On Sun, Jan 08, 2012 at 04:09:00PM -0800, Don Lewis wrote: > On 8 Jan, Poul-Henning Kamp wrote: > > > > I'm doing som data-mining on a pile of ISO images right now. > > > > I stuck the ISOs on a UFS2 on a flash-disk for speed, and mdconfig(8)'d > > them so I could mount them. > > > > The traffic pattern his "interesting": > > > > dT: 1.003s w: 1.000s > > L(q) ops/s r/s kBps ms/r w/s kBps ms/w %busy Name > > [...] > > 1 733 733 1466 1.3 0 0 0.0 98.2| md39 > > 1 733 733 23449 1.3 0 0 0.0 93.2| da0 > > > > Notice the 1:16 ratio on kBps but 1:1 ratio on ops/s ? > > > > da0's UFS2 has 32k block-size: > > > > magic 19540119 (UFS2) time Wed Jan 4 16:41:47 2012 > > superblock location 65536 id [ 4f046cf5 c30697ee ] > > ncg 104 size 19537685 blocks 19228156 > > bsize 32768 shift 15 mask 0xffff8000 > > fsize 4096 shift 12 mask 0xfffff000 > > [...] > > > > It looks like every 2k read from CD9660 turns into a 32k block > > read in the UFS filesystem, without any beneficial caching happening. > > Probably some confusion about which filesystem layer owns the cached > data. It would probably be inefficient to cache the data in both > places. The best fix would probably be for CD9660 to think that the > underlying device has 32Kb sectors. I discussed the issue with phk further. The reason for discarding the 30K of the read 32K block is that md(4) supplies IO_DIRECT flag for VOP_READ, and FFS avoids putting the read data into any cache. Most likely, we can implement a sysctl that would disable direct reads, at the cost of double-buffering the data. For obvious reasons, it is impossible to disable caching from the filesystem living on top of md(4) volume.
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:23 UTC