Hi George, I noticed that the credit2 scheduling algorithm has one runqueue every L2 cache that is shared by cores in one socket. If vCPUs, which need to communicate more with each other, are assigned to the cores sharing the same L2 cache, the L2 cache miss ratio of Guest VM may be reduced. And also, the performance of the Guest VM will be improved. Is it right? But, how to test the cache miss ratio of L2 in Guest VM? Are the tools that used to test the cache miss ratio in non-virtualized environment can be used in Xen Guest VM? Thank you very much. -- Best Regards, Gavin _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
2012/1/10 gavin <gbtux@126.com>:> I noticed that the credit2 scheduling algorithm has one runqueue every L2 > cache that is shared by cores in one socket. If vCPUs, which need to > communicate more with each other, are assigned to the cores sharing the same > L2 cache, the L2 cache miss ratio of Guest VM may be reduced. And also, the > performance of the Guest VM will be improved. Is it right? But, how to test > the cache miss ratio of L2 in Guest VM? Are the tools that used to test the > cache miss ratio in non-virtualized environment can be used in Xen Guest VM?Looking through the code, it looks like performance counters are passed-through to HVM guests, both on AMD and Intel. And it appears that Fujitsu is regularly using the performance counters on Intel at least (since Deitmar Hahn has been the one posting patches for new performance counters). It appears that there is no support for access to performance counters in PV mode, other than to use them for xenoprofile (which will not allow you to read the counters directly). So it''s worth trying whatever method you normally use to look at performance counters while running in an HVM guest. -George
Am Mittwoch 11 Januar 2012, 12:07:48 schrieb George Dunlap:> 2012/1/10 gavin <gbtux@126.com>: > > I noticed that the credit2 scheduling algorithm has one runqueue every L2 > > cache that is shared by cores in one socket. If vCPUs, which need to > > communicate more with each other, are assigned to the cores sharing the same > > L2 cache, the L2 cache miss ratio of Guest VM may be reduced. And also, the > > performance of the Guest VM will be improved. Is it right? But, how to test > > the cache miss ratio of L2 in Guest VM? Are the tools that used to test the > > cache miss ratio in non-virtualized environment can be used in Xen Guest VM? > > Looking through the code, it looks like performance counters are > passed-through to HVM guests, both on AMD and Intel. And it appears > that Fujitsu is regularly using the performance counters on Intel at > least (since Deitmar Hahn has been the one posting patches for new > performance counters). > > It appears that there is no support for access to performance counters > in PV mode, other than to use them for xenoprofile (which will not > allow you to read the counters directly). > > So it''s worth trying whatever method you normally use to look at > performance counters while running in an HVM guest.Yes, only HVM guests. On Intel (for what I sent patches) only some processors are supported. You have to add the hypervisor boot variable ''vpmu'' in your grub/menu.lst or whatever. If your processor is not supported you will see a message in the hypervisor boot logging (it''s printed from the sorce xen/arch/x86/hvm/vpmu.c function vpmu_initialise()). Dietmar.> -George-- Company details: http://ts.fujitsu.com/imprint.html