A team of researchers from Carnegie Mellon University's School of Computer Science and Meta have redesigned operating systems and hardware to drastically improve memory management in datacenters.

As memory capacity increases in datacenters, virtual memory has become a major performance bottleneck. A vast body of prior work relies on the assumption that an operating system can allocate large, physically contiguous memory to reduce the costs associated with that bottleneck. However, the team identified that in reality, memory contiguity is scarce in datacenters.

With that in mind, the team design a system, called Contiguitas, that provides ample physical memory contiguity and reduces virtual memory overheads. The team tested Contiguitas in production at Meta's datacenters and found it improved end-to-end application performance by 2 to 18%. The research showed that the system is a production-ready and future-proof solution to a long-lasting problem in datacenters.

The research received a Best Paper Award at the 50th International Symposium on Computer Architecture (ISCA), recently held in Orlando, Florida. The team was led by Dimitrios Skarlatos, an assistant professor in the Computer Science Department and head of the CAOS Group, and included Kaiyang Zhao and Ziqi Wang, both Ph.D. students in computer science; Kaiwen Xue, a master's student in computer science; and Dan Schatzberg, Leon Yang, Antonis Manousis, Johannes Weiner, Rikvan Riel, Bikash Sharma and Chunqiang Tang from Meta Platforms Inc.