"We thought that in software systems, when you release an OS and find bugs later on, the manufacturer sends patches and new revisions," he said. "If the processor doesn't work when it executes a certain instruction ... why not send a patch? It would allow you to fix the things in the field without having to recall the chips."
After more than two years of research at the University of Illinois, Urbana-Champaign, was born "Phoenix" -- hardware that sits on a computer chip to detect and help a processor recover from design defects that can cause processor crashes, data corruption, computational mistakes and processor hangs, Torrellas said.
Unlike software, where manufacturers can send out patches as often as needed to fix bugs that are discovered after shipment -- think Microsoft's patch Tuesdays -- processors have become so complex that there are no easy ways to repair design flaws in the field without recalling them or disabling a feature. Phoenix, however, allows for defects to be addressed in the field, according to Torrellas.
"[Manufacturers] could do less testing and ship the processor earlier than they would otherwise, knowing that if there were problems you could fix them later," he said. "Shipping a chip earlier is very advantageous because you beat the competition."
Torrellas began his project by researching 10 processors from Intel, AMD, Motorola and IBM to compile lists of known defects published by the manufacturers. He then began devising a hardware prototype that would allow his research team to detect these problems and repair them on the fly, he said.
"We put hardware in the chip that reads the state of the processor when it is running," he said. "When it finds a set of conditions that would imply that soon there will be a crash or a corruption of the data, we change the path of the processor so that it doesn't encounter these conditions."
Essentially, Phoenix taps into key logic signals and then -- based on downloaded defect signatures -- the hardware flags defects and then "flushes" those signals before a problem can occur. Much like the patching process with software, when a hardware manufacturer discovers a new design flaw, Phoenix can be programmed to add that flaw to its list of defect signatures, Torrellas said. So far, he said the system has been able eliminate two-thirds of known bugs completely; Torrellas is researching ways to attack the remaining flaws.
The system, which Torrellas said has gotten positive feedback from a few chip manufacturers who have read about the research, requires the addition of "a few wires and some logic" that can be added to a chip inexpensively. In addition, the system does not affect the performance of the chip, he said.
Computerworld (US online)
