Enhancing Data Locality in a Fully Decentralised P2P Cycle Stealing Framework

Peer-to-peer (P2P) networks such as Gnutella and BitTorrent have revolutionised Internet based applications. P2P approaches provide a number of benefits, however most cycle stealing projects, such as SETI@home, have concentrated on centralised methods which still require massive amounts of concentrated network bandwidth in order to scale. More recent P2P research has developed the concept of distributed hash table (DHT) P2P overlays. These overlays provide efficient and guaranteed message delivery unlike earlier P2P networks which relied on large scale replication to probabilistically find data. Our G2:P2P framework makes use of a DHT overlay to provide a fully decentralised P2P cycle stealing system. Its distributed object programming model allows direct communication between objects and it remains reliable even as the set of peer nodes changes. In this paper we describe extensions to G2:P2P which allow us to optimise object distribution for locality. The importance of optimising data locality is well understood and has received extensive research, however, in the context of cycle-stealing systems and more generally DHT based P2P networks it is completely unexplored. Whilst our work is motivated by parallel programming, it is generic in nature and may have applicability to other DHT applications.