On Mon, Aug 3, 2020 at 1:30 AM Dale <[email protected]> wrote: > > Little update here. Rich, I think you mentioned it would slow down when it > ran out of PMR space while trying to redo the shingled part. Up until now, I > hadn't ran into that issue. It seems the PMR section for this drive is > somewhere around 40 or 50GBs, maybe 60GBs. I hadn't had time for backups in > over a week so it was a good bit larger than usual. It was around 70GBs, > maybe 75. When it got close to the end of the rsync process, I noticed it > slowed down quite a bit. I'd guess about half or so. Usually it runs at > around 180 to 190MBs/sec for larger files. Pretty close to the end, rsync > was showing around 100MBs/sec at best. It was a little over on some but > mostly a little below that. Earlier in the process, it was the normal speed.
I doubt this particular drop is the result of SMR, assuming 100MB/s is the instantaneous speed. 100MB/s is still reasonable for a hard drive - on newer CMR drives I've seen the speed of dd drop from 200MB/s to 100MB/s for sequential writes as the heads move from one end of the drive to the other, and then it goes back up to 200MB/s if you start over at the beginning (badblocks testing and so on). That level of drop is probably more likely to be due to filesystem overhead and so on - fragmentation/etc. When SMR buffer overrun occurs you REALLY hit a wall and the rates drop quite a bit more than that. If it were a difference of only 50% most would probably tolerate it. Now, if 100MB/s were an updating average across the entire run then that would be a different matter, because that would mean that it was running at 200MB/s for most of the run, and then probably dropping much closer to 0 for a while so as to drive the overall average down to 100MB/s. I'm not sure where you're getting those numbers from so I don't know what period that 100MB/s reflects. For an instantaneous speed I'd consider it a completely reasonable performance for a typical hard drive when you're writing to a filesystem. If you were using dd or maybe copies of very large files on an efficient filesystem you would get better results. -- Rich
