Rocky dreams

Everyone knows that Sun's ultraSPARC RT is dead. Right? Wrong - the strategic reason for not releasing it disappeared with the Oracle deal.
Written by Paul Murphy, Contributor

Earlier this week I placed the last rock on the last wall to finish my summer landscaping project - well, except for shoveling more dirt next week. Two small walls, one along the sidewalk in the front and a little one along the back patio have embedded wire frames and concrete reinforcement but all the bigger stuff is pure drystack - and boy was I wrong thinking this was going to be easy! Oh well, live and learn right? - and, besides, I have a feeling it'll eventually appear to have been both easy and rewarding relative to my next challenge: helping take the Alberta Wildrose party from nothing to electoral success and political power.

One of life's little lessons here, of course, is that thinking or writing about doing something is a lot easier than actually doing it - and on that basis I want to share a hope about Sun's "Rock" Ultrasparc RT processor.

A few months ago, when Sun's senior people were getting their marching orders, Rock's absence from the SPARC road map they got fired off the rumor that Rock is dead, but that's never been formally confirmed by anyone from either Sun or Oracle.

Now obviously all the senior people involved are living in fear of litigation and keeping their mouths tightly shut accordingly, but the Oracle relationship changes the strategic picture on Rock to the point that it should be very much alive.

Rock is the first large scale processor to support three new technologies on release: transactional memory, hardware scout, and the second generation cool threads stuff. Since that's big change, and big change tends to be viciously resisted in the market, Sun's strategic problem with Rock was that the cost of getting it into production couldn't be met without quick market success and they didn't see a way to make that happen.

Oracle offers a solution to this: a way to skip the three to five years it usually takes to get new ideas accepted; and because that significantly changes the strategic picture I think there's reason to hope that the SPARC/Solaris solution they're planning to show on October 14th at their Openworld conference is based on Rock - not Niagara.

What Oracle can do to make Rock an instant success is simple: convert their own major products to the technology, and sell the combination on an appliance+ basis - with the plus being that these machines can also run any existing SPARC/Solaris binaries, albeit more slowly than they should. As a result customers can combine faster with cheaper at the major applications level while transitioning their IT teams to work with the new technologies - technologies the wintel mass market will be another five to eight years in getting around to copying.

The Oracle/Sun announcement that they plan to beat IBM's p595/DB2 TPC/C record at OpenWorld lends, I think, credence to this idea because it's the right benchmark for Rock and the wrong one for almost anything else. Among the reasons for that the most important is simply that only 10% of the transactions defined under the TPC/C standard require SMP, but these constitute the bottleneck for the p595 - the independence of the other 90% is, of course, the reason Sun, which specializes in high end SMP, despises the TPC/C benchmark while IBM, which exploits customer commitment to paying for big machines so they can be partitioned into little ones by building them as tightly interconnected clusters of little four and eight way machines, loves it.

Since Rock doesn't have this bottleneck and can exploit hardware transaction management on the severable queries, even a two way Rock processor with ZFS disk and a beta Oracle RDBMS with most of the lock management code stripped out, should be able to blow away IBM's p595 result on this test - something that would otherwise take 8 to 12 T5440s in a rack with at least a half dozen storage servers.

So are they going to do this? I don't know - but I'd give it perhaps three chances out of ten that they do, with another three going to the proposition that they want to but won't get the software done in time - and the remaining four on the safer bet that you'll see some T5440 gear running 800,000 to a million TPM each.

Editorial standards