Kumar receives ARO Young Investigator Award for error-resiliency research
Kim Gudeman, Coordinated Science Lab
- Kumar's the technique has already proven workable on polynomial-time applications, and will now be developed for other applications.
- ECE Assistant Professor Rakesh Kumar has received the Army Research Office's (ARO) Young Investigator Award.
- The award will fund the development of algorithmic techniques to make applications robust to numerical errors.
ECE Assistant Professor Rakesh Kumar has received the Army Research Office’s (ARO) Young Investigator Award. The award, which honors young researchers who “show exceptional promise for doing creative research,” will fund the development of algorithmic techniques to make applications robust to numerical errors.
The goal is to enable transformed applications to make forward progress and deliver acceptable outputs in spite of errors. Kumar, a researcher in the Coordinated Science Lab, is also working to employ the error-resilient techniques on applications whose traditional implementations are not error-tolerant, such as sorting and graph matching.
The approach is unique in that it does not correct errors, but continues to produce acceptable results in spite of them. In addition, many error-resilient techniques are aimed at hardware, while Kumar’s research focuses on software solutions. By alleviating unnecessary error corrections at the hardware level, the technique could significantly reduce power consumption – a primary focus of today’s computer architects.
“If you try to fix all the hardware faults, you are correcting a lot of unnecessary problems, because not all hardware faults end up affecting software,” said Kumar. “If you let all these faults propagate to the software side, applications can make good decisions about which faults are tolerable.”
The research builds upon the work that earned Kumar and his students the Best Paper in Session at SRC TECHCON in 2011. The group will continue to advance their goals by recasting applications with discrete outputs as optimization problems, which are tolerant of numerical errors.
The challenge is that there is no way to automate error correction, which means that Kumar may have to create new algorithms for every class of applications. In addition, the technique cannot overcome certain errors.
“You still have to make some parts of the code robust,” Kumar said. “For example, the algorithm isn’t going to make glue logic robust. It will still require classical error-resiliency techniques, which were also developed here at CSL.”
However, the technique has been proven to work on a very large class of applications, called polynomial-time applications. Kumar says the method also works on other classes, though he cannot yet prove it.
The Army is interested in potentially using techniques like this to guard against injection-based security attacks. The goal is to make code so robust that it continues to achieve the desire outcome in spite of potential breaches.
Kumar says the 3-year, $150,000 ARO award would not have been possible without the hard work of his students.
“I’m particularly proud of this award because it’s a recognition of all our hard work in an area that we got into after coming to Illinois,” he said. “I’m also very happy for my student Joseph Sloan, who leads this work and constantly makes our group look good.”