On Tue, Jul 21, 2020 at 01:42:54AM -0600, Andreas Dilger wrote: > > I re-reviewed the patch with the changes, and it looks OK. I see that > you reduced the prefetch limit from 32 to 4 group blocks, presumably to > keep the latency low? It would be useful to see what impact that has > on the mount time and IO performance of a large filesystem. No, I didn't change it. As before, it's 4 times the size of the flex_bg size (assuming the file system has flex bg's enabled); otherwise it's 128 (4 times 32). I'm actually worried that this is too aggressive on storage devies where LBA's which are adjacent to each other aren't necessarily adjacent to each other on the physical storage device. For example, on dm-thin, some qemu-img formats, and potentially some cloud virtual block devices.... Unfortunately, there isn't a good way to query a block device to determine whether adjacent LBA's are really adjacent in the real world. - Ted
