Hi everyone, first post here , having a lot of fun reading through the forums.
I recently built a PC and wanted to use it to familiarize myself with Chia concepts while waiting for pools to arrive. The relevant specs are:
Ryzen 5 5600X (6 core/12 threads)
B550 MOBO (one of the M2 slots is PCIe 4.0 and the other PCIe 3.0)
32GB RAM
2x Corsair MP600 Core 1TB NVMe SSD
I’m experimenting with different plot settings in plotman. Making 2 plots at a time (one on each disk, 4 threads for each) results in plots being completed in around 4 hours (with the drive in the 3.0 slot being around 15 min slower), which seems reasonable. But trying to plot in parallel results in degraded performance and a lot of IO wait. Is this consistent with the sustained read/write performance of the MP600 Core? Or should I check for any other bottleneck in the system?
I was trying to read through benchmarks of the MP600 Core, but I don’t really know which specific tests indicate good plotting performance. I also used GNOME’s Disks utility and ran some benchmarks, would happily share them if useful.
Due to how modern SSD’s work the thing to look for is sustained high volume write performance. SSD’s have a small amount of SLC cache, which can be written very fast. But if that fills up, as it always does with Chia, it has to fall back on it’s native TLC/QLC performance, and then the smartness of the controller comes into play to limit the damage somewhere between x3 and x8 slowdown. So I’m afraid in your case, it does not look ideal.
I was afraid that would be the case. Ideally, I would like this system so be able to do 6 plots in parallel across two NVMes. Do you think a pair of 980 Pros could achieve this?
Yeah, the thing is that the performance degrades rapidly because most of the plots are stuck on IO wait. So the questions would be, would a pair of 1TB 980 Pros would be able to do 6 plots in parallel without much IO wait?
I do it with same setup 1 MP600 and 16 GB of ram.
6 Plots running at a time doing about 1.6 TB a day. Its all about how you configure
I bet your degradation is due to lack of using discard in the mount options
I am currently mounting the NVMes with btrfs with discard set to async, so I doubt that’s the issue here. I may give it a shot with a manual fstrim on a cron job. How are your plot times with the 6 in parallel?
Also go to XFS performance is way better for multi threaded writes. You also killing your disk with all the extra writes and btrfs is slow. I will be experimenting with ext4 (not great stock standard) and some additional tweaks .
Thanks for the tip, will try XFS to see how it goes. I did try EXT4 before, I don’t have much data right now, but IIRC performance was very similar (if not a bit worse than BTRFS). Will report back all the relevant numbers once the new NVMe’s arrive.
, having a lot of fun reading through the forums.