Sinclair Focuses on Heterogeneous Systems as a Way Forward After Moore’s Law; New Award Recognizes His Work
7/9/2018 4:27:24 PM
As a PhD student at Illinois Computer Science, Matt Sinclair focused on using heterogeneous systems as a potential path to greater efficiency as the benefits of transistor scaling slow.
The recent graduate’s work has been recognized as part of the first-ever group of ACM SIGARCH/IEEE CS TCCA Outstanding Dissertation Awards.Sinclair (PhD CS ’17) was awarded an honorable mention in June “for enabling an efficient global address space for heterogeneous systems.” The award is sponsored by SIGARCH, the Association for Computing Machinery’s special interest group on computer architecture, and TCCA, which is the Institute of Electrical and Electronics Engineers’ committee focused on integrated hardware and software design.
Sinclair said he is humbled by the recognition for work that he likened to “eating your vegetables” – important, complex, but not always the first thing people are eager to talk about or take on.
“I kind of have to pinch myself every now and then. It’s humbling to get rewarded like this in front of your research community,” said Sinclair, who is currently working as a postdoctoral researcher at AMD Research and will become an assistant professor at the University of Wisconsin-Madison this fall.
Moore’s Law has been a defining factor in computer architecture since the 1960s. But its impending end drove Sinclair, like a lot of people in his field, to think about what follows.
“It’s not something we can rely on in the long term, to continue improving performance and energy (efficiency) like it has in the past,” Sinclair said. “I view the future of computing architecture as being heterogeneous.”
Heterogeneous systems offer the possibility of enhanced performance and energy efficiency by combining different types of processors, making use of the strengths of each to handle particular tasks. But that combination of different processors comes with its own set of challenges.
For one thing, as Sinclair points out in his thesis, current heterogeneous systems “are inefficient for emerging applications, like graph analytics workloads, that have fine-grained synchronization, relaxed atomics, and more general sharing patterns.”
Heterogeneous systems are also hard to program. Different kinds of processors are often programmed in different languages and can access data in different ways, decreasing their efficiency when combined in one system.
“Existing systems have different kinds of processors that work well in isolation but are less efficient when they need to communicate frequently with one another,” Sinclair said.
Trying to capture some of that potential efficiency, Sinclair set out to redesign the memory hierarchy of heterogeneous systems. First, he used a unified address space between the different processors to improve programmability. Next, he looked at how to design more efficient memory organizations, and then he focused on designing a more efficient cache coherence protocol. Finally, he defined a new memory consistency model that makes it easier for programmers to write programs for these systems.
Sinclair’s new memory consistency model builds on the groundbreaking work of his advisor, Professor Sarita Adve. Early in her career, Adve developed the data-race-free memory consistency model that became the standard for most major programming languages.
Already, Sinclair says, his ideas have generated interest from the hardware industry, the C++ community, and the Heterogeneous System Architecture Foundation.
“We’re finding these ideas are starting to influence both the way that programming languages and hardware are being designed,” he said. “That gives me hope that we’ve come up with something really impactful.”
Sinclair was quick to credit two of Professor Adve's other students with whom he collaboarted on papers that comprised his thesis, Rakesh Komuravelli (MS CS '10, PhD '14) and John Alsop, who is currently pursuing a PhD in Electrical and Computer Engineering.
Sinclair is also grateful to Adve.
“I’m definitely building on the shoulders of giants,” he said.