Tanvir Ahmed Khan awarded Rackham Predoctoral Fellowship
CSE PhD candidate Tanvir Ahmed Khan has received a Rackham Predoctoral Fellowship to support his work in optimizing data center processors at the University of Michigan.
To serve billions of users across the planet, modern web applications have to access huge datasets with complex application logic. This scale leads to two major issues for their operation in data centers: poor data access patterns and inefficient instruction fetch operations. Khan has devised a wide range of optimization techniques to address these bottlenecks, making use of a profile-guided approach and software-hardware codesign.
Khan’s work, which he titles “Rescuing Data Center Processors,” is driven by the observation that both data access and instruction fetch in data center applications follow a deeply repetitive pattern. This has allowed him to “profile” these processes and provide predictive performance boosts based on their past behavior. He’s applied this profile-guided design to a number of efficiency-boosting instruction-fetch technologies, including instruction prefetching (I-SPY) and replacement (Ripple), branch target buffer prefetching (Twig) and replacement (Thermometer). All of these techniques have been shown to boost the speed of data center applications, and a number of them are being adopted by real-world data center processor designers like Intel.
In addition to these performance optimizations, Khan has also worked on systems that repair bad data access patterns in real time with profiling. Huron is a system for automatically detecting and repairing false sharing, a poor data access pattern that hurts the performance of parallel programs. DMon finds and repairs general data access issues with a new approach called selective profiling, enabling in-production detection of poor data access patterns and guiding compiler optimizations to automatically improve the corresponding pattern. Finally, APT-GET automatically repairs poor data access patterns using profile-guided timely prefetch operations.
Khan’s research on data center applications’ performance optimizations has appeared in top computer architecture and systems venues like ISCA, MICRO, FAST, EuroSys, PLDI, and OSDI. His performance analysis methodology made real-world workloads twice as faster and consequently was adopted in the Arm Neoverse N1 Core, the leading data center CPU powering 49% of all Amazon Web Service machines.
Khan is advised by Prof. Baris Kasikci.
About the Rackham Predoctoral Fellowship
The Rackham Predoctoral Fellowship supports outstanding doctoral students who have achieved candidacy and are actively working on dissertation research and writing. They seek to support students working on dissertations that are unusually creative, ambitious and risk-taking.