RealTime IT News

Ravi Arimilli, IBM Fellow and Chief Architect

Ravi Aarimilli In 2000, IBM embarked on an ambitious strategy to boost Unix server market share after years of watching its sales suffer at the hands of rivals HP and Sun Microsystems.

The company trotted out its first Power4 machines in 2001 and promptly set about devouring market share from its rivals, watching its plot in the Unix land grow from 15 percent to 25 percent between 2000 and 2004, according to IDC.

But that wasn't enough. Big Blue continued to advance its processor roadmap, bringing it more closely into alignment with its company-wide initiative to provide computing resources on-demand. Customers have been asking vendors for such dynamic computing capabilities to help curb data center complexity.

The vendor unveiled Unix servers based on its new Power5 processors with great fanfare and competitive sniping more than two weeks ago, promising "jaw-dropping" performance.

The machines, available to the public August 31, were developed under the auspices of IBM Fellow and Chief Architect Ravi Arimilli. The engineer discussed the fruits of four years of labor on the Power5 with internetnews.com, as well as what IBM has in store for the future.

Q: You were given the task of making up lost ground in the Unix server market after years of being bested by HP and Sun. What is IBM doing to set itself apart?

We'll tie the technology and say things like 'Is there any way we can do a hybrid technology and add some silicon germanium into it?' We're always pushing value-adds. We were the first to come out with SOI [Silicon on Insulator]. My team also looks beyond the microprocessor core to other details. For example, my team had meetings with customers where we listened to them tell us how they needed Linux on Power to be better than Linux on Power4 was. We spent literally three or four weeks at different moments in our days with the Linux kernel team -- Linus Torvalds, Andrew Morton. They had a gun full of stuff that they wanted us to change and it was a great brainstorming exercise.

I want my team to be aware of how the bricks they are laying get used up here by the Linux operating system guys, by the middleware guys, by the WebSphere guys. I doubt the Sun or HP processor guys had meetings with Linus Torvalds. We have the ability to have meetings with these guys to improve the technology.

Q: Were these the meetings of the IT minds between IBM and customers?

Yes. On Power5, we really had time to use more technologies and sit back and talk to, for example, the CIO from American Express, the CIO from State Farm, the CIO from Home Depot. They'd get surprised because you'd walk in and ask them what they want to see in their IT farm. They're not used to anyone doing that, except for individual component vendors meeting them like a DB2 guy, or an Oracle guy.

We came in and hit them at a broad scope. The new model at IBM as a company is outbound -- we brainstorm with them. It used to be that the operating system would adapt to whatever the hardware is.

Linux has to run on so many architectures, and I learned at these meetings how difficult it is for an open source, multi-architecture system to be good. When you see that, you realize you have to help these guys do what they do better rather than them changing their design for you. So Power5 is heavily optimized towards Linux.

Q: For Power5, you and your team came up with a lot of features you are clearly proud of, including the micro-partitioning virtualization. VMware gets a lot of praise for what they've been able to do on the Intel side. But now you've been able to do something on the Unix side that no one has seen, such as run as many as 10 virtual servers on a processor.

Let's talk about VMware. If you look at what we've done on the hardware, we have truly decoupled the operating systems. But while we say decoupled, I think there is a misnomer here. For example, I can run z/OS on my laptop and I can run Linux and Windows. You can run multiple operating systems on one processor today on your laptop. That's really what I call the emulated model -- not native to the metal model.

So many people have done many emulated models where you can have three different Windows systems on a PC. If you look at what VMware does, they do more of that kind of thing, even though they talk about partitions per processor. They're not native on the hardware; they're emulated. What we did is actually become native on the hardware, so when you run a partition in this model vs. the direct-attached model, you lose no performance. That is not the case with VMware at all.

Q: So you're saying they're not true virtualization?

They're not true virtualization in the way that I think it should be defined. I have the same analogy for the many software vendors that make software that runs on top of your laptop that can run z/OS. We actually made microprocessor changes to allow many things to run native on the metal but are invisible to which metal they're on. And that's true virtualization. If you just peel it away, I can run 100 operating systems on one processor. That's not virtualization; that's just emulated fashion.

Q: What features in the Power5 systems reflect the company's broad strategy of on-demand?

The major thing I'd day that ties into that is the virtualization. One is the fact that we have micropartitioning. The other is the fact that we are virtualizing the whole system. One thing we haven't done, which is the next step to on-demand, is something that will appear on Power5+ next year: Let's say I buy six SMP machines, with 20 partitions per SMP. Well, it's great that those partitions residing within the SMP are fully virtual, with all of the I/O and memory and so forth. But the real idea of true, multi-system virtualization is making those partitions run anywhere. Not just on a physical SMP, but on any domain.

Now to do that, there are some obvious rules. All of the stuff that the SMP connects into, like the storage and network, must be twin-tailed to the other machine. It's multi-system virtualization, so active partitions migrate from one cluster to another. That ties into on-demand in my view. Because on-demand needs to do all of that stuff. We have done stepping stones to the pieces of it. But true on-demand can run any place on the Internet. Also, on-demand has the model of being network-based, which I think is its weakest link. That will go down once we get to multi-system virtualization where there will be more seamless migration. We're getting closer and closer. It's all coming. On-demand has got a story outside of IBM. But inside IBM, we are making it better with the Power5 story.

Q: Power5+ is up next on the menu. What can you do in Power6 and Power7 beyond what you've done in Power5, which seems to be a pole-vault over the Power4?

What we did on Power4 was a technology-driven design. On 5, we did technology but we also went outside the company. The model we have right now is what I call "real-time design changes into a roadmap of chips." We have a Power5+ in a laboratory running to be announced next year. It's like 200 square millimeters on a 90 nanometer design. It's a beautiful story, and the costs will be virtually nothing. But what's more important than that is that second pass of Power5+ is still open. We talk to customers, ISVs. We are putting in design changes in real time before the gate shuts for it. If we can't get features in before the gate shuts for Power5+, we'll get them in Power6. If we can't do that, we'll do it in Power7.

We really have a dynamic, real-time mechanism of listening outside and trying to squeeze it in where it makes sense. There are some things I learned from the life sciences community I couldn't quite get into Power5 -- they'll get into Power5+. We're just in the infancy stage of a huge stepping stone of engineering. Customers are becoming a part of our design team. They have a huge influence. If I were to show you all of the 30 features we added, you'd think each one in itself was trivial. That's not the hard thing. The hard thing was spending the time outside and creating the answer, which was Power5.