On Thu, Jun 4, 2009 at 10:44 PM, Hans-Joachim Picht <h...@linux.vnet.ibm.com> wrote:
> The example above illustrates the syntactic format of a memplug and menunplug > rule. These two variables must be adjusted depending on the usage and workload > of your Linux on System z installation. No general or all purpose example > configuration can be provided as this does not provide a useful setup for > production > systems. Hans, It's early for my monthly arm wrestling with a kernel developer, but anyway ;-) I consider it poor taste to illustrate the syntactic format by a single example that is semantically incorrect. If you actually did experiments, put one of those in the book. Even though they may not fit each system, they would at least fit one system. But that did not scare me away from it. Even though I am very skeptical about tuning with only half of the performance data. Unfortunately you could not attend my presentation. What I explained to the audience is that practical purpose, the rules could be rewritten as MEMPLUG = "swaprate > 20" MEMUNPLUG = "swaprate > 10000" The 2nd rules says that you should remove memory when the system swaps with more than 10 MB/s. That's heavy, but doable with VDISK. But to *remove* memory in that case seems rather silly. So I assume your intention was to remove memory when the system is *not* swapping. So MEMUNPLUG = "swaprate < 20" This conforms to the Principle of Least Astonishment. Just like we have been telling people how to size their Linux virtual machines: "when it swaps more than a little bit, make it larger. when it does not swap, make it smaller." So I guess I showed that one *can* come up with a generic rules that make sense for all systems. And if you had implemented the MIN and MAX slightly different, they would have allowed for a generic example as well. For my experiments I created a bit more hysteresis, otherwise you're wasting resources just by constant breathing of the balloon. My conclusion is that the controls available in cpuplugd are insufficient to improve memory management for Linux on z/VM. And for the experiments I did, it made performance worse: increased CPU usage and caused additional latency. Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/ ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to lists...@vm.marist.edu with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
