Just how much more power you can get out of an IBM 3090 model S is now open to great debate, until of course someone gets one and benchmarks it with a non-proprietary test. IBM has mixed up its claims to such an extent that customers will find it very hard to calculate whether or […]
Just how much more power you can get out of an IBM 3090 model S is now open to great debate, until of course someone gets one and benchmarks it with a non-proprietary test. IBM has mixed up its claims to such an extent that customers will find it very hard to calculate whether or not any particular configuration is man enough for the job they have in mind. At the UK announcement IBM was telling us that it now has a machine in the 600S that will overpower an Amdahl 5990-1400. It adds that even without running the preferred ESA operating system, which will give a significant power boost, the machine outstrips the Amdahl equivalent. Executives cited 120 RPVs, which are Relative Processing Values, and not MIPS, but which we’re pretty sure amounts to much the same. Amdahl’s new machines are usually cited at well over 100 MIPS, approaching 120 MIPS, and in a four- rather than a six processor configuration. But looking at raw CPU-bound power, the improvement of the processor cycling time to 15ns should yield an internal throughput around 12% to 13%, some way short of the 56% that IBM is claiming for the 600S running ESA, over the 600E running XA. That throughput improvement is augmented by the fact that IBM has reduced the penalty of running a single image over six processor from around 23%, down to around 16% to 17%, which buys it 6% right away. But that still gives a maximum improvement of around 19% ignoring storage changes, ESA, and microcode optimisation.
Eight-way multiprocessor Storage hasn’t only gone up to 512Mb, and the high speed cache memory to 128Kb, but the way that cache operates has been vastly reworked. Details are still hazy, but we accept IBM’s word that it broke new ground in reorganising the directory for the high speed cache. While the throughput gains from this are impossible to work out analytically, IBM was moved to claim a total benefit of 10% to 30% on the low-end machines; 20% to 30% in the middle and 20% to 40% at the top end, although it fails to qualify this by detailing the job mix. ESA is again a factor here, and we can only return to the original 3% to 12% benefit it was claimed to endow at launch time. We would expect this to have been improved via microcode enhancements. Earlier worries that ESA would never really operate as fast as claimed should be removed, since the limit of 256Mb real storage was blamed, and this is no longer a problem. All this means that IBM’s claims for overall improvement of 20% to 40% are upholdable, but we would expect the 56% gain in heavy IMS work, the 52% claimed for CICS, and the 45% for DB2 processing, recording a new world record number of relational transactions a second at 270 and just pipping Oracle Corp (CI No 982) are all a little less believable, and it really depends upon what you are comparing with what. The gradual microcode move towards supporting online database processing, may just make these claims palatable, but meanwhile it looks as if the machine’s raw power is around 96 MIPS, and no lower than 108 MIPS driving through ESA. Amdahl could still make out a very strong argument to retain a lead, if somewhat diminished. Now the race is on for Amdahl and NAS to get ESA working, before IBM gets a G model out the door, the result of which will decide who keeps the lead in the mainframe power stakes. When to look for a G – or should we call it T now? April 1989 is a clear possibility. And will the announcement include an eight-way multiprocessor? You bet.