Author:
Levente Filep
Affiliation:
Faculty of Mathematics and Computer Science, Babeș-Bolyai University, Mihail Kogălniceanu, Cluj-Napoca, Romania
Keyword(s):
Peer-to-Peer Networks, Super-Peer Topology, Distributed Computing, Middleware.
Abstract:
Regarding large-scale scientific computing, many alternative solutions to Cloud Computing Services exits, which combine existing, cheap, commodity hardware into computational clusters. The majority of these, due to their ease of deployment, are based on Client-Server architecture. Decentralized approaches employ some form of Peer-to-Peer (P2P) design, however, due to their increased design complexity, and without major benefits over the Client-Server ones, none of these systems gained wide popularity. The P2P system presented in this paper features decentralized task coordination, the possibility of suspending, migrating and resuming workload on different nodes, employs remote checkpoints to allow partial result recovery, and workload tracking, which offers the possibility to initiate communication between them. Design considerations and choices for this system are presented and discussed. The chosen topology is super-peer managed clusters arranged in an extended start topology and e
valuated by simulation. Such a system comes with enormous design complexity; however, a middleware can hide these complexities, while providing the applications a simple interface to access network resources. Harnessing idle computing resources, the system can be deployed on a combination of in-house computer networks, personal and volunteer devices, as well as Cloud-based VMs.
(More)