Threadripper Lackluster Performance

Chia Plotting
13

oh, apparently edited this out accidentally:
it’s 4x 2TB
with 2 980 Pro an two equivalent performing Samsung OEMs

3 of them are on the motherboard, one in a PCIe card adapter all performing to spec.

PCIE 4.0 vs 3.0 Speeds seem to make quite the difference.
Striped vs 4 individual seem to make no huge difference in practice, as this still looks largly bottlenecked by CPU. Will probably switch it back to striped, since 10GB+/s single file read speads are something to behold.

8x16GB RAM

So make sure your not throttling your Gen4 NVMe with a PCIe 3.0 motherboard or starve your CPU with empty Memory Channels.

With the config below, they clock in about with 12 hours per plot, so reliably in high 60s/day for me when going for 36 concurrency.

16 stage 1 plots filling up the 8 cores not blocked by the 16 2+ phase plots seems to be the best balance to me. If you do that single or in 2 plots or two threads for stage 1 on seems to be not much of a difference, but I think to notice less overhead in single threaded runs now.

My current working theory for cpu limited high parallel plotting, given phase one is not maxing out cores:
Have 1.5x concurrent thread count of physical cores at any time in evenly staggered plotting
Make sure half of them are in phase 1, so later phase threads shouldn’t exceed 2/3 of physical core count.

Here one of the 4 current swar-chia-plot-manger job configs for reference. Set minimum_minutes_between_jobs to something like 6 minutes.

     - name: main1
        max_plots: 9999
        farmer_public_key:
        pool_public_key:
        temporary_directory: P:\plot
        temporary2_directory: 
        destination_directory: F:\
        size: 32
        bitfield: true
        threads: 1
        buckets: 128
        memory_buffer: 3800
        max_concurrent: 9
        max_concurrent_with_start_early: 12
        stagger_minutes: 72
        max_for_phase_1: 8
        concurrency_start_early_phase: 5
        concurrency_start_early_phase_delay: 1
        temporary2_destination_sync: false

To boost up launch phase:

  • set stagger_minutes to 0 for each job
  • run for about two hours, or until you hit 24 concurrency,
  • set back to 72 minutes (slightly below total time/concurrency per job)
  • then restart the plot manager.

Or just have a second config file with this boost config swap/rename them accordingly.

1 Like