The Second Coming of TSUBAME
When the TSUBAME 2.0 supercomputer is formally inaugurated in December, it will officially be declared the fastest supercomputer in Japan. However, it’s not simply speed that separates this machine; boasting a raw performance of 2.4 petaflops, the new TSUBAME exceeds the total FLOPS capacity of all other government and academic supercomputers in Japan today. That kind of computational brawn will make it the platform of choice for some of the most powerful scientific applications on the planet.
TSUBAME’s other claim to fame is that it is part of the vanguard of a new generation of GPGPU-powered supercomputers making their way into large research institutions around the world. China seems particularly enthusiastic about the technology and has already installed three GPU-equipped supercomputers: Nebulae (2.98 petaflops), Mole-8.5 (1.14 petaflops), and Tianhe-1 (0.56 petaflops). France, Australia and the US are also gearing up to deploy large GPU-accelerated clusters over the next 18 months, and other countries will no doubt follow suit.
Tokyo Tech, though, has been in the GPGPU camp for a couple of years now. After flirting with ClearSpeed accelerators with TSUBAME 1.0 in 2006, by 2008 they realized that GPUs were destined to become the modern-day vector processors for HPC. That year the institute added 170 NVIDIA Tesla S1070 servers (680 GPUs) as part of the TSUBAME 1.2 upgrade, which increased the machine’s peak performance from 80 to 141 teraflops. This 1.2 incarnation also turned out to be the first GPGPU-powered supercomputer to earn a spot on the TOP500 list.
These first-generation TSUBAME machines were built with Sun Microsystems gear, based on the AMD Opteron-based x4600 servers. The second generation is quite a different animal. The dissolution of Sun’s HPC roadmap under the Oracle regime meant Tokyo Tech had to find a different system vendor going forward. That vendor turned out to be HP (who, by the way, will also be the prime OEM for the NSF-funded Keeneland GPGPU supercomputer in the US). HP, along with NEC, co-designed the second generation system, with NEC also providing on-site integration and software tuning.
As pointed out in last week’s coverage of HP’s new GPGPU gear, TSUBAME 2.0 will use the company’s latest ProLiant SL390s G7 server for its compute infrastructure. Specifically, the 2.4 petaflops of compute power will be derived from 1,442 compute nodes: 1,408 of which are the new SL390s G7 nodes, each equipped with two Intel Westmere EP CPUs (6-core, 2.93 GHz) and three NVIDIA M2050 “Fermi” modules. The system will also include 34 Nehalem EX-based nodes hooked up 34 10-series Tesla S1070 servers. Total memory capacity for the system is 80.6 TB of DRAM, plus 12.7 TB of local GDDR memory on the GPU devices.
Each node will also be outfitted with either 120 GB, 240 GB or 480 GB of solid state disk (SSD) local storage for a total of just under 174 TB. External storage is provided by over 7 PB of DataDirect Networks gear, including a 6 PB Lustre partition plus another petabyte of NFS/iSCSI-based disk. An 8 PB Sun SL8500 tape system represents the final layer to TSUBAME’s storage infrastructure.
The whole cluster is woven together with QDR InfiniBand, in a full bisection, fat tree architecture. Voltaire is supplying the networking gear, including 12 core switches (Grid Director 4700) and 179 edge switches (Grid Director 4036). The SL390s G7 nodes will use dual-rail InfiniBand, utilizing Mellanox silicon on the motherboard and an adapter card for the second rail. Half a dozen 10 GbE switches, also supplied by Voltaire, have been installed to hook the machine up to the Sun tape system.
Compared to the Jaguar supercomputer at Oak Ridge, which has essentially the same peak performance (2.3 petaflops), TSUBAME 2.0 is just one-quarter its size and will use about one-quarter the power. In fact, the 44-rack TSUBAME takes up only 200 square meters of floor space and consumes only around a megawatt of power, which helps to explain why GPU computing has become so popular in this island nation of limited land and energy resources.
At the recent NVIDIA GPU Technology Conference, Professor Satoshi Matsuoka, who leads the TSUBAME effort, reported that the system was installed over the summer and is now up and running. Currently it is undergoing stress tests, which includes Linpack benchmarking. Matsuoka said they expect to achieve between 1.2 to 1.4 petaflops on Linpack, which would place it in the number 2 spot on the TOP500 list today, behind Jaguar’s 1.7 petaflop mark.
Where Matsuoka expects the system to really shine though is with real-world applications in climate and weather forecasting, biomolecular modeling, tsunami simulations, CFD codes, and raft of other scientific codes. Tokyo Tech’s TSUBAME users have already had a couple of years developing CUDA software on the previous generation system. In a few short weeks, the first batch of second generation CUDA codes will meet their second-generation GPUs — and on a true petascale platform.
One of the most interesting applications is ASUCA, a Japanese weather forecasting code similar to the US-based Weather Research and Forecasting (WRF) model developed by NCAR and others. ASUCA, though, has been ported in its entirety to GPUs and is projected to achieve 150 teraflops on the new TSUBAME. To lend some perspective, the current weather forecasting performance record is around 50 teraflops with WRF running on ORNL’s Jaguar. The GPU-accelerated ASUCA represents an 80-fold improvement compared to a CPU-only implementation.
According to Matsuoka, although the official TSUBAME 2.0 inauguration is in December, the system will be available to users around the beginning of November. That includes over 2,000 researchers in academia, government and industry, including some outside of Japan. And until Japan’s 10-petaflop “K” supercomputer becomes operational in 2012, TSUBAME will be the country’s number one machine. Beyond that, a 30-petaflop TSUBAME 3.0 is already on the drawing board and is expected to launch in 2014 or 2015.
In the meantime, Tokyo Tech is poised to become a hotbed of petascale GPU computing for the open science community. Matsuoka has been an outspoken advocate of the technology and has managed to attract a vibrant community of software developers to the TSUBAME mission. Although Tokyo Tech can’t compete in budget and facility size with the most well-endowed research institutions in the world, it has compensated for with its enthusiasm and clarity of purpose.