PA-RPC: Introducing Payload Awareness to Improve Data Center Efficiency

RPC layer accounts for considerable fleet-wide CPU cycles in today's data centers. We observe that existing HW/SW systems to accelerate RPC are developed based on high-level characteristics of RPCs, which are insufficient to identify many optimization opportunities. Therefore we build the real-world workload generator twrk, and perform a comprehensive, payload-aware characterization study of RPCs using DeathStarBench socialNetwork microservices and demonstrates the potential performance improvements brought by the payload-awareness introduced in this work. Further, we propose payload-aware architecture and system optimizations that can leverage the redundancy patterns illuminated in our characterization to accelerate RPCs.