A Multi-Agent Elasticity Management Based on Multi-Tenant Debt Exchanges

Abstract

A multi-tenant Software as a Service (SaaS) application is a highly configurable software that allows its owner to serve multiple tenants, each with their own workflows, workloads and Service Level Objectives (SLOs). Tenants are usually organizations that serve several users and the application appears to be a different one for each tenant. However, in practice, multi-tenant SaaS applications limit the diversity of tenants by clustering them in a few categories (e.g. premium, standard) with predefined SLOs. Additionally, this coarse-grained clustering reduces the advantage of these multi-tenant ecosystems over single tenant architectures to share dynamically virtual resources between tenants based on their own workload profile and elasticity adaptation decisions. To address this limitation, we propose a multi-agent elasticity management where each tenant is represented by a reinforcement learning agent that performs elasticity adaptations based on a new technical debt perspective, and make use of debt attributes (i.e. amnesty, interest) to form autonomous coalitions that minimise the effect of the unavoidable imperfections in any elasticity management approach. We extended CloudSim and Burlap to evaluate our approach. The simulation results indicate that our debt-aware multi-agent elasticity management preserves the diversity of tenants and reduces SLO violations without affecting the aggregate utility of the application owner.