On Fri, Feb 20, 2009 at 12:51 PM, Michael MacIsaac <mike...@us.ibm.com> wrote: > >> Last time I tried 3rd level guests it was fairly painful not sure what >> subsystem to blame.
> Not that I'm a z/VM guru by any means, but ... I have found the peformance > of 3rd level Linux systems to be anywhere from perfectly acceptable to > lethargic/downright annoying. When I'm on a system that is approaching > annoying I will often do a "top". Now that the steal percentage (%st) is > reported, the values can be very telling. I will see steal rates in the > 70-85% range. As I understand it from people who are z/VM gurus, this > happens when the system is doing a lot of I/O and the third level guest has > to go in and out of SIE constantly. On second level Linux systems, the > steal percentage usually maxes out in the mid-single-digits. Nor am I. The trick with V/SIE hardware support is that popular scenario's can be done with a "virtual SIE exception" and thus remain entirely within (the outer) SIE. When it does not, you get a real SIE intercept and outer CP gets the overhead. I plan to attend session 9273 at SHARE, that might have some data about this :-) I question the idea to blame I/O. After all, CP is already involved with I/O. And we see that doing CMS is not bad at all in a 2nd level z/VM in LPAR. The performance problem shows when you do Linux for example - I am looking at data where CP overhead is 2 seconds for each second of virtual time. When you play with virtual addressing beyond the 2 supported layers in SIE, first level CP ends up doing the old game of shadow tables etc, like what you had in the early days with VM/SP. Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/
