Dimitrios Skarlatos, an assistant professor in Carnegie Mellon University's Computer Science Department, has received three research awards from Meta, totaling $200,000, that will support his research group's work to improve the scalability, programmability and security of datacenter computer systems.

"Current hardware and operating system abstractions and interfaces were built at a time when we had minimal security threats, homogeneous compute, scarce memory resources and a few users. These assumptions are not representative of today's computing landscape," said Skarlatos, who leads the School of Computer Science's Computer Architecture and Operating Systems (CAOS) group. "In this new era of computing, it is urgent that we rethink the synergy between the OS and hardware layers from scratch."

Meta awarded Skarlatos $100,000 to develop methods to safeguard operating systems against speculative execution attacks in datacenters, which prey on CPUs and trick them into accessing and leaking privileged information. These attacks pose a major threat for datacenter environments. Skarlatos' work provides new abstractions and interfaces between hardware and the operating system to eliminate speculative vulnerabilities at their root.

Skarlatos also received $50,000 from Meta to work on redesigning virtual memory for datacenters. Existing virtual memory mechanisms are decades old and were designed for processors with a few megabytes of main memory — a stark contrast to the terabytes of today's datacenter-grade processors. With processor hardware facing scalability limits and memory-intensive datacenter workloads such as machine learning becoming dominant, virtual memory has become a major performance bottleneck. Skarlatos seeks to completely rethink virtual memory across the operating system and hardware, and eliminate virtual memory overhead. 

Lastly, Meta awarded Skarlatos $50,000 to codesign hardware and software for the next generation of datacenter-scale AI systems. Heterogeneous hardware such as graphics processing units and other specialized accelerators improve the performance of machine learning inference and training workloads in today's datacenters. But they pose a major programmability challenge for these environments. Skarlatos' work will provide a new abstraction across the hardware architecture and the operating system to holistically integrate accelerators into the system stack.

Learn more about the winners of Meta Research Awards on the company's blog.