Two Crazy iSeries Ideas for 2004
January 19, 2004 Timothy Prickett Morgan
Like many people at IBM, and many loyal iSeries and AS/400 customers who want to see the OS/400 platform survive and thrive in the coming years, I spend a fair amount of time trying to figure out how to help the iSeries business grow. The health and strength of a platform’s ecosystem is directly proportional to the number of customers it has and how much money they spend. And I think I have two crazy ideas that might help the iSeries grow a bit.
Those of you who read my iDeal iSeries articles back at the end of 2002, before IBM launched the revamped iSeries line in January 2003, know I was a proponent of using the very cheap and powerful 64-bit PowerPC 970 processor, in the low-end and midrange of the iSeries line, to create inexpensive OS/400 servers. (The PowerPC 970 is also called the G5 processor when used in Apple workstations and servers.) I also spent a considerable amount of time espousing the idea that a PowerPC-based blade server for the BladeCenter chassis, which will start shipping in March 2004, ought to run OS/400, as well as Linux and AIX. These positions were based on the idea that the special OS/400 instructions that allow OS/400 to be supported on PowerPC and Power4 processors were in the PowerPC 970/G5 processors. I asked IBM repeatedly if the instructions were there, and time and again I was told that they were. As I explained late last year, in a series of stories that previewed future iSeries technologies, the top brass in the iSeries organization said that the 64-bit G5 processor did not, in fact, support that 65th bit that allows OS/400’s single-level storage to work.
I was angry at this lack of foresight, and have been stewing about it ever since. I think I have another answer, one that will help get a vast installed base of ancient CISC AS/400 users to start modernizing. I also think I have an idea that will help Apple to support larger customers and to push the IBM Power platform into new accounts.
CRAZY ISERIES IDEA NO. 1
First, a little math on the OS/400 server installed base.
As best as I can figure, IBM has sold approximately 385,000 AS/400 and iSeries machines since 1995, about 325,000 of which used a PowerPC or Power4 RISC processor. Two years ago, I estimated that there were about 550,000 installed OS/400 servers in the world, a number that IBM gave the nod to. With sales of between 55,000 and 60,000 for new iSeries servers, and a rampant server consolidation program going on among IBM’s customers, it’s hard to say how big the installed base of OS/400 servers is right now. It could be 450,000 or 500,000, or maybe it’s still holding at 550,000–depending on how many machines have been unplugged. Say there are still 500,000 AS/400 and iSeries machines chugging along out there. If you subtract 325,000 RISC-based AS/400 and iSeries machines from a total installed base of 500,000 OS/400 servers, you end up with 175,000 vintage CISC-based OS/400 servers that are still running in the world.
This is a vast market, and it represents between five and six years of current iSeries shipments (at rates set for 2001, 2002, and 2003). It is hard to say how much money such customers might spend on a new iSeries server (if they could be convinced to buy one), but at a $10,000 average configured system cost (including real-world configurations, all tools, and a lot of storage), that’s $1.75 billion, or about one year’s worth of current iSeries sales. If IBM thinks these vintage AS/400 customers are going to buy an iSeries Model 270 with a geared-down S-Star processor and spend tens of thousands of dollars on a configured machine–and Big Blue’s marketing strategy clearly indicates that it does believe this–the company is flat-out wrong. These companies are the new disenfranchised IBM midrange customers, like those companies that stuck with their System/36s way back when (and I hear that many of them are still running these boxes).
I’m oversimplifying here because I am not a systems engineer, but my understanding is that OS/400 requires some extra bits to deal with memory tags, which, in turn, tell OS/400 how to cope with single-level storage. The CISC-based AS/400s were a 48-bit processor and apparently used a 49th bit to deal with the memory tags that allow the single-level storage architecture of OS/400 to work. OS/400 V3R7 was recoded in Java and C++ to be a 64-bit operating system with extra circuitry to deal with the 65th bit for memory tags, and its kickers, OS/400 Versions 4 and 5, followed in this path. So here is what I want to do, and it is a very simple thing to say, but it may not be all that easy for IBM to do, technically or politically: Create a 49-bit OS/400 operating system that will run on current 64-bit PowerPC hardware and will support these old CISC-based OS/400 applications in 48-bit mode. Then support this OS/400 on Apple’s G5 servers, on the IBM BladeCenter PowerPC 970-based blade servers, and maybe even the low-end of the pSeries server line. I don’t think bringing OS/400 Version 3 forward to PowerPC is possible, so this entails backstepping OS/400 Version 5 to 48-bit application support.
If the memory tags require special circuitry and instructions on the chip, emulate those instructions in microcode and take a performance hit. Those old Motorola 68000 processors, which are the forebears of the PowerPC processors (and were used inside AS/400s as coprocessors for peripherals), could do this, and I suspect that, although IBM has never said so, the PowerPC and Power processors can store emulated instructions in special ROMs or memory circuits, too. Create the instructions at boot time and store them in L1 or L2 cache. There is probably some way to do this. If Intel can emulate 32-bit X86 applications on the 64-bit Itanium with a 30 to 50 percent performance penalty, I have to believe that IBM can emulate 48-bit OS/400 applications on a 64-bit PowerPC 970/G5 processor.
Let’s assume the smart people at IBM can figure out how to do this. Take a look at the new Xserve G5 server and Xserve RAID arrays from Apple, which were just announced at Macworld two weeks ago.
The Xserve machines run Mac OS X, a derivative of the Berkeley Systems Design (BSD) variant of Unix. The Xserve RAID array can be attached to Mac OS X machines, as well as to Windows and Linux boxes. The core functionality of the new Xserve server looks a lot like other rack-mounted X86 machines. It fits in a 1U form factor (like prior G4 Xserves) and can support two-way symmetric multiprocessing using the G5, also known as the PowerPC 970. The G5 is a trimmed-down implementation of the 64-bit Power4 processor from IBM that supports a frontside bus for each processor that runs at half the clock speed of the chip’s core. (Just like the Power4 processors do inside IBM’s own iSeries and pSeries machines.) That means a 2 GHz G5 processor has a frontside bus running at 1 GHz, delivering 16 GB/sec of bandwidth. This is a lot more than the 167 MHz frontside bus for the 1GHz PowerPC G4 could deliver. The G5 processors are available running at 1.6 GHz and 1.8 GHz, but when they will be available in the Xserve machines is unclear.
The Xserve G5 supports up to 8 GB of DDR SDRAM with ECC protection, up to 750 GB of internal disk storage (three 250 GB Serial ATA drives), has two Gigabit Ethernet ports on the motherboard and two FireWire 800 ports, one FireWire 400 port, two USB 2.0 ports, and a serial port. The G5 server can support one full length PCI-X card running at 133 MHz or two short PCI-X cards running at 100 MHz. The Xserve G5 runs Mac OS X, which is based on the open-source BSD distribution of Unix. A base machine with a single 2 GHz G5 with 512 KB of L2 cache, 512 MB of main memory, and 80 GB of disk costs $2,999. A two-way version of the machine with 1 GB of main memory costs $3,999, and a two-way machine with 2 GB of main memory and three 250 GB disks costs $6,599. A single drive bay version of the Xserve G5 with more airflow for cooling sells for $2,999 with two 2 GHz G5s, and is known as the cluster node.
The Xserve RAID array is a 3U chassis that can support up to 14 hot-swap, 7200 RPM Ultra ATA drives. While many storage array makers have focused on using the fastest SCSI disk drives running at 10K RPM or 15K RPM for their arrays, which are used primarily as direct-attached storage used to be in the old days, Apple has taken the path of using low-cost ATA drives for server storage (much as many NAS array makers have done). The Xserve RAID attaches to servers via a Fibre Channel link, just like other modern arrays do. Each ATA drive attaches to its own channel, which, in turn, links to the Fibre Channel fabric, which Apple says means that capacity and I/O performance scale as drives are added to the Xserve RAID devices. The Xserve RAID supports RAID levels 0, 1, 3, 5, and 0+1 using its internal controllers, and can support RAID 10, 30, and 50 by mixing hardware RAID on the arrays with software RAID on the Xserve servers. The base machine has four 250 GB ATA drives, 256 MB of cache, dual RAID controllers, and dual Fibre Channel server links for $5,999.
Apple says that, based on suggested retail Web pricing as of December 30, it can deliver 3.5 TB of storage in the Xserve RAID for $10,999, or $3.14 per GB. This is one third the price of much higher performing SCSI arrays from Dell, and is a lot less than what Apple reckons is the competition from IBM, Hewlett-Packard, and Sun Microsystems in the entry disk array market. A Dell/EMC CX200 array with 2.1 TB of storage costs $18,999, or $9.05 per GB, by Apple’s math, and an HP StorageWorks 1000 array with the same capacity costs $23,925, or $11.39 per GB. IBM’s FASt200 3542-1R array is a similar 3U RAID box, and it costs a whopping $51,895, or $24.71 per GB, in a 2.1 TB configuration. A Sun StorEdge 6120 array requires two 3U chasses to get to 2.04 TB and costs $74,600, or $36.57 per GB.
This is a pretty inexpensive potential OS/400 server and storage array. If 48-bit OS/400 could be created and ported to these machines, IBM and Apple could work together to repackage a two-way motherboard and a RAID array into a tower chassis (with the motherboard and RAID array tipped on their sides) that was smaller than the iSeries Model 270. The Apple components might even fit in the same Model 270 box, for all I know. All I am saying is that it is an interesting possibility, and considering the age of those vintage AS/400s, a so-called “eServer Xserve” running OS/400–even at a severely degraded performance–would be a killer upgrade for these OS/400 customers.
CRAZY ISERIES IDEA NO. 2
As best as anyone can figure, there are about 25 million Mac desktops in the world, with about 10 million of them on Mac OS X (the 32-bit and now 64-bit version of the operating system based on Unix), and about 15 million of them running its predecessor, Mac OS 9. Apple has been a complete laggard when it comes to delivering powerful Mac servers, and even the current Xserve machines might be underpowered to run real-world applications.
What a lot of Mac shops probably need is a much bigger Xserve than Apple can deliver. A two-way is nice, but you can’t run database applications on it. This is probably why, in the installed base of 10,000 Mac OS applications, there are not a lot of heavy-hitting ERP-style applications. To be fair, a lot of newspapers, graphic-arts studios, digital rendering, and multimedia streaming services are based on the Mac platform. These companies would probably like a single, consolidated platform on which they can support fewer servers to run their Mac workloads (whether OS 9 or OS X) and on which they could install Unix, Linux, or OS/400 applications to run their businesses. Enter stage left: the IBM-Apple, co-branded eServer Xserve.
What IBM, Apple, and Motorola promised way back in 1991 was an integrated Power platform that would run Windows, AIX, Mac OS, and OS/400. They obviously did not deliver on this. But porting Mac OS X to the eServer Power line (which is embodied in the current iSeries and pSeries lines) could fulfill at least some of that promise. IBM and Apple could offer distinct eServer Xserve machines through a common channel for Mac customers. Alternatively, IBM and Apple could simply support Mac OS X partitions within logical partitions on the Power4 and Power5 lines of machines, alongside OS/400, Linux, and AIX. They could do both. They could also deliver G5-based Integrated Xserve Server cards for iSeries machines running Mac OS X, and they could offer Integrated Xserve Adapter cards for linking Macs to the iSeries storage subsystem and Integrated File System, much as IBM’s own xSeries Windows servers can do today.
At the right price–meaning something a bit lower than IBM charges at the street level for its pSeries line–Mac shops might take a shining to much bigger Power-based servers than Apple can deliver. And if Apple charges a bit more than the $999 it currently charges for an unlimited user license to Mac OS X Server edition–maybe it can charge two or three times as much to cover machines with two or three times as many users–it works out for both IBM and Apple. The point is, neither IBM nor Apple wants Mac shops to buy big Windows boxes for their bigger jobs. This strategy gives Mac shops a native option on bigger iron, and it helps prop up the entire Power chip ecosystem on which both Apple and IBM depend to a large extent.