Pricefair: On fair scheduling of heterogeneous resources

We consider the problem of all-or-nothing resource allocation for tasks with diverse resource demands in resource-constrained heterogeneous systems. We observed that allocation approaches based on max-min fairness or admission control, despite their desirable properties, may fall short of satisfying task demands and utilizing the full potential of the heterogeneous system. This becomes particularly clear in high-demand scenarios where certain resource types may be insufficient to satisfy all task demands. We present Pricefair, a new all-or-nothing resource allocation mechanism for task demands in heterogeneous systems. The key contribution of Pricefair lies in its approach to high-demand scenarios when specific types of resources are exhausted. Rather than denying those tasks, Pricefair tries to allocate the surplus of that type of resource to less-loaded, better-performing, although more expensive, nodes in the cluster. In such instances, to maintain fairness, Pricefair employs a dynamic pricing model. By using the concept of price parity, Pricefair ensures that all users with resources in the exhausted tiers share the price increase and pay proportionally based on their resource requests. Through our experimental evaluation in realistic, high-demand scenarios, we demonstrate the effectiveness and benefits of Pricefair.