Achieving Accountable MapReduce in cloud computing
Future Generation Computer Systems
| Hi-index | 0.00 |
Accountability implies that any entity shall be held responsible for its own specific actions and behaviour, so that the entity is part of a larger chain of accountability. PeerReview (Haeberlen et al., 2007) is a software library that provides accountability for distributed systems suffering from Byzantine failure. A key assumption in PeerReview is that a message sent from one correct node to another will be eventually received. In real world environment, however, message loss is commonplace and unavoidable due to the dynamics and uncertainties of the current Internet and it prevents a message from always reaching its destination. Beginning with this point, we have comprehensively analysed the behaviour of PeerReview with the assumption that, eventually, a message will be lost (i.e., never reach the destination). We have demonstrated that PeerReview would be unable to maintain its completeness and accuracy under such circumstantial conditions. Therefore, there is a presentation of six possible errors and the corresponding causes from which they originate. We also propose Message loss Tolerant PeerReview (MLT-PR), which gives a second chance for the falsely indicated nodes to clarify themselves. We re-evaluated PeerReview and MLT-PR with two newly defined metrics, node accountability and system accountability, which are employed to assess the degree of system accountability. Simulation results show that message loss decreases the performance of PeerReview in terms of both metrics and that on average; MLTPR outperforms PeerReview by 31.9% in terms of System Accountability.