Hello, Michael Kelly, le mer. 22 avril 2026 21:33:01 +0100, a ecrit: > I haven't sampled what percentage of calls to rumpdisk_device_write are > actually aligned so can't really quantify the performance loss. I cannot > however measure any significant difference in average compilation times with > and without these changes so possibly the misaligned writes dominate.
That probably isn't the highest cost at the moment indeed, simply because we currently still mostly use one-page reads/writes, in which case copying 4K data is really fast compared to the RPC etc. costs. When we will enlarge this amount and reduce the RPC etc. costs, the copy cost will show up. At least for now we can use such browntape solution so that at least things work fine. And then I really believe we should discuss with the BSD people to understand how they cope with it. Samuel
