Interview with Vijay Lund and Jim Herring

At Fall COMMON, September 18, 2005

Vijay Lund, vice president of server and storage development, IBM Systems and Technology Group

Jim Herring, director of iSeries product management and business operations

Vijay Lund: Let's begin with Power 6. Next year, we will introduce Power5+, which is Power5 in more advanced 90 nanometer technology. Not this year.

Now, people have asked me what is a nanometer. It's one billionth of a meter. So 90 nanometers is pretty small. So we're doing great in the development laboratories with Power 5+. It's just a matter of when we bring hardware, software, price, performance, yield, and cost all together. So some time next year. Power5+, I think that's doing great.

Now Power6 is a much longer term effort. We started this work right after Power4, and we generally kick off two generations at a time. We have the ability to do that. So this one is back from the factory, it's been built, of course that's the chip [handing it to us]. It's up and running in the laboratory and we're very happy the progress on Power6.

Now this chip is pretty impressive. We started to build on themes of Power4 and Power 5, this is Power6--three quarters of a billion transistors. That's a lot. I don't think anyone is attempting three quarters of a billion. And then someone from the crowd asks, that's a big number. Three quarters of a billion, using 65 nanometer technology. And I said, let me think about this, as to how big this number really is. If you bought 500 newspapers, and you counted every letter, through 500 newspaper, that would add up to about three quarters of a billion. And like I said, the transistor size is so small, so we can put lots of cache on this thing.

Timothy Prickett Morgan: What's 'lots' mean?

Vijay Lund: Lots is lots.

Timothy Prickett Morgan: How many stages are in the Power6 pipeline?

Vijay Lund: See, now you want me to talk about details about which I can't. We'll start to talk about this starting next year.

Timothy Prickett Morgan: As far as I can tell, you're changing the pipeline, it's different than Power5, and almost certainly shorter so you can crank up the clock speed a bit more.

Vijay Lund: Tim, don't jump that far. Everybody in the industry has given up frequency because its' too tough, you can't do it. Well, we don't quite think so. If you really look at this from a system architecture point of view, we can push frequency a lot further than anyone else in the industry.

So you see the big module, in the palm of your hand. We've got four chips on there, we've got cache chips on there. We've been looking at cost very seriously. How do you bring down the cost using small building blocks, how do you build using small building blocks. How do you build big computers using small building blocks.

The package is ceramic, of course, and the package is laid out the way it is [is had lots of little gold knobs, fatter and shorter than CPU pins, on one side and flat ceramic with capacitors on the other] because this is how the chip communicates. This actually sits on a package. This is IBM's C-4 technology, which we've always had from the days of the mainframe. It's very reliable, very robust. Others do other things, we do what we do best. So we built this package as a low cost building block. Now this will go into our low end and small systems. On the high end systems you'll see a different package, which I don't want to talk to you about, I don't want to pre-announce that. But that will have cache chips on it. But again, same concept, use a small building block. And we've found clever ways of how to communicate between chips, much like the days of Power4.

So lots to come, these little things you asked about are decoupling capacitors. Good designs have a fair amount of decoupling capacitors for noise and crosstalk. We're proud of this, this is good. We're doing well, we're working on challenges that come with a first pass design. We're working them through. So we're very happy. So I thought I'd bring this along with me and bring by some excitement.

Timothy Prickett Morgan: Has IBM made any statements with the general performance of Power6?

Vijay Lund: No.

Timothy Prickett Morgan: I have heard the jump to Power6 from Power5 will be bigger than the jump from Power4 to Power5.

Vijay Lund: It's going to have lots of performance. So we've always followed at least 100 percent between generations. 2X has been the minimum.

Timothy Prickett Morgan: Can you get the thermal envelope down from where you are?

Vijay Lund: You see, that's the issue. And that's what we're working very hard to do in the laboratory. I think we found the sweet spot. I think we found a secret sauce here, which lets us push frequency. There's a reason that some folks in the industry have stopped pushing frequency. Frequency is one component of performance, it's how fast the clock ticks, and how much work you do in a clock cycle. So if you're clock is slow, you better do a lot of work. If you're clock is fast and you do a lot of work, that's pretty damn good. So we're not letting up on frequency. We're going to push this thing. So it starts with the semiconductor technology. Transistors can give you a lot of performance, but they also burn a lot of power. So right there you started to work thermal things that help us as we grow these transistors up into our systems.

Timothy Prickett Morgan: How hard is it to get yield with something that has transistors that are so small?

Vijay Lund: Well, we do a lot of technology learning. The more product you push through the line, the more learning you get. It's just as simple as hard work and sweat.

Timothy Prickett Morgan: So, over time, you can get equivalent yields you're getting with 130 nanometer technology with 65 nanometer technology?

Vijay Lund: Yes, of course.

Timothy Prickett Morgan: I would just think there would be diminishing yields as things get smaller and smaller.

Vijay Lund: We haven't seen that as yet. Maybe those days are coming, but we haven't seen it.

Timothy Prickett Morgan: I though it might come with 65 nanometers because the gate size is getting so small. I mean, you're talking atoms here.

Vijay Lund: That's right, about four or five atoms. It takes a lot of learning, and that's why it takes so long to get products out the door. You don't want to get this thing out there and have it ending up being so expensive. For iSeries, we want to pick the sweet spot for cost, for quality--everything. So this is something we are very proud of.

Jim Herring: Think about the billions of dollars of investments it has taken to do that. You don't wake up over night thinking 'I know how to build that.' It takes a tremendous amount of R&D, soup to nuts from R&D to fabrication and the whole thing. For me as a product guy, that's what excites me. You're going to have leading edge technology in iSeries and pSeries for a long time coming.

Timothy Prickett Morgan: How much system is in that chip. For a long time, starting with Power4 and moving into Power5, you have been making what is effectively a system on a chip.

Vijay Lund: We've had lots of debates and discussion on what's next from Power5 to Power6.

Timothy Prickett Morgan: My quick math in my head says there's about 300 million transistors doing something else on there besides being Power cores. Maybe 250 million transistors.

Vijay Lund: Maybe it's 300 plus. Maybe it's 300 plus with lots of cache. It could be the L2. It doesn't have to be not the L2. You've got the L3 directories in here. You've got a fair amount of stuff in here. You've got lots of system function. Things that we have done in software that are performance sensitive we are putting back on the chip. I can't preannounce this to you, but there are system functions we'd rather do in hardware than in software. Execution related items, floating point related items. That's what we're going to do with this chip.

Timothy Prickett Morgan: What about Power5+? What is in it?

Vijay Lund: It's going to be a rev. A little bit of a frequency boost, of course, and we'll look at a few functions here in a quality perspective.

Jim Herring: And from an i5 lineup, we don't see a radical change in platform models. As I told you in Chicago, we are so pleased with the product line in terms of cost, function, quality--you name it. It's doing extremely well in the market. Profits have bumped up a bit.

Timothy Prickett Morgan: The thermal envelope must have come down a bit with Power5+, since you're moving to the new 90 nanometer process.

Vijay Lund: Yes. This is true.

Timothy Prickett Morgan: So is there a way to get a Power5+ onto a blade? Is it cool enough? That is what Intel's doing with Montecito--trying to get a dual-core Itanium down to a blade form factor.

Vijay Lund: We're having those debates in the laboratory. We're not ready to say anything definitive right now, but it's a possibility.

Timothy Prickett Morgan: So what other things do you work on?

Vijay Lund: Games, storage. Hardware is hardware.

Timothy Prickett Morgan: I meant in terms of iSeries technology. What are you working on, other than Power5+ and Power6?. What about different ways of connecting the systems. InfiniBand interconnect is still coming, right?

Vijay Lund: InfiniBand is coming, iSCSI is coming.

Jim Herring: We're also working on the integration between iSeries and xSeries. Those are probably the biggest connectivity things. We'll have new revs of RAID adapters, bigger faster RAID adapters. There will be RAID 6. So there will be more hardware things we're doing next year.

Vijay Lund: Better quality, lower cost. From a processor point of view you've got the engine that I described, and the engine is solid.

Timothy Prickett Morgan: And we're moving away from IOPs as well, Jim, right?

Jim Herring: Actually, here's what has happened. The same phenomenon that Vijay took you through on our main processor chip is the same as on our IO adapters as well. We are putting more and more function on them, to the point where our IOPs, which used to have a significant amount of function on them--from networking IOPs, data link control, all that kind of stuff--that has been outboarded, for the most part, out to IO adapters. So what IOPs have really become is a direct memory access (DMA) device, which automates data flow from main memory to the IOP then out to the IOA and then the peripherals attached to them. What we've figured out is that we can eliminate that DMA device--the IOP--and then DMA directly from main memory off to an IOA, with no loss in performance. So this wonderful asymmetric multiprocessor that we've build up for years and years, we still have it, it's just two tiers instead of instead of three tiers. And so we're excited about that from a product cost perspective. And also the other thing that's great about putting stuff on the chips, is that anytime you can take electrical components, and bring them very close together, the reliability goes up. So, again, if I get rid of a piece--an IOP in this case--we've just made the system better.

Timothy Prickett Morgan: Are you moving all of the workload overhead it out to the IOA or is some of it getting put back to the CPU?

Jim Herring: Very little gets put on the CPUs. We used to DMA from main memory into the IOP, and from the IOP over to the adapter. Now we're just taking that DMA and linking to the IOA.

Vijay Lund: This is expensive real estate [holding a Power6 chip]. This is Manhattan, so you've got to be very careful with putting overhead on it.

Timothy Prickett Morgan: That Power6 is Lower Manhattan. Trust me, Upper Manhattan is not so expensive. That's more like Cell.

[Laughter.]

Vijay Lund: So these things are along the lines of reliability. We're doing a fair amount of reliability work on this, system level reliability work. Instruction retires, those types of things. Pretty heavy stuff from our mainframe systems. We're starting to just import these things, like we've done before for Power4 and Power5. There is density on this chip, so we can do things. So those are things we're trying to work on. Like I said yesterday, it's reliability and cost. Performance--we've got plenty of performance.

Timothy Prickett Morgan: As far as iSeries shops go, I think they've got enough performance.

Vijay Lund: This chip is probably fine.

Jim Herring: We do, but I got to be honest with you, especially at the high end our product line, as hard as it is to believe that 165,000 CPW isn't enough for some people, at some point it won't be enough. We have a lot of large customers that are doing millions of transactions per day with their iSeries, so they don't ever want to run out of headroom, so this kind of development is very, very important to them.

Timothy Prickett Morgan: But I'm talking about low end machines, and these are more important for the rest of the iSeries customer base. It's great to have some huge customers that are big deals and great accounts doing millions of transactions. But most customers aren't like that. The pSeries has many more of those big boxes, I would guess. I don't know if it's an order of magnitude, or what, but it probably is.

Vijay Lund: We're going to build them out of these little things. There will be another module that has L3 chips on it, just one processor. So we've learned how to do that. Where we can get the interconnecting between these chips like there was no loss in performance. We've learned how to do that.

Timothy Prickett Morgan: So how do you do it?

Vijay Lund: We've been very clever at this. We've done Power4 in a very clever way, and Power4 was my baby. And then Power5, and this, too.

Timothy Prickett Morgan: Power4 was clever.

Vijay Lund: Our chip to chip interconnect was 1GHz, when the rest of the industry was practicing 200 MHz, 300 MHz. We were up there 1 GHz between chips. We're just going to keep going.

Timothy Prickett Morgan: Does the architecture of the machine change with Power6? I mean you have all this sophisticated interconnect, and then all kinds of backplane . . .

Vijay Lund: Great question. The system architecture, Power, is fabulous. It's got lots of room, lots of room. I don't know how Intel concluded that their X86 was out of gas ten years ago. I think that was a blunder on their part. They concluded that, because ten years later, they say IA-32 can be extended with EMT-64. So the Power architecture is pretty sound. It's solid. We've got lots of headroom. We seen that with the 390 architecture, too. Our 390 architecture is solid. The microarchitecture, we keep refining that, inside the chip. That's where you look at the pipeline depth and you look at things that matter. But architecturally, this is solid.

Timothy Prickett Morgan: Do you see a need to go beyond 64 cores?

Vijay Lund: We think we better provide enough headroom so that we never have this issue of customer consolidation ever a problem. If customers want to consolidate they can. We want to build big things out of small things. You just keep stacking as many small things as you want.

Timothy Prickett Morgan: But the single system image size. There's no reason to get much larger than we are now, is there?

Vijay Lund: The largest SMP today in the world, I think, is a 128-way.

Timothy Prickett Morgan: But Fujitsu is not going to 128-way in the next generation of their servers. There are indicating that they are stepping back to 64-sockets as they double the cores.

Vijay Lund: These are decisions that we have to toy with.

Jim Herring: Exactly.

Timothy Prickett Morgan: It gets increasingly expensive and difficult to glue all that together to make a big system.

Jim Herring: But I think what's important, and Vijay's team has done a great job of that, is not precluding it in our system architecture.

Vijay Lund: You can do it if you want. You can go build these things out. So we've got the hooks in here to go do those things. We may chose not to build these systems.

Alex Woodie: What's the multicore story with the Power6?

Vijay Lund: By his math, there's 300 million transistors are doing something else, and all I said is 300 plus. So , it's pretty good stuff. This is good stuff.

Timothy Prickett Morgan: Everybody's intrigued about what it is inside Power6. I'm not convinced you can do four cores so easily, or 8 or 16 or whatever like some chip makers are talking about.

Jim Herring: Irrespective of how many cores we do, or how many transistors, whatever technology, the important thing is that those RPG applications we talked about in the Town Hall yesterday, they're going to run unchanged on whatever we build. Fundamentally, our commitment to application investment is just as strong on Power6 as it was on Power5 as it was on Power4. So you're right, some of this stuff is hard. Vijay's guys are incredible at solving hard problems.

Vijay Lund: You put your finger on something that's very, very profound, let me tell you, Tim. It's like in a car, when you watch the speedometer. You can't go 60 without clocking 40. You can't just go to 60 or 80, without clocking 20, 40, 60. You've got to learn two cores before you learn four cores. We've been at work on this since 2001. There are many issues, and we've got to go deal with these things. And when you learn how to deal with two, then you go to four. You don't just start with eight, or four.

Timothy Prickett Morgan: So Power6 is early 2007 maybe?

Vijay Lund: In 2007. There's a fair amount of stuff before then. We've got these game chips, too.

Timothy Prickett Morgan: But I can't run OS/400 on a game chip. We can talk about it all day long. Maybe we can, and you just won't let us, but that's a separate issue.

Vijay Lund: These are pretty impressive chips, too.

Timothy Prickett Morgan: No question about it, and I'm impressed with Cell. Which brings me to the next point: Are you ever going to put Cell on a server? And if so, what on earth would it be doing? I can see it for visualization, other types of co-processors.

Vijay Lund: Remember, I told you about microarchitecture. There are microarchitectural features, thoughts, learning--they can be applied. Right away. You know these special purpose engines? If you look at a Cell chip, it has eight special purpose engines. Accelerators. That's pretty good. I view Cell as an accelerator function. It's a booster function.

Timothy Prickett Morgan: What can it boost?

Vijay Lund: You look at the application, and you say, you need a specialized engine for this, you need a specialized engine for that. And as long as you can talk directly to the metal, you can do almost anything. Just like we learned with Cell. So we're learning from Cell. Cell pretty much does a lot of graphical rendering types of applications, and we need these eight special purpose engines to go after these things. We're looking for applications, too, in the server space.

Timothy Prickett Morgan: On that chip literally, or using that style of chip design?

Vijay Lund: The design style. That's how it starts. You can't start broad brush. You've got to start with the principles of how these applications of tomorrow--in the case of Cell, these gaming applications--can exploit these special purpose engines. And we're learning quite a bit from that. So I see the principles of that being applied.

Timothy Prickett Morgan: Are these Cell principles in the Power6 that chip, or are they in Power chips after Power6?

Vijay Lund: Let's just say we'll talk at the next COMMON. That's a good question though. We've spend a fair amount of time on this. Power5 was pretty good too. It's pretty cool.

Timothy Prickett Morgan: I'm not upset with Power5. It is a very good processor.

Vijay Lund: And now we have Power6+ kicking in the laboratories. And we are getting going with the next thing beyond that. It's a fair amount of work.

Timothy Prickett Morgan: When did you start Power7?

Vijay Lund: Traditionally, whenever we bring up the next generation of technology, we start the next one. So it's been a few months that we have been thinking about Power7. These things are five year programs. So you want to start now if you are talking about new chips for 2010. These are just early thoughts that we start to sketch. They're on the whiteboard.

Timothy Prickett Morgan: And I am guessing that Power7 probably looks like some things that are coming out of Austin, like the Project TRIPS morphing processor.

Vijay Lund: We are taking ideas from Austin, Poughkeepsie, Rochester. We're tying them all together. Rochester has got a fair amount of learning by making Blue Gene. IBM has a fair amount of learning with Cell, and with Microsoft's Xbox 360.

Timothy Prickett Morgan: But you can't really use those parts in servers.

Vijay Lund: Not as it is, just yet. We've got to learn the principles first.

Timothy Prickett Morgan: I still think Cell would look like a very interesting workstation.

Vijay Lund: Cell could be an interesting blade. It's finding its way into many applications today.

Timothy Prickett Morgan: It would make a hell of a graphics processor.

Vijay Lund: That is for sure. We're seeing that for sure.

Alex Woodie: Any chance iSeries will get a native GUI?

Jim Herring: I don't think so. We're a business system that you attach clients to. We are looking at ways to decrease the cost, decrease the price, of connecting up consoles to system managers. But a native graphics attach doesn't make much sense to us when you can buy a $200 or $300 Linux terminal to attach and run your system off.

Timothy Prickett Morgan: So what else is cooking for the iSeries, Jim?

Jim Herring: We've got a lot of software next year. We'll be delivering the next release of the operating system, i5/OS, in the first half of next year. It's got oodles of cool stuff. We're primarily focused on the principles you heard yesterday: Commitment to openness, collaboration, innovation. We're trying everything. Our operating system has in the past delivered on those things. We're going to continue to do so.

We're going to continue to enhance our relational database. There are a couple of key areas there, one is in terms of making our support for SQL more user friendly as well as more application friendly, supporting more interfaces. We're going to supporting our on-demand performance center for the database, so that even if you're not a database administrator, you can still look at how SQL and your database performance is with a graphical shot of your system. Today, there are really good drill down capabilities in iSeries Navigator, and we want abstract that up a level or two so that your basic system administrator can understand how the database is doing.

Timothy Prickett Morgan: And that on demand service is not a service, it's a tool that in the system?

Jim Herring: Yes, it is part of the system. The other thing we're doing for openness, for example, is we're taking RPG and allowing you to go directly to Web services. So you'll be able to take RPG business logic, wrap it in Web services, and tie it to other applications, to other systems, tie it to partner systems, or whatever. We're also having a Web performance problem determination function built into the system for our Rational Observer product.

Timothy Prickett Morgan: How are you going to be doing Web services? Is it an XML wrapper?

Jim Herring: It's a direct connection. And I honestly forgot the name. It's an Apache function is all I remember. [Editor's note: this sounds like the Project Xerces XML parser.] So there's a way to tie a piece of application, if you will, to Web services, and we're using that for all the ILE languages, in fact, not just RPG.

Also, with V5R4, we're making it easier for people to program SQL inside RPG. And that has to do with a lot of enhancements that we've made with SQL on this platform for all of Version 5. It's pretty significant.

I mentioned iSCSI before. We'll use it to attach xSeries servers and BladeCenter directly to iSeries. It's basically an Ethernet cable and an Ethernet network versus having a proprietary xSeries adapter and High Speed Link.

Alex Woodie: So you'll be able to get rid of the Integrated xSeries Adapter?

Jim Herring: Yes, we'll slowly migrate off that. We'll make it available another year or so.

Timothy Prickett Morgan: Will this replace the Integrated xSeries Server co-processor as well?

Jim Herring: It will, as a matter of fact. We'll support it for some amount of time, but it will be eliminated.

Timothy Prickett Morgan: The IxS had limitations, too, in terms of connectivity and performance. It was also one more thing you had to make.

Jim Herring: Yes, it was. Because of the form factor and everything else, it can only reach a 2 GHz processor now. So with iSCSI, we can expand a bit.

Here's another thing that's a notable change in our installed base. Five years ago, we went full-on with tower-based systems, and every system we shipped was a tower. If you wanted a rack-mounted server, you had to do unnatural things to it. Now we're seeing the shift exactly in the opposite way. Almost all of our 550s are rack mounted, and a large majority of our 520s are, too--not so much the Express versions. So one of the interesting things we're going to be shipping an integrated disk drawer next year that is rack mounted, with a much better density. We'll go from like 12 to 14 disks up to 24.

Timothy Prickett Morgan: Will they be SAS drives?.

Jim Herring: Not next year. We have it on the road map, but we're not doing it next year.

Timothy Prickett Morgan: Are they too expensive?

Jim Herring: No. We're just very conservative on iSeries when it comes to disk technology. Because of single-level store and its importance to the reliability of the system, we're pretty careful about it.

Timothy Prickett Morgan: You just want someone else to burn them in first?

Jim Herring: That's a good way to put it, Tim. I usually don't use disk and burn in the same sentence. But, I want someone else to be on the bleeding edge of that technology, not us.

Probably the biggest area we're doing things in is around availability, not so much unplanned downtime, because we're pretty good at that already, but planned downtime. So we're doing things like virtual save sys, so that you can spool off your save image off the disk, then spool to tape, so you don't have to acquiesce your entire system for a long period of time.

Timothy Prickett Morgan: How much faster is it to do that?

Jim Herring: It's very similar to virtual media installed today. It's very quick to install an operating system release versus doing it off the CD.

Alex Woodie: So that will be a virtual tape offering?

Jim Herring: Yes, it will be part of the operating system. So instead of spooling it off the tape drive, the application, or the system, will just see a virtual tape drive.

We'll also be able to do a native save and restore of spool files. This is an important thing because a lot of customers do their compliance around files. So now you're going to be able to do save and restore of spool files and you've got the compliance center to show that I've got this spool out there, to tape or optical or whatever.

Timothy Prickett Morgan: Did you go to a third party to add that in or did you write it yourselves?

Jim Herring: We wrote it ourselves.

You'll see new versions of everything from Domino to WebSphere, just like always. We continue to develop and continue those products. We've also been doing a lot around service-oriented architecture, a lot of the products both inside of i5/OS, like WebSphere Express, and outside of i5/OS, will support that functionality. That's one of the big reason that we wanted to get this RPG to Web services thing done so it's in time for our focus on service-oriented architecture.

One other thing. We've had an integrated Java Virtual Machine since 1998, with V4R2. We're very proud of it in terms of its scalability, robustness, all that sort of that. We still hold the record as far as I know as far as pure Java transactions on the platform. But that virtual machine didn't scale down very well on our smaller systems, no matter how much power Vijay put into them, they really did not perform as well at the low end as we wanted them to. So we are going to implement, in addition to our integrated JVM, another Java virtual machine with a smaller footprint. It's the IBM standard JVM. We're really excited about what this is going to mean. So it's the same one running on Linux and AIX, and it's got a smaller footprint, takes up less memory, and performs very well on low end machines.

Timothy Prickett Morgan: What's the difference? I thought it was about 500 CPW to be in the game with JVM?

Jim Herring: To get in the game. To play it well, you need 2,000 to 3,000 CPW, and especially for a lot of our existing customers, that was too much.

Timothy Prickett Morgan: And where will this new one put you in terms of CPW?

Jim Herring: That's a good question.

Timothy Prickett Morgan: I know. That's why I asked it.

[Laughter.]

Jim Herring: We think about half as much. It won't be CPW as much as will it be memory footprint, that's what really got us.

Timothy Prickett Morgan: What was the memory footprint before? Or did you have to be at 2 GB?

Jim Herring: You had to have at least 1 GB dedicated to it, and a lot of our customers are running on 1GB machines, period. I don't have the specifics on the exact number, but I do know it's smaller. So that's exciting. And we'll continue to support both for some amount of time to make sure we've got compatibility from release to release.

Timothy Prickett Morgan: Will that be standard across the line?

Jim Herring: Initially we'll do the integrated one that we have today, that will be the default JVM for V5R4. With the next operating system release, we may choose, depending on how this thing rolls out, to make the new one the default. And then at some point, probably five years from now, we'll sunset the old one.

Timothy Prickett Morgan: That's even on the big boxes?

Jim Herring: Yes. We've had our hands on it for maybe six months and we're just getting it integrated into the OS.

Alex Woodie: Have you considered adding an extra processor just for Java?

Jim Herring: Well, we sort of have that today. We sell our bigger systems with sort of your choice of how you want to run the processors, whether you want to run i5/OS Enterprise Edition or i5/OS Standard Edition or no i5/OS at all, where you can run Linux or you can run AIX--and you can do whatever you want with those Linux and AIX processors. So we have that concept that zSeries does for that and Integrated Facility for Linux today on larger systems. What would be interesting to do is to explore that on a smaller scale, to take our micropartitioning technology, and that's something that we're investigating.

Timothy Prickett Morgan: Can you increase the number of micropartitions as you are ramping up the processor speed?

Jim Herring: Any way you want to. Today we've got ten micropartitions. There's not an architectural limit. It's really a matter of processor. How much processor do you need to run each system instance.

Timothy Prickett Morgan: For a lot of server consolidation workloads, carving up a 64-way into 254 partitions, each individual partition is still a little too big.

Jim Herring: We support 254 partitions on our 64-way, and that's a lot. That's four 750 CPW machines per processor.

Timothy Prickett Morgan: Yes, I'm saying for a lot of workloads, you need 100 CPW. That's my point.

Jim Herring: So that's the beauty of i5/OS, right? Because we have subsystems to be able to manage that as well. I guess if there's no architectural limit, I could see the marketplace saying we need to have 100 CPW partitions, then we'll figure out how to do it. I don't have that on the roadmap right now.

Timothy Prickett Morgan: Nobody else is scaling down that far, carving up that many.

Jim Herring: zSeries has a lot.

Timothy Prickett Morgan: Again, it's theoretically unlimited, but the practicality is a couple hundred per server. But I don't know if customers are really using it that way. Do they really put hundreds of Linux partitions on the mainframe.

Vijay Lund: Even on zSeries, the exploitation is not as rapid as we'd like.

Timothy Prickett Morgan: I just think, customers who have zillions of servers, all running at 5 percent, that they would consolidate like crazy.

Jim Herring: But you have to remember, especially for iSeries customers, the workloads typically run a lot higher than 5 percent on average. So even though that's the industry average, that's not the iSeries average. The plumbing in our hardware and our firmware is so important, while it's 5 percent average, it peaks at 80 percent. So you need to be able to manage that flow, that dynamic flow, and that's something that's implicit in Power 5 and Power6as well as our firmware.