The Great Leap of the Jaguar

Portland, Nov 16, 2009 / Hans W. Meuer, Prometeus & Universität Mannheim, hans@meuer.de

At the third attempt, it finally made it. The supercomputer at the National Center for Computational Science in Oak Ridge, named Jaguar, now tops the 34th TOP500 list of supercomputers with 1.759 petaflops. In the previous two lists, the Jaguar had already gotten quite close to list leader Roadrunner from IBM (the first system worldwide to break the petaflops barrier in summer 2008) – the difference was less than 4 percent Linpack performance.

How did the Jaguar manage to leap to first place with 69 percent more Linpack performance than the Roadrunner? Since July, the Jaguar’s number of compute cores has been increased from 129600 to 224162 and – what’s substantial – it has been equipped with AMD’s new six-core processors ‘Istanbul’, which have only been available since August, and 2GB of memory per core. Each compute node features two Opterons with 12 cores and 16GB of shared memory. The whole system has 300TB of memory and 10PB of hard disc space.

Another factor was that two of the Roadrunner’s CUs (connecting units) have been dedicated to other tasks reducing the system’s Linpack performance from 1105 teraflops in June to 1042 teraflops now.

In June, Oak Ridge and Cray had hinted that they hadn’t wasted valuable computer resources on the Linpack, but had instead used the Jaguar for work on substantially scientific problems – in retrospect, a clever (bluff) move. They passed on a close first place in June in order to surprise everyone with a truly conclusive advantage over the Roadrunner now in November – nobody doubts that the Jaguar could have topped the list in June. Accordingly, Adolfy Hoisie of Los Alamos described the circumstance of the Roadrunner placing first instead of the Jaguar as the “miracle of Hamburg” at the ISC’09 Conference. The Jaguar now achieved 1.759 petaflops of computing power solving a linear system of equations with n = 5474272. It finished calculating after 17 hours and 17 minutes.

It’s worth mentioning that this is the very first time – after 34 TOP500 lists – that the company Cray (“The Supercomputer Company”), well known for being specialized in supercomputers only, has achieved first place. During the early stages of TOP500 and even before that, since 1976, Cray had been the market leader for supercomputers. Consequently, when the first TOP500 list was released in June 1993, 41 percent of the listed systems were from Cray. That’s an unbelievably high percentage because the list’s second, Fujitsu, only had a market share of 13.8 percent back then.

But now, 16 years after the start of TOP500, the dynamic HPC market has changed completely: in the course of the years, Cray has lost its position as market leader and has become a niche computer company serving the needs of high-end government laboratories and academic users. Other companies, first and foremost IBM and HP, have clearly taken the lead in the list. Cray had undergone a severe crisis because of the total collapse of the vector computer market, a crisis it still hasn’t recovered from. In the current list – Nov 2009 – Cray only has 19 systems, which equals a share of 3.8 percent. However, 14 of these systems are in the TOP100 and two of them even hold excellent positions in the TOP3. So, it seems possible the former leader Cray is on the rise again, especially now, motivated by the success of topping the TOP500 list. In any case, Cray will keep its first place until the next TOP500 list is released at the ISC’10 (www.isc10.org) in Hamburg next year, between Mai 31 and June 3.