On Thu, Oct 22, 2020 at 8:31 PM Christopher Unkel <cunkel@xxxxxxxxxxxxxx> wrote: > > Hello all, > > While investigating some performance issues on mdraid 10 volumes > formed with "512e" disks (4k native/physical sector size but with 512 > byte sector emulation), I've found two cases where mdraid will > needlessly issue writes that start on 4k byte boundary, but are are > shorter than 4k: > > 1. writes of the raid superblock; and > 2. writes of the last page of the write-intent bitmap. > > The following is an excerpt of a blocktrace of one of the component > members of a mdraid 10 volume during a 4k write near the end of the > array: > > 8,32 11 2 0.000001687 711 D WS 2064 + 8 [kworker/11:1H] > * 8,32 11 5 0.001454119 711 D WS 2056 + 1 [kworker/11:1H] > * 8,32 11 8 0.002847204 711 D WS 2080 + 7 [kworker/11:1H] > 8,32 11 11 0.003700545 3094 D WS 11721043920 + 8 [md127_raid1] > 8,32 11 14 0.308785692 711 D WS 2064 + 8 [kworker/11:1H] > * 8,32 11 17 0.310201697 711 D WS 2056 + 1 [kworker/11:1H] > 8,32 11 20 5.500799245 711 D WS 2064 + 8 [kworker/11:1H] > * 8,32 11 23 15.740923558 711 D WS 2080 + 7 [kworker/11:1H] > > Note the starred transactions, which each start on a 4k boundary, but > are less than 4k in length, and so will use the 512-byte emulation. > Sector 2056 holds the superblock, and is written as a single 512-byte > write. Sector 2086 holds the bitmap bit relevant to the written > sector. When it is written the active bits of the last page of the > bitmap are written, starting at sector 2080, padded out to the end of > the 512-byte logical sector as required. This results in a 3.5kb > write, again using the 512-byte emulation. > > Note that in some arrays the last page of the bitmap may be > sufficiently full that they are not affected by the issue with the > bitmap write. > > As there can be a substantial penalty to using the 512-byte sector > emulation (turning writes into read-modify writes if the relevant > sector is not in the drive's cache) I believe it makes sense to pad > these writes out to a 4k boundary. The writes are already padded out > for "4k native" drives, where the short access is illegal. > > The following patch set changes the superblock and bitmap writes to > respect the physical block size (e.g. 4k for today's 512e drives) when > possible. In each case there is already logic for padding out to the > underlying logical sector size. I reuse or repeat the logic for > padding out to the physical sector size, but treat the padding out as > optional rather than mandatory. > > The corresponding block trace with these patches is: > > 8,32 1 2 0.000003410 694 D WS 2064 + 8 [kworker/1:1H] > 8,32 1 5 0.001368788 694 D WS 2056 + 8 [kworker/1:1H] > 8,32 1 8 0.002727981 694 D WS 2080 + 8 [kworker/1:1H] > 8,32 1 11 0.003533831 3063 D WS 11721043920 + 8 [md127_raid1] > 8,32 1 14 0.253952321 694 D WS 2064 + 8 [kworker/1:1H] > 8,32 1 17 0.255354215 694 D WS 2056 + 8 [kworker/1:1H] > 8,32 1 20 5.337938486 694 D WS 2064 + 8 [kworker/1:1H] > 8,32 1 23 15.577963062 694 D WS 2080 + 8 [kworker/1:1H] > > I do notice that the code for bitmap writes has a more sophisticated > and thorough check for overlap than the code for superblock writes. > (Compare write_sb_page in md-bitmap.c vs. super_1_load in md.c.) From > what I know since the various structures starts have always been 4k > aligned anyway, it is always safe to pad the superblock write out to > 4k (as occurs on 4k native drives) but not necessarily futher. > > Feedback appreciated. > > --Chris Thanks for the patches. Do you have performance numbers before/after these changes? Some micro benchmarks results would be great motivation. Thanks, Song > > > Christopher Unkel (3): > md: align superblock writes to physical blocks > md: factor sb write alignment check into function > md: pad writes to end of bitmap to physical blocks > > drivers/md/md-bitmap.c | 80 +++++++++++++++++++++++++----------------- > drivers/md/md.c | 15 ++++++++ > 2 files changed, 63 insertions(+), 32 deletions(-) > > -- > 2.17.1 >
