On Thu, 2010-08-05 at 20:55 +1000, Nick Piggin wrote: > On Wed, Aug 04, 2010 at 09:04:03AM +0800, Zhang, Yanmin wrote: > > On Tue, 2010-07-20 at 13:12 +1000, Nick Piggin wrote: > > > On Thu, Jul 08, 2010 at 04:56:27PM +0800, Zhang, Yanmin wrote: > > > > Nick, > > > > > > > > I work with Andi Kleen and Tim to investigate some scalability issues. > > > > > > > > Andi gave me a pointer at: > > > > http://thread.gmane.org/gmane.linux.kernel/1002380/focus=42284 > > > > > > > > Where can I get your latest patches? It's better if I could get patch tarball. > > > > > > > > Thanks, > > > > Yanmin > > > > > > > > > > Hi Yanmin, > > > > > > Sorry for the delay. I have a git tree now, and it has been through > > > some tress testing. > > > > > > http://git.kernel.org/?p=linux/kernel/git/npiggin/linux-npiggin.git > > > > > > I would be very interested to know if you encounter problems or are > > > able to generate any benchmark numbers. > > Nick, > > > > We ran lots of benchmarks on many machines. Below is something to > > share with you. > > Great, thanks for doing this! > > > > Improvement: > > 1) We get about 30% improvement with kbuild workload on Nehalem > > machines. It's hard to improve kbuild performance. Your tree does. > > Well that's nice. What size of machine is this? Did you run it on an > ACL enabled filesystem? > > > > Issues: > > 1) Compiling fails on a couple of file systems, such like CONFIG_ISO9660_FS=y. > > Yes there are a couple that broke, which I still need to fix up. > > > > 2) dbenchthreads has about 50% regression. We connect a JBOD of 12 disks to > > a machine. Start 4 dbench threads per disk. We run the workload under > > a regular user account. If we run it under root account, we get 22% > > improvement instead of regression. The root cause is ACL checking. > > With your patch, do_path_lookup firstly goes through rcu steps which > > including a exec permission checking. With ACL, the __exec_permission > > always fails. Then a later nameidata_drop_rcu often fails as > > dentry->d_seq is changed. > > > > With root account, it doesn't happen. We mount the working devices > > under /mnt/stp/XXX. /mnt is of root user. So the exec permission > > check is ok. > > Yes if running with root, this should have the same effect as the > rcu-walk aware ACL patch. > BTW. dbench has a nasty call to statvfs() > which is a huge cost (which should be fixed in future versions of > kernel+glibc). You can try switching the statvfs(2) call in fileio.c > to statfs(2) and see if performance improves. I change the statvfs call to statfs, and get 40% improvement with your patch. pure 2.6.35-rc5 also gets 40% improvement. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
