Collaborative Research: FoMR: Taming the Instruction Bottleneck in Modern Datacenter Applications

Data centers are the power plants that drive the digital economy. These warehouses filled with acres of computers, operated by the likes of Google, Facebook and Amazon, run the websites and apps that we use every day. Yet, as vast as these computing resources are, the amount of code they are tasked with running is larger still and growing at around 20% each year. This project will explore new hardware and software mechanisms to allow large data center programs to better fit within their limited computing resources, to improve data center program performance and energy efficiency, and reduce the toll that power-hungry data centers take on our planet. This project will unfold in three stages. The first stage will leverage new hardware mechanisms for profiling program execution, to obtain an accurate picture of how a data center program is running at the hardware level. The second stage will use that profiling information to adjust the code of the program at runtime to make better use of on-chip caches. The third stage will use the profiling information from the first stage, integrated with new hardware support for code prefetching and code-aware cache replacement, to drive cache utilization even higher than in stage two. Taken together, these optimizations will accelerate data center programs and improve data center energy efficiency. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.