Interview: Head of 1st ARM Supercomputer Project
Computing power is growing by leaps and bounds, everyone knows that. However, in the realm of high-performance computing, faster chips are not the only measure of success. Alejandro Ramirez, leader of the team at the Barcelona Supercomputing Center (BSC) developing the world’s first ARM-based supercomputer, spoke to Opinno about the need to save energy and invent a whole new scientific computing environment.
The computer under development at the BSC is a prototype system using 64-bit, quad-core Nvidia Tegra3 ARM processors paired with the company’s graphics processors. Until now, ARM processors did not have the double-floating point precision necessary for scientific computing. This computer is one of the first steps in the European Union’s Mont Blanc project which aims to develop petaflop-level computing systems that use a small fraction of the energy current systems consume.
“The goal is to achieve a large energy savings,” said Ramirez. “ We want to have the same amount of calculations but use four times less energy; four to six times the energy savings depending on with whom you compare. The system that we are ultimately going to design is planned to be capable of 50 petaflops. But the prototype that we are constructing will be much less powerful. The objective is 50 petaflops with a power usage of seven megawatts. For comparison, the number-one supercomputer in the world is capable of 10 petaflops and uses 12 megawatts.”
As ARM processors are renowned for their computing-to-power ratio (a typical ARM processor consumes less than a watt while an Intel Core i7 consumes around 100 watts), they have found utility in low-power applications such as smartphones, tablets an embedded systems.
Even though the architecture is widely used in low-power systems, many challenges remain for ARM to be used in high-performance and scientific settings.
Linux will be the operating system of choice the BSC’s efforts as it is open source and has been compiled to operate on the ARM architecture. There are also many flavors of Linux that are purpose-built for massively parallel supercomputers such as SUSE and LAC Linux.
“One of the primary advantages is the operating system software support. One of the requirements is to use a system that is well-known by programmers. There are other more efficient alternatives, but the support of the Linux operating system is one of the fundamental elements,” said Ramirez.
While Linux is ready to go for the new platform, the applications are not. Ramirez spoke about the need to write software for the platform as simulations and other mathematically intensive applications are the bread and butter of supercomputers. Currently, all existing software is compiled to run on X86, Itanium and SPARC systems.
“…While ARM supports Linux well, what it doesn’t support is the rest of the software environment used in high-performance computing,” he explained. “The software libraries and the mathematics [programs] provided by vendors like IBM or Intel with their machines are optimized for their systems. This software doesn’t exist for ARM. We have the operating system, but the majority of the software is missing.”
Nvidia’s relationship with the BSC’s Mont Blanc efforts revolve around the use of Nvidia’s advanced CUDA architecture which offloads many of the computational tasks to graphics processors. As with scientific applications, the BSC is breaking new ground as ARM drivers for Nvidia’s chips― which don’t exist. Ramirez said Nvidia is working closely with the team in order to write the necessary software. It is hoped that an initial version will be ready first quarter 2012.
“We still don’t have drivers for the graphics processors we use for calculations,” he said.
Ramirez predicts ARM supercomputing will catch on quickly as it will provide a practical way to develop low-power supercomputing for use in the field. He thinks geologist will be keen on the technology due to it being more practical to deploy on-site in mineral exploration. Ramirez also sees applications in biotechnology and pharmaceutical research.
When asked about HP’s efforts in developing low-power commercial servers, Ramirez predicted their efforts would not be as challenging as those of the BSC. The reason behind this is most server software such as web servers and databases do not rely on the same mathematically-intense software that supercomputers use.
Ramirez said that with this new frontier, experience needs to be developed from scratch.
“The chip makers do not have experience producing supercomputing ARM chips nor do the supercomputer manufactures,” he said. “This is really the first time. We are opening new doors.”