My CPU is a Threadripper 3960X with 24 cores, 48 threads. I have 128 GB of RAM.
For plotting, I use a 2TB Inland Premium NVMe. I’m assuming this is the bottleneck in my system. Can you suggest how I can make the most out of this configuration?
Currently using the default Plotman settings:
# Optional: default is 1
# Don't run more than this many jobs at a time on a single temp dir.
# Don't run more than this many jobs at a time in total.
# Don't run any jobs (across all temp dirs) more often than this, in minutes.
Since I only have 1 plotting drive, but a CPU/RAM that should be able to handle at least 12 plots at a time, would you suggesting upping tmpdir_max_jobs to 12?
That should get you a ~6.5hr plot time. If that’s the case and the above settings result in little to no io/wait time, bump tmpdir_stagger_phase_limit, tmpdir_max_jobs and global_max_jobs each by 2 until you see unacceptable IO wait times. …AKA your NVMe is totally pegged.
Keep in mind that, no matter what you call it (or folders), in your setup, you have a single tmp dir. So, tmp dir jobs = global jobs. Also, keep in mind that for now, only phase 1 is multi-threaded. So you won’t see 6threads/plot through the duration.
Thank you @danarbraz ! Yes, this system is currently only used for plotting. I’ll give those settings a try. By IO wait times do you mean the “io” column in “plotman interactive”, or are you using something else to determine those?
Yes. That’s IO wait. You don’t want that being in the high double-digits in mins consistently . It’s ok to have some IO wait but the more you have, the less efficient which will mean less plots over time.
…and also know that your destination matters (the speed or lack of it). The final copy alone can lead to wait times. Totally fine but you need to be able to see the diff between wait time during plotting vs wait time for a completed plot to copy to the destination drive. If you only see wait time increase in stage 4, that’s going to be a result of time to copy to dst.