O Passado Também Importa: Um Mecanismo de Alocação Justa de Múltiplos Tipos de Recursos ao Longo do Tempo

Shared computing systems are composed by different resource types, such as CPU and memory, and hold users with different resource constraints. While some users execute short workloads in which fast allocation is essential, others execute long workloads that require more resources. Among the different proposals to allocate resources in this scenario, Dominant Resource Fairness (DRF) is notable for satisfying some desirable properties, such as truthfulness and Pareto efficiency. However, these proposals focus only on instantaneous fairness, ignoring users heterogeneity. This paper proposes DRF with state (SDRF). SDRF satisfies the fundamental properties of DRF, besides enforcing a new notion of fairness that look at past resource allocations. We verify SDRF with both theoretical analysis and simulations using Google cluster traces. Results show that SDRF reduces users' average waiting time and improves fairness by increasing the number of completed tasks for users with lower demand with low impact on high-demand users.