Job Dispatching and Scheduling for Heterogeneous Clusters -- A Case Study on the Billing Subsystem of CHT Telecommunication

Many enterprises or institutes are building private clouds within their own data centers. Data centers may have different batches of physical machines due to annual upgrades, but the number of machines is fixed most of the time. Consequently it is crucial to schedule jobs with different resource requirements and characteristics to meet different job timing constraints, in such heterogeneous yet most of the time static environments. This paper describes a cloud resource management framework that dynamically allocates and reallocates computation resources for jobs that have different requirements, including deadline and priority. This framework makes decisions according to specified policies, and the framework provides four default policies for system administrators to choose to fit their specific needs. The framework is designed to be component-pluggable. The components of the framework can be hot-swapped, i.e., Replaced without shutting down the services. In addition, the framework can work as an individual cloud computing system, or as an extension of an existing cloud system. Our experiment results demonstrate that our system is capable of dynamically adjusting the resource allocation plan according to run-time statistics collected. The system also tolerates hardware failures, and will dynamically reallocate workers to compensate for the downtime in order to finish the jobs before deadline. Our experiments also suggest a trade-off between priority and deadline.