--- Carlo Marcelo Arenas Belon <[EMAIL PROTECTED]> wrote: > On Fri, Dec 22, 2006 at 08:05:02AM -0800, Martin Knoblauch wrote: > > Hi Folks, > > > > in order to fix bz#84 for Linux. > > http://bugzilla.ganglia.info/cgi-bin/bugzilla/show_bug.cgi?id=84 > > I think that the fix for this bug should actually include adding 2 > more > metrics, as the problem as stated isn't really that ganglia isn't > reporting > the right count of CPUs, but that there is no way to know if it is > virtual or > real CPUs for inventory and in some sort also scheduling reasons. > > This way "cpu_num" could be kept as the number of available CPUs, as > is > implicitly described to do in the current documentation for this > metric and > will have "cpu_cores" and "cpu_sockets" as the number of available > cores or > available sockets. > > of course for HPC, the number of "effective" CPUs is a function of > all those 3 > and the type of code that is being run, so we should leave up to the > end users > to figure that out while giving them all the information they need > for that. > > the advantages of doing it this way, are that the code is greatly > simplified, > all possible use cases are covered and the metric is kept backward > compatible. > > comments, anyone? > > Carlo Carlo,
modulo the naming of the new metrics, I completely agree with you. In order to make an educated "guess", we need all three components. And we should not forget that more and more clusters in use are running non-HPC workloads, where the virtual CPUs may actually be of use. One thing we should at least keep in mind is the fact the number of CPUs may no longer be a constant - CPU hotplugging is available on Linux and some of the proprietary Unixes. Same for memeory. And the cpu-frequency has been variable for years. Cheers Martin ------------------------------------------------------ Martin Knoblauch email: k n o b i AT knobisoft DOT de www: http://www.knobisoft.de
