roman, you understand incorrectly 🙂
yes, you and sever are right, that every task can be moved to PHYP, and it can do the same things, as VIO. But in this case PHYP will shortly be as VMware — the same bloatware. And more — we will receive the same problems with performance, as VMware has. As far as I remember, there are no VMware benchmarks, e.g. for TPC-C, because VMware forbids the publications of these benchmarks. Do you know, why? I suppose, you know 🙂
The 2nd problem with bloat hypervisor is resources. As for now PHYP is executed on its own CPU (PPC604e afair) and has its own flash memory. It doesn’t use server’s CPU and storage, because they are much more important for business applications, as for virtualization.
Having such small PHYP we receive a very good performance in micropartitioning and other techonologies, which don’t require I/O (mostly storage) virtualization.
But I/O virtualization is also important. Because of that we have VIOS. VIOS is a sort of appliance, which does only I/O virtualization. It per se can’t virtualize your CPUs or RAM. While I/O virtualization requires usually more resources as a small hypervisor, VIOS was moved to a normal userspace — we need to create an additional LPAR, which uses our RAM, CPUs, storage. It has such advantage, that we use «normal» Power CPUs, and I/O virtualization works with good (suitable) performance. Because we have a «normal» LPAR, we can install two VIOSes and make a cluster of them, which makes better availability. If we’d have all of these functions only in PHYP, we could not make a cluster — we’d be required to buy a second PHYP, but it can only be done for Hi-End models. In case of I/O virtualization on the level of LPAR you can implement HA even on Power710.
Now back to the basics 🙂 Why NPIV and IVE are not our bright future.
NPIV is a very good technology, but everything has its own problems. It seems, that you never worked with NPIV in real BIG enterprise environment. If you have 50 LPARs on the host, you need to make 50 vfchosts for storage and 50 vfchosts for tapes on one VIO. And the same on the second VIO. It is quite easy to make them and arrange to the client LPARs. No problem. The problem begins with client LPARs — you’ll receive a lot of performance problems and lot of strange problems with FC. Do you know how much memory do you need for this configuration? I don’t want to look at the Infocenter — but you can do it for me. I will say only, that it will be more than 32GB per VIO afair. The official answer of IBM — «Memory is cheap, buy more». Now compare this to the requirements of VSCSI. It’s a homework for this WE 🙂
The 2nd problem with NPIV, which I was pointed from one of the readers of my blog, is that we need to install on all of these 50 LPARs storage drivers. If we are for simplicity (and against complexity) we’d rather install all of these drivers on one (two) VIO(s) and manage all problems with them in one (two) point(s).
As for IVE — it’s the worst invention of IBM. Please read once a redbook on IVE and you understand, why is it so. I never recommended to use IVE to any of my clients, and as far as I can see, nobody uses it.
What awaits us in the future is VIOS NextGen with its shared storage pools. As for now it’s the only «next step» in the IBM virtualization. The 2nd big step is SDMC — it has nothing common with VIO and PHYP (as for now), but nevertheless it is our future.
Sorry for english — I don’t have russian keyboard on my job and can’t reconfigure my PC 🙁