Andreas Dilger wrote: >On Mar 04, 2002 00:32 -0500, Chris Mason wrote: > >>Ok, I'm not going to be able to replicate the entire test, but I can >>at least demonstrate the high number of subdirectories is slowing down >>the creation time. I'm guessing it is either caused by the >>subdirectory inodes not being in cache often enough, or increased >>log traffic. >> >>Try setting the number of subdirectories to 10. If this fixes the >>reiserfs performance problem, we can look into solutions. >> > >Hmm, interesting. When Andrew was doing MTA performance testing on >ext3, he found that _increasing_ the number of directories improved >performance. I don't have the thread handy, but I _think_ it had to >do with VFS locking on the directories - more directories means that >more operations can be done in parallel since Al made this part of >the VFS thread-safe. > >Cheers, Andreas >-- >Andreas Dilger >http://sourceforge.net/projects/ext2resize/ >http://www-mddsp.enel.ucalgary.ca/People/adilger/ > > > It is a performance problem for us that Viro giant locked directory ops, because reiserfs performance is highest for putting everything in one directory because we don't have linear directory search issues. (having subdirectories when there isn't inherent meaning to them is an ugly kludge old fashioned filesystems often require;-).)
We had this problem with squid. Hans
