On Mon, Feb 16, 2009 at 12:32 PM, Johansen.Klaus KYJ <[email protected]> wrote:
> What I suggest with cpuplugd is basically to let Linux (in your own > words) "behave social". I only looked at cpuplugd for dealing with virtual processors, and my comments about waste of energy applied to that. I would have to think deep about using it to drive cmm. Playing with the timed cmm release was not very productive. >> For some applications, using cmm_pages in a kind of scheduled >> way (eg squeeze during day shift and let air out for nightly >> backups) may work well as a manual compromise. > > If our current guests were running something newer than SLES9, I would > certainly consider to schedule a cron job to "echo 1 > > /proc/sys/vm/drop_caches". The nice thing with cpuplugd, on the other > hand, is that it can adjust dynamically during the day - it doesn't have > to be scheduled. Clearly you realize that drop_caches will not return the pages to VM, so you would still need to use cmm after that. When you free up memory with cmm it comes out of the same pool, so I'm not convinced there is an advantage to do both. > Generally I want to avoid my Linux guests swapping... Not even on VDISK, > which takes up memory better utilized as main storage. And I do guess, > that the LRU algorithms will help Linux some of the way, when it decides > which pages to keep in memory. It depends on the server utilization. As you say, high performance swap in VDISK is made up of the same bytes as virtual machine memory. The required total of high performance swap and Linux' real memory is determined by the application. The ratio real/total is determined by utilization. When the server is busy 80% of the time, you have less drive to share memory than when it is active only 5% of the time. > I guess that's somewhat CMM2 does... Unfortunately I have not figured > out how to test CMM2 in a larger scale in our environment. But I'm > certainly looking forwarded to try it out. Yes, that's why I felt the idea of CMMA was very cool. Unfortunately it lacks instrumentation and documenation, and the interaction with z/VM is missing. We'd need to see this beyond an artificial lab workload that was designed to show that it works. The value of CMMA is to share memory among virtual machines. So I was disappointed to see an IBM presentation on CMMA performance without any reference to z/VM storage management. When I commented that I was missing the VM storage data in the picture, the presenter wrote it down as a "good idea to look at" :-( Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/ ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
