MUTECO: A Framework for Collaborative Allocation in CPU-FPGA Multi-tenant Environments

CPU-FPGA collaborative environments are progressively being adopted by Cloud Warehouses. In this environment, multiple clients share the same infrastructure to maximize resource utilization with energy efficiency and scalability. However, such a provisioning of resources is challenging, since kernels may be concurrently assigned to both CPU and FPGA in a scenario where available resources and workload characteristics drastically vary. To make the best use of resources in this complex environment, we propose MUTECO: A MUlti-TEnant COllaborative resource provisioning framework. MUTECO optimizes considering both multitenancy and CPU-FPGA collaborative execution, in contrast to existing approaches that focus on collaborative single-tenant or non-collaborative multi-tenant workloads. MUTECO is highly configurable and integrated to the Hypervisor layer, so it can be tuned to optimize convergence time, performance, and energy, according to different scenarios that comprise number of tenant requests, the incoming kernels' behavior, and the available resources. Over a varied set of scenarios, MUTECO outperforms in up to 2.91x and 2.39x the current non-collaborative and single-tenant approaches.