Hi all, I have posted most of my benchmark findings in this latest blog post. Feedback welcomed!
Nice! That’s additional info based on what I did here as well:
One thing I’ve definitely learned is that extra ram is basically pointless, there’s virtually no benefit. Don’t build any extra RAM into machines that you plan to use for plotting!
I also agree with you that super extra mega fast disks don’t actually help that much:
The difference between a single NVMe and RAID0 of two NVMes was only roughly 12 minutes.
I noticed this when comparing SATA (a measly 6gbps max!) versus NVME runs. I saved one hour, or 13% of overall plot time. That’s not nothing, but it’s not exactly changing the world, either.
It would be interesting to repeat some of your runs for 2 threads as well.
Oh, Nice! Definitely going to give this a look at.
I was wondering about the difference between SATA and NVMe. I will look at this also. Thanks!
Yes and it seems to me that using RAID (especially software RAID) is adding a lot of potential hassle in terms of making it harder to just pick up drives and move them from one computer to another. Which unless you have @codinghorror 's 10GB ethernet can be a big advantage!
Great post! My feedback would be: while it’s interesting to see the single (non parallel) plot times and benchmarking, I think it would be much much more useful to see parallel plot times as that’s what most folks are doing.
Obviously parallel plot times is much more complicated as they don’t start and finish at the same time (maybe plots per day is the right metric?).
Also think parallel plotting will have very different performance characteristics than single plotting as the bottlenecks could be in completely different parts of the system.
Yes, this will be another project. But it takes much more time to gather the data. Thanks for the input!
I kind of disagree; once you’ve tuned the single plots, parallel will follow the same parameters, just slower due to contention. And you can measure that contention (I/O load, CPU load, RAM load).
Yeah I was thinking the same and that’s probably generally true, but I reckon that once you parallel (especially if you’re hitting it hard) you can get some wierd non-linear results around the limits. I haven’t figured out how to measure IO load yet, maybe that’s my problem!
Hey, there’s this site that has user contributed hardware plotting benchmarks, if you can find your set up or something similar you might be able to have a good point of reference. Hope this helps https://chiametrix.com/