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. Are you disk bound or CPU bound at this point? > I remount all file systems on the testing path with noacl option, and > get the similar results like under root account. > > 3) aim7 has about 40% regression on Nehalem EX 4-socket machine. The > root cause is the same thing like 2). Thanks for subsequently porting and testing the ACL patch. I saw some performance gains on reaim on 2 socket 8 core machine, although it would depend on the workfile used. Thanks, Nick -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>