A Reconfigurable Fabric for Large-Scale Datacenter Services
Recorded 11 February 2015 in Lausanne, Vaud, Switzerland
Event: IC Colloquia - EPFL IC School Colloquia
The ending of Moore’s Law is starting to have a profound impact on datacenter operators, who have traditionally relied on steady advances in processor performance and efficiency to make improved services economically viable. The Catapult reconfigurable fabric at Microsoft ushers in a new datacenter architecture that marries programmable software with efficient and low-power programmable hardware (i.e., FPGAs). Catapult is deployed on a bed of 1,632 servers and was shown to double the ranking throughput of a large-scale web search workload (Bing).
Exploiting a large-scale reconfigurable fabric such as Catapult required radical changes to conventional software-hardware contracts, programming methodology, and coordination between hardware and software teams at Microsoft. Some challenges encountered were: (1) the need to accommodate rapid changes to models and algorithms, (2) the need for flexible but well-defined contracts and interfaces between high-level software and low-level accelerator engines, and (3) mitigating the productivity challenges associated with programming low-level accelerators. This talk will present some of the solutions used in the Bing pilot, and offer suggestions and directions for future work.
Watched 5151 times.Watch