On Tue, Sep 15, 2009 at 11:18 PM, Robert J Brenneman <bren...@gmail.com> wrote:
> Admittedly - not 8TB in a 200G box, as Lee tried to do, and it was on > z/VM 5.1, so it didn't have the system execution space stuff of later > z/VM releases. It did teach the lesson that more page packs can only > get you so far. At some point the system data structures needed to > support the enormous guest just wont fit. This may be a reasonable > calculation to make within CP as a sanity check. If a factor of 2 does not make a difference, then try an order of magnitude. :-) One of the problems with booting Linux is that it determines the size of the virtual machine by testing pages rather than ask CP about it. If I remember right, it tries the first page of every architectured segment. And to make it worse, it uses a test that also forces CP to initialize the page frame. Which means that CP must also allocate a PGMBK to hold the page tables to span that segment. So for each MB of virtual machine storage, 3 pages must be allocated. When I get the math right, the 8 TB virtual machine will very quickly require 96 GB worth of page frames. That needs to come from somewhere... A decent paging subsystem can fill up a single 3390-3 in a minute or two. And although we tell people that you need to add one 3390-3 page pack for every GB of Linux server you define, there's still folks who think we talk nonsense because with the first few Linux guests their z/VM system did not page at all. But once you do start to page, page space utilization growth is not subtle. It's more like shifting your cup of coffee towards the edge of the table. Rob