Evaluating Dynamic Resource Management for Bulk Synchronous Parallel Applications

Bulk Synchronous Parallel (BSP) applications comprise distributed tasks that synchronize at periodic intervals, known as supersteps. Efficient resource management is critical for the performance of BSP applications, especially when deployed on multi-tenant cloud platforms. This project evaluates and extends some existing resource management algorithms for BSP applications, while focusing on dynamic schedulers to mitigate stragglers under variable workloads. In particular, a Dynamic Window algorithm is implemented to compute resource configurations optimized over a customizable timeframe by considering workload variability. The algorithm applies a discount factor prioritizing improvements in earlier supersteps to account for increasing prediction errors in future supersteps. It represents a more flexible approach compared to the Static Window algorithm that recomputes the resource configuration after a fixed number of supersteps. A comparative evaluation of the Dynamic Window algorithm against existing techniques, including the Static Window algorithm, a Dynamic Model Predictive Control (MPC) algorithm, and a Reinforcement Learning (RL) based algorithm, is performed to quantify potential reductions in application duration resulting from enhanced superstep-level customization. Further evaluations also show the impacts of window size and checkpoint (reconfiguration) cost on these algorithms, gaining insights into their dynamics and performance trade-offs.

Degree: MS Project Defense (CS)