CSL Associate Professor Rakesh Kumar has just received a major accolade.
His 2005 paper “Interconnections in Multi-Core Architectures: Understanding Mechanisms, Overheads and Scaling,” co-authored with Victor Zyuban and Dean M. Tullsen, has been
named the winner of the 2020 ACM SIGARCH and IEEE-CS TCCA ISCA Influential Paper Award.
The point of such “test-of-time” accolades is to honor work that doesn’t just dazzle its initial readers, but proves over time that it has significant and lasting value. According to the website of the ACM Special Interest Group on Computer Architecture, the ISCA award “recognizes the paper from the [International Symposium on Computer Architecture] Proceedings 15 years earlier that has had the most impact on the field (in terms of research, development, products or ideas) during the intervening years.”
The paper—which examines area, power, performance, and design issues for on-chip interconnects on chip multiprocessors—also happens to be the most cited paper from the 2005 ISCA.
Why did it resonate so broadly? Kumar explains that the paper made three significant contributions that impacted three separate groups of researchers.
First, the paper offered a previously lacking quantification of the high overheads of interconnection architectures, and how the overhead of the interconnection changes as the number of cores increases and other changes are made.
“The numbers looked quite bad! Essentially we showed that the overhead of interconnection can be very high, and that people need to pay close attention to how they build interconnection architectures,” says Kumar, who is an associate professor of electrical and computer engineering.
That aspect of the work garnered significant attention from circuit designers and architects who were focused on architecture and implementation of interconnection networks.
“We pointed out that a naive implementation, or what was state-of-the-art then—that kind of interconnection won’t cut it. So people did a lot of innovation subsequently on reducing the overhead of interconnection.”
The paper’s second contribution was that it showed that architectures and interconnects need to be co-designed.
“Let’s say that you decide on the architecture of a system and then slap an interconnection on top of it,” says Kumar. “That design is going to be considerably worse than a design where you have thought about the architecture and interconnect from the ground up simultaneously.” Many system architects and chip architects responded to the paper’s strong argument for co-design.
Finally, the paper’s third contribution—which Kumar characterizes as “a much more nitty-gritty kind of contribution”—was its introduction of a methodology for modeling the area, power, and latency overheads of interconnection implementations. That work proved to be helpful to other researchers working on modeling.
“If you look at the papers that cited our paper, some of the most important [ones] were modeling papers,” says Kumar. “People subsequently developed models for the area, power, and performance of different kinds of chips.”
ISCA is widely regarded as the premier conference in computer architecture. The Most Influential Paper Award, which includes an honorarium of $1,000, would normally be presented at the annual conference’s awards lunch. Thanks to COVID-19, though, ISCA 2020 is going to be held virtually. Kumar and his co-authors will be recognized online this year, and the conference organizers tentatively plan an in-person recognition at the 2021 conference.
The ISCA award is the second test-of-time award Kumar has received this year. In January, he was honored with the 10-Year Retrospective Most Influential Paper award at the 2020 Asia and South Pacific Design Automation Conference.