E5 2682 V4 *2 very low cpu usage

  • E5 2682 V4 *2

  • Samsung REG ECC 8GB14 + 16GB2

  • DELL R730

  • Samsung 860 QVO1TB 3 2TB1

  • DELL SAS 1.8TB*2

  • windows sever 2019

  • Chia 1.1.6

running 23 plots now 12 in phase one 4 threads 4000 ram

I dont know what is happening.
If you can ask my question thx a lot

Your problem is in your ssd’s, the write speed on those after the cache gets filled is like under 100 MB/s, so basically hdd speed. They wont handle multiple plots per disk

If you want to run a lot of plots on one disk, you need nvme drives with high sustained write speeds

1 Like

Not only are the SSDs and issue but if @Victor922 does not know what NUMA nodes are he needs to learn.

Obviously those QLC SSDs are junk for Chia. Need TLC at the least, MLC optimally. Enterprise MLC for massive TBW if you can afford.

I also have no idea what those are, never seen this term before, then again maybe not surprising as I just got my first ever Xeon system today

Care to explain?

You have two processors. Windows is going in picking randomly what it thinks is good to assign to whatever work. If you have PCI-E lanes, you have to figure out which lanes go to which processors to get best performance, assuming you are putting NVMEs in the PCI-E lanes. You aren’t in your case, but stick with me.

Each CPU is made to use certain PCI-E lanes. Say you slot in one drive in slot 1, and CPU 0 uses slot 1 as a primary. Well that’s good, you should get good performance if CPU 0 does the work. But you have to specify in SWAR, hey use CPU 0 when plotting on this SSD.

If you have CPU 1 use slot 1 PCI-E devices, you might have a 10-20% performance dip.

For you, certain SATA slots could benefit from being mapped to certain CPUs. However, the bigger issue for you is those QLC SSDs you have. They are NOT good for Chia. They have good TBW, 2880TB last I checked, but speed is bad.

Priorities for you are:

  • Better SSDs
  • More memory (you need 128gb to take advantage of the 64 threads you have)
  • More SSDs (you need 2TB per 16 threads, so 8TB total need for you to get maximum parallel plots!)
  • Research your motherboard and understand which CPUs use which PCI-E slots and SATA ports
1 Like