The convergence of P2P and Grid Computing Foster and Iamnitchi IPTPS 2003 Alan Sussman CMSC 818S March 13 2007 Notes Project schedule by Thursday groups ideas Postpone Ninja paper until after break and after I get back from IPDPS Death taxes etc Grid and P2P computing target the same general problem organization of resource sharing within virtual communities And take the same general approach create overlay structures that exist on top of underlying organizational structures Grid computing addresses infrastructure but not yet failure whereas P2P addresses failure but not yet infrastructure Grids vs P2P Grids sharing environments persistent standards based services to share resources within distributed communities resources owned by physical organizations shared via locally defined policies what who and when a virtual organization VO P2P applications that take advantage of resources at the edges of the Internet unstable connectivity network addresses etc so no reliance on central servers create overlay networks in implementations Target communities incentives Grids motivated by needs of scientific users needing access to remote resources datasets and or computers users willing to devote effort to create and operate infrastructure some degree of trust and accountability VOs can t impose solutions on local owners P2P more grass roots file sharing and clientserver cycle sharing large scale in number and diversity of users need to detect and deal with resource misuse Resources Grid resources can be more powerful diverse and better connected than P2P resources so need explicit administration according to well defined policies QoS important but increases integration cost leads to needing information services to determine resource properties Availability tends to be higher and more uniform Grids can integrate desktop resources too but such resources are aggregated into administrative domains e g Condor to integrate into Grid Most P2P resources less powerful home or desktop computers Applications Grid apps very diverse across communities remote access to supercomputers desktop grid apps virtual laboratories remote instrument access P2P systems often vertically integrated file sharing cycle sharing BOINC is an exception diversity comes from different design goals scalability anonymity availability etc Grid apps often data intensive high network disk throughput rates but P2P apps aren t even file sharing maybe not true for P2P streaming video if that ever happens Scale and failure Scientific Grids modest scale in number of participants institutions pooled computers simultaneous users so scalability not a big issue until recently since centralized solutions sufficed P2P communities can consist of millions of nodes for file sharing and cycle sharing so need robust self management of large numbers of nodes But total amount of activity in P2P systems not quite as big as it seems 1 2TB per day in file sharing systems at end of 2001 3 4TB per day in D0 experiment Grid example Services and infrastructure Grid community has created persistent multipurpose infrastruture services e g authentication authorization discovery resource access data movement etc overlays on resources and services maintained by participating institutions but still enforce local policies P2P systems have focused on integrating simple resources e g files cycles on individual computers protocols provide specific vertically integrated services each P2P app has its own protocols and interfaces to its services Summary Both Grid and P2P have the same vision a worldwide virtual computer where access to resources and services can be obtained when and where needed Need to combine strengths of both to address failure using scalable self configuring protocols provide persistent and multipurpose infrastructure supported in an organized distributed way to achieve robustness performanced and trust
View Full Document
Unlocking...