Those write speeds below 100MBps are quite common on the last ~500GB of free space.
I remember that early plotting was done with SSD buffers between the plotting nvmes and the farming hdds, so that the plotters could resume their work. Ultimately the optimum was to have multiple transfers going out of the plotting servers to different jbods/nas/hdds, from those temp SSDs.
It doesn’t look like GPUs will speed things up too much, so this strategy could be copied.
3 Likes
85 MB/s is about what I get writing to my drives in a DS4246 as well. I imagine you’re doing daisy chaining too?
Anyway your question is a very good one that not many people are thinking about yet. GPU plotting will be closer to 2-5 minutes on pretty reasonable hardware. Even at full HDD write speeds you’ll need to be writing to multiple HDDs in parallel or be okay with your plotter sitting idle part of the time (which isn’t a big deal tbh, you’re still way faster than CPU plotting). A fast SSD buffer drive would help as well (like a few SSDs in RAID0).
2 Likes
I solved my slow write speeds, I was able to achieve 165MB/s to about 150MB/s as the drive filled. I had just converted my farm over to Linux from Windows and my drives were still NTFS, once I formatted my test disk with EXT4 I was able achieve the higher write speeds. So that helps. So when replot 2.0 start, I will convert the rest of my disks to ext4. I should be able to keep up with my two plotters at that speed, until I start playing with the GPU plotter then I will probably have to setup a SSD buffer. Let the fun begin.
2 Likes
Be careful, as Linux will use whatever available RAM, and move those plots first there, thus your mv, etc. command may/will show higher rates than the actual ones (rates of removing file from the src, not transferring it to the dst). I use bpytop, as it shows direct disk io. This is more pronounced toward the last 500 GB or so of free disk space.
The same thing happened to portable plots. I will need to write a script that plots one thing, delete one thing, copy over. Since the size is smaller this time, I will need to check if the deletion is needed. If not, then copy straight.