On Fri, Mar 11, 2011 at 1:50 PM, Michael MacIsaac <mike...@us.ibm.com> wrote: > > Hello list, > >> We are activating cpuplugd process for dynamic CPU and memory >> management >> for Linux guests running in z/VM. We have found a reference in >> Virtualization cook book for SLES11 SP1 how to make necessary >> configuration >> for CMM modules within Linux.But couldn't find any reference about the >> configuration to be done within z/VM for CMM. Is any configuration >> required >> or z/VM comes with CMM enabled by default. > > I've been waiting for some performance person to chime in on what I perceive > as a disparity, but it has not happened. So I will ask the question - > Aren't we talking about two different technologies here - cpuplugd and > CMM/VMRM? Yes, the new Virtualization Cookbooks describe cpuplugd in the > section "Utilizing the cpuplugd service". This was based on presentations > from Hans-Joachim Picht, et al, but it does not involve CMM and VMRM. In > previous versions of the book there were sections on those, but it was > agreed that they should be removed - probably the references to loading the > cmm module at Linux boot time should also be removed (perhaps this was the > source of the confusion).
Yes, two different things. I've been telling people that you need to combine VM and Linux metrics to get it right. VMRM tries with just the VM data, cpuplugd tries with just the Linux data. IMHO the results confirm my claim... > So I believe the answer to the original question is - you don't need to > configure CMM and VMRM for the cpuplugd service to work. That section > should pretty much stand on it's own. You also need to distinguish the two different tuning parts (memory and CPU) which leaves these to consider: * VMRM for memory: uses just the VM memory metrics to tell CMM in Linux to use less memory when there is less available. While recent enhancements support setting a minimum to avoid killing Linux, it remains a "close your eyes and cross your fingers" approach. Even the believers suggest you don't use it for serious business workload. * cpuplugd for memory: apart from the examples being wrong (and maybe never tried) it is just too simplistic. The controls are either too aggressive or too soft. Typically cpuplugd is too slow in giving back resources. * cpuplugd for cpu: it's a bit like a solution looking for a problem, and to some extent it creates its own problem (it confuses the z/VM scheduler about the priority of the workload). The cases where it would make sense would be addressed better by proper configuration, and in several other scenarios it makes things worse. It would be more relevant to Linux in LPAR. > And please, could we have a bit less FUD without really digging into the > question? Thanks. I'm not sure what your request is... the devil is in the detail, as always. If it didn't even sound like a great idea to the casual observer, then we would not even have this discussion. I'm perfectly happy to explain, but can't do without going into the details. Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/
