Data center resource management for in-network processing

Nowadays there exist several In-Network Processing (INP) solutions that allow tenants to improve their application performance in terms of different metrics: Daiet inventors claim to achieve an 86.9%-89.3% traffic reduction by performing data aggregation entirely in the network data plane. Other solutions like NetChain and IncBricks let programmable switches store data and process queries to cut end-to-end latency; CloudMirror allows client applications to specify bandwidth and high availability guarantees. For the time being, it seems that there is still no valid resource allocation algorithm that takes into account the presence of a network having a data plane that supports (partially or completely) INP. This thesis has mainly two goals: (i) model and evaluate an Application Programming Interface (API) through which applications can ask for INP resources and (ii) discuss the importance of a scheduler which can reject INP requests and propose their server-only equivalent when needed (e.g., high switch utilization).