Practical Aspects Using a Distributed Queuing System

As more and more workstations interconnected by fast local area networks become available, the popularity of heterogeneous processing increases. On the other hand, it is not easy to spread tasks on machines diiering in software availability, hardware resources, and work-loads. It is hard to keep track of all tasks and even harder to collect their results. In our contribution we present experiences with a queuing system and algorithms for scheduling and for mapping tasks onto various workstations which are located at diierent institutes and are used by hundreds of scientists. Depending on computational speed, number of processors, work-loads, availability, and reliability, the distributed queuing system schedules jobs according to a list of resource requirements. Thereby, availability reeects the fraction of daytime a machine is freed from exclusive usage by its owner. Reliability is a soft factor, responding for frequency of faults, network connectivity and irregular interactive usage. This mapping algorithm distributes tasks depending on the availability of workstations and thus provides a kind of load balancing. Nevertheless, it has enough exibility to provide preferential usage for the workstation owner. To support parallel applications computational equivalents, set up to favor global load balancing on currently available nodes, are introduced. In addition, overlap regions are deened to allow parallel processing on heavily loaded nodes which are expected to be idle in grace. In addition to this more or less traditional batch, queues with speciic application contexts are in use. These queues are resource driven, too. They accept pure application input. Invocation of executables is hidden. This has been proven to be useful for more involved applications, like the set up of a parallel virtual machine.