Hi, I recently setup a test ppc64le VM, and incidentally noticed some things. Since the page size is 64KiB, the sector size and node size with a default mkfs.btrfs also became 64KiB. As a result, I saw quite a lot of small files stuffed into a single leaf (inline data) compared to the default 16KiB, where often the inline process seems to give up and put even very small files less than 1KiB into a data block group, thus taking up a full 4KiB extent (that's on x86).
Further, due to how compression works on 128KiB "segments" (not sure the proper term for it), there is quite a metadata explosion that happens with compression enabled. It takes quite a lot more 16KiB leaves to contain that metadata, compared to 64KiB leaves. I'm wondering if anyone has done benchmarking that demonstrates any meaningful difference for 16 KiB, 32 KiB and 64 KiB node/leaf size when using compression? All the prior benchmarks back in ancient times when default node size was changed from 4KiB to 16KIB I'm pretty sure were predicated on uncompressed data. It could be a question of how many angels can dance on the head of a pin, but there it is. -- Chris Murphy
