    
|
|
Instruments monitor planets in the hopes of answering the myriad of questions about our universe and its first stars. Yet many scientists are forced to turn off instruments and not collect science data because transmitting it to the ground for analysis involves expensive ground operations. But what if the instruments could stay on all the time through on-board computing? This possibility among others is what motivates the Remote Exploration and Experimentation (REE) project, funded by the High-Performance Computing and Communications (HPCC) Program. Since its inception in 1992, REE has staked its claim to high-performance, fault-tolerant, low-power computing in space and never looked back. It is this kind of computing that will enable sensors to be on all the time. Right now, instruments in orbit around the earth can see a ground station 15 minutes out of a 90-minute orbit. "If you want continuous bandwidth, you end up paying for huge ground operations and station costs around the world to keep in contact with your mission," REE project manager Robert Ferraro tells us. In addition, "more power devoted to sending data home means less instrument capability," states Hans Thomas, who works on NASA's Telerobotics Program. "For deep space and planetary missions, it's vital to have on-board, data pre-processing." COTS in space For that reason, REE is now focusing its research on a data processing, computing platform that can be adapted to different mission requirements. It is leveraging commercial, off-the-shelf (COTS), scaleable computing technology for use in space. COTS in space will yield several benefits, including ever-ready instruments with higher throughput that allows scientists to get more quality down the pipe. "The real benefit is in increasing the science return and capability for a fixed dollar," Ferraro points out. "If you have a more autonomous spacecraft, you have less need for ground operations. If you have spacecraft that sorts volumes of data and only sends the useful data to the ground, that means your data collection is more efficient and you spend less in operating the spacecraft to retrieve the same amount of data." Another advantage to transferring commercial technology to space is access to commercial software such as compilers and operating systems. To develop a system now, scientists write algorithms for a vehicle and then turn them over to another group who readies the science application for space. With the extra tools that come with COTS, scientists won't have to turn over their algorithms. They can write their own applications and even make changes to the system in flight, which brings scientists closer to their applications. Don't touch that bit Sounds good enough. So why haven't scientist bought into this idea much sooner? Well, they have, to a degree. Even now, every instrument and science collection platform has some computing element. Among the reasons many scientists don't want more computing on board is that they want every bit of the science data from their instruments. Ferraro predicts that when the initial commercial system flies in space, very little money will be saved. Scientists will want to process double—data on board and at home—because they'll want to be sure to retain every bit of data. John Davidson, REE deputy project manager, pinpoints a cultural resistance: Scientist are behind by three decades. In the '60s, the objective "was to deliver flight-qualified, 100-percent-reliable systems. They were sure to make systems armor-protected against cosmic rays—an expensive process. But then, money was no object. We were competing in the race to space." "Still today, there is a consortium of NASA and Air Force programs requesting proposals for a radiation-hardened, next-generation RISC processor. It will be 2001 by the time they ready that processor for use in missions. Processor technology will be three or four years old. At the same time, commercial technology will have moved forward by a factor of 10 beyond what is available in the radiation-hardened form," Ferraro points out. To woo scientists to commercial computing in space, they must be convinced of the advantages. For that reason, REE has developed five science-application teams to work on the ground in a simulated environment in which commercial testbeds are stressed in space-like situations. "Let the scientist determine the best use of this technology; they know better than us," says Ferraro, who hopes to win over scientists in advocating this new technology. You may or may not know Goddard Space Flight Center's (GSFC's) John Mather, Stanford's Peter Michelson, University of Washington's Allan Gillespie, Jet Propulsion Lab's R. Steve Sanders, or GSFC's Steve Curtis. But they are the makers of new applications for COTS in space that range from telescopes to solar terrestrial probes—the people to see in learning the benefits of high-performance computing on board. Science requirements from these teams will be instrumental in the design of the final testbed, which will be completed in September next year. In March, two vendors were awarded a design contract by REE—SEAKR Engineering Incorporated and Sanders, a Lockheed Martin Company. These vendors are developing the testbeds for the simulated space environment on the ground. They were given this product definition: Design a testbed that delivers 30 million operations per second (MOPS) per watt. According to Ferraro, this is about ten times better than the flight system on Mars Pathfinder.
The two vendors addressed the product definition with opposite approaches. While Sanders is leveraging message passing— processors communicate by sending messages down special links, SEAKR is using a shared memory approach in which memory in a parallel computer, usually RAM, can be accessed by more than one processor, usually via a shared bus or network. Whatever the method, the overal computing model needs to demonstrate high bandwidth. That is why REE is considering technology that merges processor logic with memory cells on a single chip, known as processor-in-memory. The value in this approach is that it allows for very-low-power, high-speed accelerators that transfer data from one physical structure in the central processing unit to another. Ultra-low-power components are key to REE, whose goal is to eventually deliver a fault-tolerant, data system, including mass storage capable of operating on less than a watt of electrical power. Laughing at cosmic rays If a testbed is to fly in space, though, the burgeoning data-processing tool must prevail in a bruising environment full of cosmic rays. In a radiation environment, commercial parts have random transient errors that can disable the machine. So the testbed vendors have to propose hardware- and software-based strategies for fault-detection. They must provide a fault-isolation capability so a partially failed architecture can reboot. To accomplish this, they'll be studying the failure mode across the architecture, including the processors, interconnects, memory and glue logic. According to Ferraro, "by reengineering the memory to self-correct single-bit errors, you cover a large percentage of the faults that are likely to occur when a computer system has an upset."
In addition, software-implemented fault tolerance (SIFT) is employed to mitigate the errors that result from not using radiation-hardened parts. SIFT is an application technique in error detection and correction that helps compensate the failure of any single component of the system. In a failure, the system checks the registers and memory and then provides a correction. Using the captured state, it restores the application. Developing a system with near non-stop, error-free operation despite failures is key to REE's research. The outcome of this research on fault-tolerance, which recognizes hardware failures as inevitable, is far from certain. Still, REE hopes to have a flight prototype that can function in the presence of transient errors for about 95 percent of mission needs. Next stop—space As a result of the testing accomplished by the science teams, a legacy, flight-computing tool will lift off sometime in 2003. Having been fully stressed by applications that are data- and CPU-intensive, it should be primed to deliver scaleable, flexible computing that can meet mission-critical needs.
Davidson states, "With large leaps in processing power, we're going to have instruments even smaller and more capable than Sojourner. They won't require ground intervention. They will be capable of solving problems and making decisions in remote space."
|
