While today’s large-scale data centers such as those run by Amazon, Facebook, Google, and Microsoft are built using commodity off-the-shelf servers, recently there has been an increasing trend towards server customization to reduce costs and improve performance. One such trend is the advent of “rack-scale computing”. We use this term to refer to emerging architectures that propose servers or rack-scale computers comprising a large number of tightly integrated systems-on-chip, interconnected by a high-bandwidth and low-latency network fabric. In the near future, we expect to see "rack-scale computers" with 1,000s of cores and terabytes of memory, connected with bandwidth and latency comparable to today's smaller-scale NUMA servers.
These new architectures challenge traditional assumptions and raise several interesting research questions, ranging from hardware architectures and system design to programming languages. In this talk, I will highlight some of these challenges and discuss some recent efforts that we have been undertaking at Microsoft Research Cambridge to design a new network stack for rack-scale computers.