

Enable/configure or disable ASM optimizations. Possible values: true, false, "intel", "ryzen", "bulldozer".

argon2-impl (since v3.1.0)

Allow override automatically detected Argon2 implementation, this option added mostly for debug purposes, default value null means autodetect. This is used in RandomX dataset initialization and also in some other mining algorithms. Other possible values: "x86_64", "SSE2", "SSSE3", "XOP", "AVX2", "AVX-512F". Manual selection has no safe guards - if your CPU doesn't support required instuctions, miner will crash.

astrobwt-max-size

AstroBWT algorithm: skip hashes with large stage 2 size, default: 550, min: 400, max: 1200. Optimal value depends on your CPU/GPU

astrobwt-avx2

AstroBWT algorithm: use AVX2 code. It's faster on some CPUs and slower on other

max-threads-hint (since v4.2.0)

Maximum CPU threads count (in percentage) hint for autoconfig. CPU_MAX_USAGE.md