On Tue, Feb 15, 2005 at 06:38:09PM -0700, Ian Cunningham wrote: > Brooks Davis wrote: > > >On Tue, Feb 15, 2005 at 05:27:53PM -0500, Robert E. Parrott wrote: > > > > > >>Is there a way to make ganglia hyperthread - aware in v3.0? > >> > >>We have a cluster of dual Xeons with hyperthreading, which ganglia > >>reports as 4 processes. This is annoying when using the web frontend, > >>since full load appears to only be 1/2 load, and people think there are > >>twice as many CPUs available. > >> > >>With 2.5.7, I hacked the code using a patch from OSC to only report 2 > >>processor per real processor when hyperthreading was found enabled. > >> > >>However, in 3.0 this is not enabled, so I've backed down to 2.5.7 again. > >> > >>Is there a reason this kind of patch is not in place? Are there options > >>I;m not aware of in the config files? > >> > >> > > > >There is no sensible way to do this that will always work. The > >problem is that some users will want the current behavior and other > >will certainly not. It depends on their applications. A small, but > >non-zero set of applications exists for which HTT really is nearly as > >good as two CPUs. The real issue is that a number of cpus variable is > >no where near sufficient to represent the issues involved in representing > >the hierarchy of CPU like things on your machine. It's actually the > >case that as long as the FSB speeds match, there's no reason why your > >CPUs even need to be running at the same internal clock rate, on the > >current x86 architecture. > > > >I think a configuration option for gmond to divide the number of CPUs by > >something before reporting would be a decent, low effort method of > >allowing people to report what they want. > > > I have not looked at the method ganglia uses to gather the cpu count on > linux, but I do know that on 2.4.21 linux, /proc/cpuinfo shows > hyperthreaded cpus as sharing the same physical id and runqueue numbers. > I agree with Brooks that this should be configurable, but I do not > recommend dividing number of cpus, rather count the unique physical ids.
Not all OSes provide the ability to show this information. Hence my suggestion to simply allow people to divide. An option to only show physical CPUs could also be useful, but would not be implementable in all cases. -- Brooks -- Any statement of the form "X is the one, true Y" is FALSE. PGP fingerprint 655D 519C 26A7 82E7 2529 9BF0 5D8E 8BE9 F238 1AD4
pgpOeEN17QTIS.pgp
Description: PGP signature
