A Scalable Super-Peer Approach for Public Scientific Computation

Many types of distributed scientific and commercial applications require the submission of a large
number of independent jobs. One highly successful, and low cost mechanism for acquiring the necessary
compute power is the ``public-resource computing'' paradigm, which exploits the computational power
of private computers. Recently decentralized peer-to-peer and super-peer technologies have been
proposed for adaptation in these systems. We designed a super-peer protocol for the execution of jobs
based upon the volunteer requests of workers, and a super-peer overlay for performing two kinds of
matching operations: the assignment of jobs to workers and the download of input data needed for
job execution. This paper analyzes a dynamic and general scenario, in which: (i) workers can leave the
network at any time; (ii) each job is executed multiple times, either to obtain better statistical accuracy or
to perform parameter sweep analysis; and, (iii) input data is replicated and distributed to multiple data
caches on-the-fly. A simulation study was performed to analyze the super-peer protocol and specifically
evaluate performance in terms of execution time, utilization of data centers, load balancing, and ability
to efficiently scale with the number of jobs and the network size.