I've filed a bug about the performance of fuser(1M) (6878048: fuser performance is suboptimal, which hasn't shown up on bugs.opensolaris.org as I write this.) I've also coded up a fix that makes the performance more reasonable. (I've requested a sponsor on request-sponsor, but I figured I'd add more detail on this list in the hopes of finding someone interested in picking this up.)
The data look like this: AS segments current with fix 0 10.93 5.72 200000 40.08 5.73 400000 69.22 5.74 600000 98.42 5.78 800000 128.05 5.81 1000000 156.22 5.75 1200000 184.31 5.69 1400000 211.76 5.74 The problem with the current implementation is that dofusers() (the kernel function involved) is looking at every address space segment in every process to determine whether it matches the vnode in question. (This is similar to what rm_assize() was doing when reading /proc/<pid>/psinfo until the fix for 6801244 went in.) The fix I've implemented adds a hash table to struct as to keep track of the vnodes that have been mapped in to a process. dofusers() walks that hash table rather than the whole AS. This solution is still going to be O(number of vnodes) rather than O(1), but that seems unavoidable. Chad _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
