> Ok, I'll ask. Why wouldn't one attach an OSA card directly to a Linux > guest?
Ties a guest to a particular piece of hardware (failure point), and forces the guest to handle all the recovery, ARP management, etc. Having CP do it for multiple guests is a much more resource efficient approach. It also ultimately uses up a lot more OSA port triplets, which aren't free. A couple 10G cards aren't cheap, but they're cheaper than multiple 1G cards, and you'll get better utilization out of the 10G cards. You also get a better chance to move any network processing outside the Z box -- 10G cards in switches tend to have LOTS of horsepower, and they're way, WAY cheaper than an IFL. > Seriously, why shouldn't this be done? Consumes lots of memory in each guest (16M per OSA is the default, I think). Also forces the guest to be dispatched more often to handle all the I/O (particularly noticeable when using layer 2 where the guest has to handle ARP and other stuff). It's also a lot more management-intensive to have to figure out all that port mapping in the first place, and then figure it out again in a DR situation where the hardware is different. With VSWITCH, you change it in one place and it's done for all the guests on that VSWITCH whether you have 1 or 100.