On Wednesday, May 09, 2001 06:57:59 AM -0700 Hans Reiser <[EMAIL PROTECTED]> wrote: >> plan A might help, but I doubt it. Any operation on the file ends up >> logging the inode, which ends up searching for the inode in the tree all >> the time. I'd love to make some stats on how often searches are done for >> stat data, and come up with caches to lower that number. > > you are saying it is atime that kills our performance? Do you remember > our long ago discussion of making atime dirties of leaves not be logged? > Any reason other than lack of time for not doing that? atime hurts us, but so do the general fact that anytime we log an inode, we have to search for it in the tree. This increases our CPU and ram utilization. There are two fixes. First, make our own dirty inode lists, so we can allow inodes to get dirty (right now we always log inodes, to prevent deadlock against kswapd). This would make atime much less of a performance hit (back to the way it was in 2.2.x). After that's done, I think we should investigate item caches. Both are issues for the medium future. -chris
