Chip, Heal Thyself

Computerworld - In 1994, Intel Corp. shipped its new Pentium processor with a tiny flaw in a region of the chip devoted to floating-point division. The company shrugged off the problem as extremely minor, but users screamed bloody murder, and Intel eventually recalled the chips, at a cost of $475 million.

Someday, such a financial and public relations disaster might be avoided by equipping chips with Tortola, a technology being developed by a young researcher at the University of Virginia in Charlottesville. Indeed, Intel and IBM are collaborating with computer science professor Kim Hazelwood as she works to create a way for users to download hardware fixes as easily as they do software patches.

Hazelwood says engineers and computer scientists have traditionally sought to optimize the performance of computer systems by concentrating on just one layer at a time  hardware, operating system, application software and so on. But she thinks the solution to a growing number of problems in microprocessors is to build a better bridge between software and hardware and to treat the two symbiotically. She calls this bridge a virtual-execution environment, or VEE.

Heres how Tortola works: Hardware sensors (which often already exist on the chip) send information about things such as temperature and voltage anomalies to the VEE, which sits between the hardware and the executing software. The VEE is programmed to react to these signals and to alter the executable code as it runs. For example, if it sees a temperature spike, it might substitute some sequence of instructions known to be less taxing to the processor. It could also respond to performance problems such as memory cache misses or resource contention by reallocating the workload among chip resources.

Hazelwood has proved the concept on a well-known problem in which certain patterns of instructions cause chip circuits to turn on and off rhythmically and rapidly. That might happen, for example, when the same sequence of instructions is executed repeatedly in a program loop. This rapid switching can degrade the reliability of the processor and lead to incorrect results.

Hazelwood has shown that the VEE can sense when this is occurring by watching for voltage spikes and dips on the chip and then judiciously inserting no-op instructions  place-holder instructions that perform useless computations  to break up the on-and-off pattern. The VEE can also unroll an offending loop to lengthen the time between voltage spikes and dips to an acceptable frequency. The altered application code is stored in the VEE and is used instead of the original code as the program runs.