On Wed, Feb 11, 2009 at 6:11 PM, Bill Holder <[email protected]> wrote:
> That's not necessarily true. When a Linux task completes, Linux does not > necessarily inform CP that the link between Linux virtual and Linux real is > broken and that any saved contents of the Linux real page can be discarded. Yes, I believe it is rather risky to make such statements only based on reading the glossies and filling in the blanks. It works better to do measurements to confirm the assumptions. For the average Linux server, most of the virtual machine is "LRU managed memory" (either page cache, JVM heap, DB buffers, SGA, etc). Once you start paging in z/VM, then *all* that memory will ultimately land on z/VM paging space. This is why we say that typically for each GB of Linux virtual memory and in-use swap VDISK, you must add 2 GB of z/VM paging space. Only exception is when you don't page ever in z/VM (only very few shops can afford that). And in some cases it can even be worse. Because the paging cleanup in z/VM is not immediate, it may take a while before the slot is freed. We've even seen a virtual machine take twice as much pages on DASD than it has paged out... > Even with CMM or CMMA enabled, Linux is not obligated to inform CP of every > page transition, and even when it does, CP may not "notice" a page is in a > state where old contents (backed on CP paging DASD) can be discarded before > Linux reuses it for some new mapping. As I stated above, you need measurements to confirm assumptions before you share them. Since there is no instrumentation for CMMA, I am reluctant to comment :-) But when my guess is as good as yours, I think it is very likely that for Linux "LRU managed memory" CP would not have a clue when the page went through a cycle of in-use, free, in-use. So CP will continue to hold onto the previously paged out content (until we page it out again). Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/
