Dataset for Hare

Client-facing services are latency-sensitive and thus require consistently low response times to attract and retain users. We identify that these applications usually have varying object sizes, dynamic workloads, and complex query-processing functions. Switch-based hotspot offloading is a trendy solution for latency-sensitive applications to achieve high system throughput with an acceptable P99 query response latency. However, the existing switch-based dynamic hotspot offloading approaches such as NetCache and DySO suffer from low switch resource utilization, so they cannot offload large hotspots or achieve fast and precise offloading. Besides, they are specialized for particular systems, so we cannot directly generalize their designs to other applications. To this end, we propose Hare, a systematic framework that consists of three invocations to address these issues. First, Hare uses a MAT-based cross-stage structure called e-MAT to store and perform hit-checks for large hotspots on the switch data plane. Second, Hare uses a switch-server co-offloading mechanism to support fast and precise offloading. Third, Hare enables generally automatic offloading by decoupling application-related query processing with hotspot offloading. Compared to state-of-the-art approaches, Hare supports 8.86× ∼ 9.97× larger hotspot size, achieves 1.27× ∼ 6.61× higher system throughput, and can recover the system throughput and the P99 query response latency within 8s.