An Improved Multiple Faults Reassignment based Recovery in Cluster Computing

In case of multiple node failures performance becomes very low as compare to single node failure. Failures of nodes in cluster computing can be tolerated by multiple fault tolerant computing. Existing recovery schemes are efficient for single fault but not with multiple faults. Recovery scheme proposed in this paper having two phases; sequentially phase, concurrent phase. In sequentially phase, loads of all working nodes are uniformly and evenly distributed by proposed dynamic rank based and load distribution algorithm. In concurrent phase, loads of all failure nodes as well as new job arrival are assigned equally to all available nodes by just finding the least loaded node among the several nodes by failure nodes job allocation algorithm. Sequential and concurrent executions of algorithms improve the performance as well better resource utilization. Dynamic rank based algorithm for load redistribution works as a sequential restoration algorithm and reassignment algorithm for distribution of failure nodes to least loaded computing nodes works as a concurrent recovery reassignment algorithm. Since load is evenly and uniformly distributed among all available working nodes with less number of iterations, low iterative time and communication overheads hence performance is improved. Dynamic ranking algorithm is low overhead, high convergence algorithm for reassignment of tasks uniformly among all available nodes. Reassignments of failure nodes are done by a low overhead efficient failure job allocation algorithm. Test results to show effectiveness of the proposed scheme are presented.

💡 Research Summary

The paper addresses the problem of performance degradation in cluster computing when multiple node failures occur. Existing fault‑tolerance schemes are largely designed for single‑node failures and become inefficient under multiple faults. To overcome this limitation, the authors propose a two‑phase recovery framework that combines a sequential load‑balancing phase with a concurrent job‑assignment phase, both driven by a dynamic rank‑based algorithm.

In the sequential phase, the system first gathers the current load of every operational node. Each node is then assigned a rank: the node with the smallest load receives rank 1, and the node with the largest load receives the highest rank N. A rank table is constructed (Algorithm 1). The load‑redistribution procedure (Algorithm 2) pairs the highest‑ranked node with the lowest‑ranked node, computes the average load of the pair, and transfers the excess load from the heavier node to the lighter one. The process repeats, moving inward through the rank list, until the loads of all nodes converge to a near‑uniform value. Because each iteration only requires communication between two nodes, the algorithm incurs minimal messaging overhead and converges in far fewer iterations than traditional pairing, tree‑based, or partitioning schemes.

After the sequential balancing completes, the concurrent phase handles two kinds of work: (a) tasks that were originally assigned to failed nodes, and (b) newly arriving jobs. The algorithm continuously checks for pending failure‑node jobs or new jobs (while loop). For each pending job, it selects the node with the lowest current rank (i.e., the most idle node) and dispatches the job to that node (Algorithm 3). This phase runs in parallel with normal computation, allowing the system to absorb new work without disrupting the already balanced state.

The authors implemented the scheme on a small testbed consisting of four homogeneous machines, each running an MPI “hello world” program. They compared three scenarios: (1) a naïve uniform distribution without the rank‑based scheme, (2) the proposed rank‑based approach, and (3) a baseline MPI distribution. The results (Table 3 and Figure 3) show that the proposed method reduces response time by roughly 30–40 % relative to the naïve approach, and the load after balancing is almost equal across all nodes, eliminating the need for further adjustments.

A literature review situates the work among several existing load‑redistribution techniques, including the Match‑maker algorithm, tree‑based redistribution, subset partitioning, the Most‑to‑Least‑Loaded (M2LL) policy, and binary‑tree domain decomposition. The authors argue that their pairing of highest‑rank with lowest‑rank nodes yields a higher “splitting ratio,” meaning that a larger portion of the load imbalance is corrected per iteration, which directly translates into faster convergence and lower communication cost.

The paper acknowledges two primary limitations. First, the design assumes a homogeneous cluster where all nodes have identical processing capability; extending the method to heterogeneous environments would require weighting ranks by node capacity. Second, the rank table is maintained globally, which could become a scalability bottleneck in very large clusters due to synchronization overhead. The authors suggest future work on weighted rank calculations, asynchronous rank updates, and predictive load modeling to address these issues.

In conclusion, the proposed dynamic rank‑based load redistribution combined with a least‑rank job assignment mechanism offers an effective, low‑overhead solution for multiple‑fault recovery in cluster computing. It achieves rapid convergence, minimal messaging, and significant reductions in response time, while remaining simple to implement and transparent to end users. The approach is well‑suited for homogeneous clusters and provides a solid foundation for further enhancements targeting heterogeneous and large‑scale systems.