The problem of scheduling G groups of jobs on m unrelated parallel machines is considered. Each group consists of several identical jobs. Implicit in a schedule is a partition of a group into batches; all jobs of a batch are processed jointly. It is possible for diierent batches of the same group to be processed concurrently on diierent machines. However, at any time, a batch can be processed o...

The most popular framework for distributed training of machine learning models is the (synchronous) parameter server (PS). This paradigm consists n workers, which iteratively compute updates model parameters, and a stateful PS, waits aggregates all to generate new estimate parameters sends it back workers iteration. Transient computation slowdowns or transmission delays can intolerably lengthen...