View source for Contention-adaptive scheduling of MTC applications on multi-core infrastructures

{{Projectproposal
|Contact person=Thilo Kielmann
|Contact person2=Stefania Costache
|Master areas=
|Fulfilled=Yes
}}
Many-task computing (MTC) applications, composed of large numbers of tasks with small execution times and various resource requirements, are becoming more common. Such applications can be found in the High Performance Computing (HPC) and data analytics workloads. Scheduling these applications on current multi-core cloud infrastructures is difficult not only due to changes in per-task resource requirements, tasks might be CPU or I/O intensive, but also due to variations in resource availability, e.g, network bandwidth. The commonly used approach is to run application tasks on a fixed number of slots per node, leading to inefficient resource usage and poor application performance.
The goal of this project is to study the characteristics of different MTC workloads (HPC and data analytics) and propose on-line scheduling algorithms to efficiently run tasks with different resource requirements on the same node while considering resource contention. The algorithms will be implemented and evaluated in existing resource managers for such workloads, i.e., YARN or Sparrow.