Decentralized Task Allocation in a Particle Physics Computing Network
Benjamin Smith, Harvard University
Particle physicists are faced with the task of performing reconstruction — turning the output of a particle detector into the types and properties of particles in an event — and analysis — turning events into a scientific result. As both of these are very computationally intensive, the particle physics community has developed a tiered hierarchical computing network. Currently, jobs are assigned based upon type, with each node in the network responsible for one or more job types. This paper presents a decentralized task allocation algorithm based upon the threshold model used in insect task allocation that is capable of dynamically balancing the load on nodes in the network without the use of a centralized controller. The use of this algorithm results in a 50% improvement in network performance in a simulated simplified ATLAS computing network.
Abstract Author(s): Ben Smith