IBM Corp isn’t concentrating only on hardware for forthcoming AS/400s. IBM hopes to put the finishing keystrokes on Version 3 Release 7 of OS/400 sometime during 1996. If IBM gets it out the door concurrently with the faster PowerPC AS chips, add another 25% to 30% to our performance estimates. (IBM’s software experts feel confident that they can speed up OS/400 in a big way.) But we wouldn’t be surprised if Version 3.7 significantly lags the second generation RISC AS/400s, just like V3.1 was almost a year behind the first black box machines. With the shrunken A30 chip (which will undoubtedly have a different name), IBM will also begin using the symmetric multiprocessing features in the A30 to build AS/400s that incorporate more than four central processors. IBM ran into a wall attempting to create AS/400s that used more than eight CPUs. So the AS/400 division called in the world’s symmetric multiprocessing experts, IBM’s mainframe gurus from the Hudson Valley, to help it to figure out how to build bigger AS/400 systems. The problem with using symmetric multiprocessing beyond six central processors is that performance severely degrades with each additional processor.
Faster processors
The situation very quickly gets to the point where extra processors don’t yield extra throughput. The trick to fix this is subtle, but no less ingenious. What IBM does is use faster processors for numbers 7 through 12 than are used for numbers 1 through 6. It’s tough to balance the processors, but IBM has gotten it to work. Although it is difficult to say how big these symmetric multiprocessing AS/400 systems will get, the aggregate power in the high end boxes will likely range from 280 (eight- way) to 425 (12-way) RAMP-C relative performance units. If V3.7 performs as expected, that range will jump to 350 to 550 relative performance units. If AS/400 customers show that they need the power, IBM may also provide tightly-coupled AS/400 clusters that will bear more resemblance to IBM’s mainframe parallel Sysplex than it does to the AS/400’s nascent OptiConnect clustering scheme. (This project is extremely iffy, so don’t bet your budget on it.) The big development in AS/400 processors will come in 1998, when IBM hopes to announce a very fast PowerPC AS chip. This chip will have elements in common with IBM’s high-end PowerPC 630 chip, although it will be developed solely by the AS/400 division. (The AS/400 division has taken back its chip development programme from the IBM-Motorola Inc-Apple Computer Inc consortium known as Somerset because of delays at Somerset.)
By Timothy Prickett
The super AS chip may be used by the RS/6000 group; it all depends on what happens with the PowerPC 620 and 630 projects, both of which are losing steam quickly. No word yet on how much power the super AS chip will have, but we expect that uniprocessors based on it will offer about 80 RAMP-C relative performance units. If IBM gets V3.7 ready for these chips, performance could be a lot better. They will be even higher still if Frank Soltis, the father of the AS/400, adds Very Long Instruction Word processing to OS/400. The other major component of the AS/400 system that will have to be enhanced as IBM adds faster CPUs and memory chips to the system is auxiliary storage. Not only will disk drives get smaller, faster and more capacious, but the technology that binds them together into disk arrays will also be radically improved. IBM and plug-compatible disk arrays for AS/400s all support fast, wide SCSI interconnection. This extension of the SCSI interface allows for data to be transferred between disks and CPUs at a maximum of 20M-bytes per second in burst mode. (Real world data transfer rates are often a lot lower, just like actual throughput in a data communications link is often a lot lower than the theoretical maximum.) But fast, wide SCSI will probably not be fast enough, especially for very large AS/400s or for clusters of AS/400s. IBM has developed a faster disk technology called Serial Storage Architecture, or SSA, that
has four times the bandwidth of fast, wide SCSI. Serial Storage Architecture disk subsystems have been announced for the RS/6000 line of minicomputers; IBM says that it has no plans to incorporate Serial Storage Architecture disk subsystems in the AS/400 line right now. That could, of course, change. But we think that it is much more likely that future AS/400s use an even faster disk interface technology, Fibre Channel.
Internet
The reason that AS/400s will most likely use the Fibre Channel interconnection scheme is raw speed. Initially, Fibre Channel will support burst data transfer rates of up to 100M-bytes per second. Down the road, Fibre Channel will be enhanced to support data transfer rates as high as 200M-bytes per second. As client- server and Internet technology takes off in the commercial marketplace, Fibre Channel will probably take the lead simply because there is no practical alternative disk technology. Seagate Technology Inc is already making Fibre Channel disks; disk arrays supporting Fibre Channel are expected to start appearing on the market sometime during the first half of this year. Sun Microsystems Inc, which co-developed the Fibre Channel specification with Compaq Computer Corp, has already begun using it in high-end Sparc servers. Compaq will undoubtedly soon offer Fibre Channel subsystems for its big ProLiant servers. IBM, Hewlett-Packard Co, Digital Equipment Corp and others have indicated that they will also use Fibre Channel where it is appropriate. IBMers familiar with the arguments now going on inside Rochester and San Jose say that, in the long run, IBM will have to adopt Fibre Channel for the AS/400 to keep up with competitors. From the January 1996 issue of The Four Hundred. Copyright (C) 1996 Technology News of America Inc. All rights reserved.
