Monday, November 05, 2007

RANT: Very Soon This Will NOT Be a Supercomputer



This is seaborg. It's an IBM SP3 with 16 way nighthawk2 processor nodes - that is each node has 16 processors - and there are a total of 414 nodes. Currently it is ranked 158th on the Top500 list of supercomputers. At one time it was #2, iirc. We are getting set to retire this NERSC workhorse. We have newer and shinier toys. It was named for Glenn Seaborg.

seaborg and I have a special relationship. When I go on rotation, it has serious issues. I don't even have to go login for it to freak out: within hours of me going on rotation, it'll just do its thing. Once it had three system wide outages in four days. I think that the damned thing has Munchhausen Syndrome. It became the vernacular around here that if seaborg crashed, I hugged it. In fact, I once pulled a prank on a coworker by having the operations group page her stating exactly those words: Will Hugged Seaborg (see the byebye to Tavia post). It even got noticed at IBM and they'd grumble if there was some way to get me off rotation just so their up time figures - a critical benchmark for contractual reasons for them and important to us and our users for other reasons - would stay higher. In a way, I am going to miss that big turkey. Then again, all the extra sleep I am going to get is going to more than offset that. ;)


I bring this all up because quite soon, seaborg will no longer be considered a supercomputer. You might be scratching your head at that. Let me throw you for another loop: that Cray right there is not a supercomputer either. Huh? It was. It isn't now. isn't a supercomputer, a, well, supercomputer? No, not really. Supercomputing, or High Performance Computing (HPC), is a moving target. Right now, to get on the Top500 list at all you have to be able to do a sustained 4 teraflops using the Linpack benchmark. Seaborg does a sustained 7 TFs. Very soon it will fall off the bottom end of the list. So what, you say, how does that not make it a supercomputer? A supercomputer is one of the fastest computers on the planet, or so goes the general definition. If its not one of the fastest computers, its not a supercomputer. Right? Sometimes people talk about the top x percent of computers performance wise. That's silly. There's millions of computers and yet very few are HPC machines, much less than a percent. For better or worse, a working definition is if you can get on the Top500 list, you're an HPC system. That also means that the definition moves. The fastest machines make it onto the top and push off those at the bottom. The minimum score keeps going up. A Cray-1 is slower than your desktop if its new. The only way that its comparable to your desktop even is that its memory subsystem is approximately as quick ( a bit sad, that, but...) So if you were to take your desktop back to 1975, you'd have a screaming #1 computing platform in the world.

Now, the reason I bring this up is that I am constantly bombarded in the IT world with companies claiming to have a supercomputer in a box. Or the most recent two that prompted me to write the post are here and here. 25 gigaflops would have been a screamer in the early 1990s. It's not even 1% - ONE PERCENT - of what's needed to make the cut now. That's not a supercomputer. Hell, a couple servers will outrip that thing. The second press release is another one that makes me roll my eyes: there have been multiple technologies that have come and gone that have promised TEH WAY DUDE for HPC. Most are corpses or fossils that litter our path. We'll see if the good professor's tech works out, until then wait and see. However, if its hand held, I doubt it'll be considered a supercomputer at the time its actually built. All I'd have to do is string together 256 of them or find some way to use that $tech in a way that won't work handheld and...you get the point.

Technology is constantly changing. Supercomputers are not like PCs. An Intel 386 based PC from circa the late 1980s is still a PC. A sucky one, but still a PC. A Cray-1 is not a supercomputer any longer. Seaborg won't be one quite soon too. So can we dispense with using supercomputer as an indicator of how stupid your marketing department is? just because your toy in your hand can out perform the greatest, fastest thing of the 1970s or even 1980s or whatever that isn't current, does NOT make it a supercomputer.

Ok. Incoherent rant over. ;)

No comments: