The AS/400’s Grandfather Talks Past, Present, and Future
June 23, 2008 Timothy Prickett Morgan
When Frank Soltis, the chief scientist for the i platform, is introduced these says as the Father of the AS/400, he sometimes likes to point out that this is not exactly a correct title for him. Technically speaking, Soltis says, he is the grandfather of the AS/400, and in fact is a man who is celebrating many anniversaries relating to the computing platform this year–not just the 20th birthday of the AS/400, which happened on Saturday but which IBM is celebrating throughout the month of June around the world.
As I told you would happen in my own homage to the AS/400 platform last week (see Happy 20th Birthday, AS/400!), Soltis gave a speech he is fond of giving about the past, present, and future of the AS/400 platform and its predecessors, which he had a key hand in designing and has championed within Big Blue for four decades. Tango/04 Computer Group, a maker of system monitoring tools for the i platform, and Midland Information Systems, a reseller of Power Systems and solutions for these boxes, jointly hosted an event that gave Soltis his soapbox last Thursday morning, and more than 1,000 people showed up to listen. This is a pretty remarkable turnout for a Webcast, and there was no way that Soltis could field more than a few of the (I kid you not) more than 1,000 questions that he apparently received during the Webcast. This is as close to rock star status as you can get in the AS/400 world.
And for good reason. As Soltis explained, in October, he will be celebrating the 30th anniversary of the introduction of the System/38, the machine code-named “Pacific” that embodied many of the ideas that sprang forth from Soltis and other IBM researchers as well as cut against the grain of other ideas about business computing from IT academics, other vendors, and sometimes even Big Blue’s other systems divisions. But that is not the only anniversary that Soltis is celebrating. Come November, it will have been 40 years since Soltis moved to Rochester, Minnesota, after getting his PhD in computer engineering from Iowa State University. This was an important move, because in 1968 IBM was looking to create the technology that would be deployed in the kickers to the System/3 minicomputer, and that set Soltis on his way cooking up the technologies that would eventually be used in the Pacific system and would be extended further with its “Silverlake” kicker a decade later, which we know as the AS/400, of course. IBM has changed the names of the Power-based machines that run OS/400, i5/OS, and now the i platform over the years and changed its technology underpinnings sufficiently that you might even consider Soltis the great-great-great-great grandfather of the Power Systems running i. (That’s seven generations in 40 years: System/38, AS/400, AS/400e, eServer iSeries, System i5, System i, and Power Systems running i.) Having just recently become a great uncle at the ripe age of 43 myself, I am sure Soltis doesn’t want to think about it this way. . . . but it is true nonetheless. And you can add another “great” in there if you want to account for the jump from CISC-based AS/400s to PowerPC-based AS/400s in the summer of 1995.
Soltis said that since April, when IBM finally and completely converged the System i and System p server platforms into a single Power Systems line supporting i, AIX, and Linux operating systems, he and his fellow IBM colleagues have been traveling the globe, speaking to and listening to i customers and partners about the convergence. And he said in the Webcast that the reaction that IBM is getting is very similar to the one that the company had two decades ago when the Silverlake platform was launched. Soltis said that the high-end customers in 2008 are “extremely happy” with the convergence, that they saw it coming and to my mind, they compelled IBM rather strongly to make the changes so there would be absolute parity on hardware pricing and consistency in terms and conditions on hardware, software, and support between i and AIX platforms.
Not everyone feels that way, of course, particularly small and medium businesses that use OS/400 and i5/OS platforms to run their businesses. “Some of these are a bit unhappy, not about the product, but more about the feeling that they have lost something,” explained Soltis. Of course, this was parallel to the reaction that the AS/400 received. System/38 shops, who needed more computing power and who wanted to preserve their applications, understood and were enthusiastic about the AS/400 launch–they not only got it, but they got what they wanted. “I can’t say the same thing about System/36 customers,” said Soltis with a laugh, many of whom were irate because of performance issues relating to RPG II applications and because they had to learn a substantially different operating system–OS/400 was far more CPF than it was SSP. And to that end, many System/36 shops didn’t buy AS/400s for a few years, particularly since new and used System/3X gear was widely available for competitive prices.
One of the secrets to the longevity of the AS/400 platform, no matter what you call it, has been its ability to absorb new technologies as they come along without compelling customers to make changes to their applications. (This is the technology-independent machine interface, or TIMI, that Soltis is famous for in the System/38 design, as is the single-level storage memory architecture used by the System/38 and its kickers.) Only three times in the history of the IBM midrange line has the TIMI changed sufficiently to have customers recompile their underlying code for the new hardware, something that happens transparently and automatically, unlike every other platform ever created. Anyway, this under-the-covers recompilation happened first in the jump from the System/38 to the AS/400 in 1988, once again in the jump from CISC AS/400s to PowerPC AS/400s in 1995, and finally this year with the move to i 6.1. Not only does this interface allow IBM to change the underlying hardware when it sees fit, but it also allows IBM to plug in other computing environments.
The AS/400 had three complete runtime environments managed by OS/400: System/38, System/36, and native AS/400. Over time, IBM has added the PASE AIX runtime, Java virtual machines, and more recently PHP engines. Plenty of system programs, like the TCP/IP stack and PHP, run in PASE. So does SAP‘s i variant of its All in One ERP suit, which I just learned from Soltis last week. The flexibility of the AS/400 design has allowed IBM to add new capabilities, and many other vendors of proprietary minicomputers no longer exist today. (Pretty much, it is down to System z BC machines, the i platform, and some Hewlett-Packard Integrity machines running OpenVMS on Itanium.)
So what does the future of the AS/400 hold? Even Soltis does not know. “I don’t know anybody who can predict out to 20 years–I certainly can’t,” said Soltis on the Webcast. “But I can look out a few years.” And so, he did, although when he looked out a few years, he actually just looked at the new 1 petaflops supercomputer made by IBM for the U.S. government’s Los Alamos National Laboratory dubbed “Roadrunner.” Roadrunner is a hybrid processor massively parallel supercomputer based on a blade server form factor and it is the first machine to surpass the 1,000 teraflops performance mark on the Linkpack Fortran benchmark test; it went live two weeks ago and took the top spot on the Top 500 supercomputer rankings last week. (See Beep, Beep: Roadrunner Linux Super Breaks the Petaflops Barrier from The Linux Beacon for the feeds and speeds on that machine, and see The Top 500 Super Ranking Now Counts Watts as Well as Flops for our coverage on the Top 500 rankings, which appeared later in the week in The Unix Guardian.)
Soltis is not just speaking out of the side of his mouth when he says that supercomputing technologies eventually make their ways into data centers and, eventually, onto desktops. It is obvious that you could put a man on the moon with the computing power embodied in one Sony PlayStation 3 game console–you might even be able to upgrade the air traffic control system in the United States with a few dozen of these. (I jest–I hope.) But Soltis is also an expert on supercomputing, as it turns out, and is a professor of electrical engineering and computer science at the University of Minnesota where he teaches graduate courses on high performance computer design. So even when he might be exaggerating for effect (as he does from time to time), you have to give serious weight to what he says.
The hybrid nature of supercomputers is the new thing that Soltis was referring to, not the massive scale, as a pointer to the future of the i platform. So don’t think that there is going to be a 1 million core OS/400 box out there in the world. (In fact, I just last week learned that OS/400 and its successors only scale to 32 cores in a single system image. But back in 1995, IBM already created a parallel system and database architecture, called DB2 Multisystem, and as I have said many times before, IBM could create a massively parallel i super-database-engine if someone was willing to pay for it.) IBM Rochester is the place where the BlueGene and Roadrunner supercomputers are designed and built, the former being a massive cluster of 32-bit PowerPC engines (BlueGene/L uses dual-core chips, BlueGene/P uses quad-core PowerPC engines), while the latter is a mix of Opteron blades and Cell processors (used in the PlayStations) ganged up such that the Cells are really co-processors for the Opterons. (Actually, the Cell chip is itself a Power4-style 64-bit core that has 8 auxiliary co-processors, so you have co-processors on the co-processors in the Roadrunner box.) Other supercomputer makers use co-processors, sometimes special ones called field programmable gate arrays, or FPGAs, that are programmed to do specific functions.
Looking ahead to the Power7 processor generation, which is expected in the 2010 timeframe according to Soltis and other IBMers I have spoken with, Big Blue will be adding processor cores to the Power chip (four seems like a logical number, but IBM has not said) as well as plenty of adjunct processing to boost all kinds of processing that is currently done by central CPUs or external peripherals. As Soltis pointed out, this will be a bit of a “back to the future” moment for IBM Rochester, which designed the original CISC processors used in the System/38 and AS/400 to have a minimal amount of oomph, but lots of bandwidth for transaction processing and pushed number-crunching and other functions out to I/O Processors, or IOPs, in the system. Such a machine is called an asymmetric processor, and to a certain way of looking at it, so is the Roadrunner supercomputer created by IBM. Soltis is not saying that IBM is moving back to using external IOPs in the future Power7 servers–don’t misunderstand. It is more like, if I understood his hints correctly, that IBM is bringing IOP-like functionality back into the Power7 processor complex, but not putting it directly on the main core.
Here’s why. With the Roadrunner machine, IBM has hit 1 petaflops of performance using 6,948 dual-core Rev F Opterons that run applications, which are boosted by 12,690 Cell chips, which have 12,690 Power4-ish cores married to 101,520 co-processors on the Cell chip. As Soltis points out, this approach allows certain algorithms to be boosted by a factor of between 10 and 12. To put that in plain English, if you have a 5 GHz Power6 core, that would be the same as boosting it to 50 GHz or 60 GHz clock speeds. We can’t even get close to 10 GHz without rewriting the laws of thermodynamics and physics, although I can remember Intel once talking about 10 GHz chips. You will find it amusing that Soltis says the IOP approach in the System/38 and AS/400 design allowed the machine to do somewhere from 10 to 12 times as much work as that modestly powered central CISC processor could do all by its lonesome. . . . Good ideas don’t change. People change things, because they think that is the only way to get their paychecks.
Without being too specific, Soltis hinted that the Power7 chip would probably have lots of this adjunct processing capacity, and not a particularly high thread count. Here’s why. Putting two virtual threads on each core helps performance a lot–maybe as much as a 50 percent boost over a single thread per core. But moving to four threads per core doesn’t help as much, and putting 8 or 16 threads on a core gives what Soltis said were minimal gains. “What happens when the industry gives you 1,000 threads on a processor? How about a million? How are you going to program that?” asked Soltis rhetorically. With much delicacy, finesse, and smarts. But since we can’t crank clock speeds, adding threads is the only way we can boost performance, and so we are going to have to figure it out. Massively parallel supercomputers will lead the way. “There’s going to be some new languages and possibly some new operating systems needed and some extended languages–extensions to RPG, COBOL, C++, and Java to take advantage of the new hardware,” Soltis said in the Webinar. “This hybrid architecture is going to find its way into our business systems.”
Some of these special purpose functions are already being tested out in the Power6 chip, including a memory addressing algorithm in i5/OS V5R4 that is implemented in Power6 hardware running i 6.1, significantly boosting performance–by about a factor of 10.
“The beauty of the i is that it can incorporate these new structures without impacting anything from the past,” Soltis said with a certain amount of pride.