Headlines

Headlines

Computational Powerhouse Hidden in Island Jungle

Seemingly unglamorous RF atmospheric research requires surprisingly complex calculations performed by the equivalent of supercomputing systems.

Perhaps one of the more unusual applications of serious computing power is innocuously located within the sweltering jungles of Puerto Rico. Inside a modern, air conditioned data center in the Aeronomy Department of the Arecibo RF Telescope Observatory, scientists and engineers study the upper regions of the Earth’s atmosphere. Their goal is to help improve the reliability of radio and satellite communications for the military, security, and global telecommunications industries.

Achieving this goal requires compute-intensive systems to handle the complex algorithms and exponential manipulations associated with high-resolution RF signal analysis. What follows is a brief history of the decade long challenge to implement the most appropriate computing technology, from early computers to today’s multicore parallel architectures.

Resolution Matters

Using atmospheric radar, engineers transmit a coded RF signal into various layers of the atmosphere. They then compare it to the received signal. In the past, a digitally modulated Barker Code was used to compress the pulse width of the bursted transmitted signals, notes Arun Venkataraman, head of the Arecibo Observatory Computer Department. Compressed pulsed signals achieve wider pulses and better signal-to-noise ratios than un-pulsed signals—a necessity for high-distance resolution. The Barker pulse employs binaryphase- shift-keying (BPSK) modulation, which is why it’s also used in many spread-spectrum applications like the 1- and 2- Mbit/s IEEE 802.11x wireless-communications standards.

Continuing refinement in both RF performance and electronic technology have resulted in the adoption of higher-resolution—hence longer code length—signals like the coded-long-pulse (CLP) sequence. This pulse sequence is so long that it must be decoded in sections. This requires a compute-intensive decoding process that involves the calculation of numerous autocorrelation functions. The resulting data is used to determine the intensity or concentrations of the reflective atmospheric layers all the way up to the most upper regions (i.e., the ionosphere). All of this information is critical to understanding how radar signals interact within the different layers of the atmosphere.

Achieving the computing performance necessary to decode and analyze the CLP signals requires the use of two sets of Intel® Xeon® processors 5500 (16 cores each or 32 cores total).

How were these compute-intensive calculations handled before the advent of modern multicore processing systems? What options—ranging from array processors to field-programmable gate arrays (FPGAs) and graphics processing units (GPUs)—are being explored for the even-greater computational needs of the future?

Past Computational Technology

In the 1970s, some of the earliest atmospheric radar work at Arecibo used array processors from a company called Floating Point Systems (Portland, OR). Array processors are specially designed to process arrays (i.e., matrices of numbers). Several of the floating-point-system (FPS) array processors were connected to a 24-bit Unisys Aries minicomputer. One of the FPS processors was used for control while the other functioned as a co-processor to calculate Fast Fourier Transform (FFT) sequences.

As computational needs increased, the FPS and minicomputer were replaced in the 1990s by a SkyBolt system. It included Intel’s 8000 and 9000 series processors. According to Venkataraman, multiple SkyBolt systems were needed to achieve the necessary processing power. Because it proved extremely difficult to synchronize all of these systems together, the effort was abandoned. Interestingly, interconnect issues remain a challenge for today’s parallel-processing multicore processors.

For a while, the engineers at Arecibo struggled with general- purpose machines. They actually built platforms based on Motorola’s G-5 processors—a 64-bit PowerPC architecture that was the successor to Motorola’s PowerPC 7400 series. The engineers even created a cluster of Apple G5 processors connected with Sun’s Grid tools to create a parallel processing environment. But even these successes weren’t enough to meet the growing computational needs. Finally, the Intel Xeon processor-based platforms were used to achieve the current compute power. But will these workhorses be sufficient for the future?

Figure: Computing servers with the Intel® Xeon® processor 5500, the first processor with the company’s Nehalm architecture. Also shown – on the right - are a bank of Aberdeen network disk servers.

Future Computations

To meet the ever-increasing need for greater computing power, the computer department at Arecibo considered buying a Cray supercomputer. But careful analysis by Venkataraman suggested that the same processing performance could be had for less money by complementing the existing array-processor architecture with several graphics processing units. As a result, computer engineers are currently evaluating nVidia’s 512-core Tesla series of high-end-performance GPUs. Their goal is to use the inherent parallel architecture of the graphics chips to facilitate greater levels of processing speed and performance. Both the Tesla and the soon-to-be-released nVidia Fermi—with thousands of cores in a board—are being considered.

Using GPUs is nothing new at Arecibo. In addition to the Aeronomy department, Astronomy is another field where intensive data calculations are needed. According to the observatory’s website, “In 1993, the first pulsar in a binary system was discovered, leading to the important confirmation of Einstein’s theory of general relativity and a Nobel Prize in 1993 for astronomers Russell Hulse and Joseph Taylor” (http://www.naic.edu/accomplish.php).

Shortly after that time, according to Venkataraman, Taylor bought a special-purpose vector array processor called the Super Harvard Architecture Single-Chip Computer (SHARC)—a high-performance digital signal processor (DSP) from Analog Devices. This type of machine was needed to handle coherent dedispersion from the pulsars under study. Dispersion occurs due to the great distances between pulsars and radio receivers on the Earth. What starts out as a sharp, well-defined signal from a pulsar becomes a flattened smear of frequencies when it reaches the Earth. Coherent dedispersion is a technique that greatly improves the timing of pulsars. It requires numeric transformations that involve the multiplication of millions of complex exponential functions. This is an extremely processor-intensive set of calculations. The Arecibo Observatory works with the Green Bank Ultimate Pulsar Processing Instrument (GUPPI) at the National Radio Astronomy Observatory (NRAO) in West Virginia to perform these complex calculations.

John Blyler can be reached at:
jblyler@extensionmedia.com