Exploiting Time-malleability in Cloud-based Batch Processing Systems


Existing cloud provisioning schemes allocate re- sources to batch processing systems at deployment time and only change this allocation at run-time due to unexpected events such as server failures. We observe that MapReduce-like jobs are time- malleable, i.e., at runtime it is possible to dynamically vary the number of resources allocated to a job and, hence, its completion time. In this paper, we propose a novel approach based on time-malleability to opportunistically update job resources in order to increase overall utilization and revenue. To set the right incentives for both providers and tenants, we introduce a novel pricing model that charges tenants according to job completion times. Using this model, we formulate an optimization problem for revenue maximization. Preliminary results show that compared to today’s practices our solution can increase revenue by up to 69.7% and can accept up to 57% more jobs.

In ACM SIGOPS Workshop on Large-Scale Distributed Systems and Middleware (LADIS) co-located with SOSP
Luo Mai
Luo Mai
Assistant Professor

My research interests include computer systems, machine learning and data management.