Optimal Number Of Checkpoints Research Articles

Performance optimization of parallel-distributed processing with checkpointing for cloud environment

In cloud computing, the most successful application framework is parallel-distributed processing, in which an enormous task is split into a number of subtasks and those are processed independently on a cluster of machines referred to as workers. Due to its huge system scale, worker failures occur frequently in cloud environment and failed subtasks cause a large processing delay of the task. One of schemes to alleviate the impact of failures is checkpointing method, with which the progress of a subtask is recorded as checkpoint and the failed subtask is resumed by other worker from the latest checkpoint. This method can reduce the processing delay of the task. However, frequent checkpointing is system overhead and hence the checkpoint interval must be set properly. In this paper, we consider the optimal number of checkpoints which minimizes the task-processing time. We construct a stochastic model of parallel-distributed processing with checkpointing and approximately derive explicit expressions for the mean task-processing time and the optimal number of checkpoints. Numerical experiments reveal that the proposed approximations are sufficiently accurate on typical environment of cloud computing. Furthermore, the derived optimal number of checkpoints outperforms the result of previous study for minimizing the task-processing time on parallel-distributed processing.

Optimizing the Level of Confidence for Multiple Jobs

Correct operation of real-time systems (RTS) is defined as producing correct results within given time constraints (deadlines). As RTS are becoming more susceptible to soft errors, employing fault-tolerant techniques is crucial. Roll-back Recovery with Checkpointing (RRC) is an efficient fault-tolerant technique. However, RRC introduces a time overhead which depends on the number of checkpoints. The imposed time overhead may cause deadline violations. Therefore, it is important at design time to have a metric to evaluate to what extent a time constraint is met such that RRC can be optimized. In our previous work, we introduced the usage of Level of Confidence (LoC), i.e., the probability to meet a given deadline, and showed for a single job that there exists an optimal number of checkpoints which results in the maximal LoC. In this paper, we assume given is a deadline and a set of jobs that employ RRC, and the objective is to find the optimal checkpoint assignment that maximizes the LoC. We show that our previous work is not sufficient for multiple jobs. Therefore, we derive an expression to compute the LoC and propose an efficient method to maximize the LoC for multiple jobs.

Low power fault tolerance scheduling scheme in DVS-enabled real-time systems

This paper investigates the power management and fault tolerance in DVS-enabled real-time systems based on the EDF scheme. The power management is achieved via dynamic voltage scaling DVS and fault tolerance is achieved via checkpoint scheme. We estimate a critical speed which effectively decreases the energy consumption by the associated frequency dependent and independent devices. The proposed low power fault tolerance scheduling scheme includes two parts: 1 an offline solution to compute the optimal number of checkpoints and the running speed; 2 an online speed reduction mechanism based on the actual workload by using the slack time. The experimental results show that the proposed approach which is compared to the previous method Melhem et al., 2004 significantly reduces the energy consumption.

Elastic full-waveform inversion using the second-generation wavelet and an adaptive-operator-length scheme

Elastic full-waveform inversion (FWI) can generally yield subsurface model parameters of high spatial resolution. However, the computational and storage burden is tremendous. To mitigate this problem, we developed the following three strategies: First, we derived a time-space-domain adaptive staggered-grid finite-difference method. The size of the calculation was reduced by using smaller operator lengths for high S-wave velocities. Second, we evaluated a multiscale elastic FWI scheme based on the second-generation wavelet, by which scale decompositions of data and model space were implemented. With the scale increasing, spatial sampling intervals and time steps became larger. Therefore, the minimum computational cost could be achieved by combining the variable-operator-length scheme with the multiscale scheme. Third, to economize the memory consumption, we extended the efficient boundary storage and checkpointing schemes into elastic FWI. Only wavefields of several parts in the boundary area between two adjacent checkpoints were stored and used to reconstruct source wavefields of the inner area. Furthermore, the optimal number of checkpoints could be automatically determined by minimizing the amount of storage. We validated these new strategies for the SEG/EAGE overthrust model. The synthetic example suggested the feasibility and robustness of the new elastic FWI method in terms of improving computational efficiency and alleviating large amounts of data storage. We also analyzed the accuracy of our FWI method. The inversion results revealed that our wavelet-based method had better reconstructed structures and higher convergence rate than the conventional filtering-based multiscale FWI method.

Checkpoint Management with Double Modular Redundancy Based on the Probability of Task Completion

This paper proposes a checkpoint rollback strategy for real-time systems with double modular redundancy. Without built-in fault-detection and spare processors, our scheme is able to recover from both transient and permanent faults. Two comparisons are conducted at each checkpoint. First, the states stored in two consecutive checkpoints of one processor are compared for checking integrity of the processor. The states of two processors are also compared for detecting faults and the system rolls back to the previous checkpoint whenever required by logic of the proposed scheme. A Markov model is induced by the fault recovery scheme and analyzed to provide the probability of task completion within its deadline. The optimal number of checkpoints is selected so as to maximize the probability of task completion.

Optimal Checkpoint Placement on Real-Time Tasks with Harmonic Periods

This paper presents an optimal checkpoint strategy for fault-tolerance in real-time systems where transient faults occur in Poisson distribution. In our environment, multiple real-time tasks with different deadlines and harmonic periods are scheduled in the system by rate-monotonic algorithm, and checkpoints are inserted at a constant interval in each task. When a fault is detected, the system carries out rollback to the latest checkpoint and re-executes tasks. The maximum number of re-executable checkpoints and an equation to check schedulability are derived, and the optimal number of checkpoints is selected to maximize the probability of completing all the tasks within their deadlines.

On the Complexity of Removing Z-Cycles from a Checkpoints and Communication Pattern

Communication-induced checkpointing protocols are mechanisms used to produce checkpoints and communication patterns which enjoy desirable properties, such as No-Z-Cycle (NZC). NZC guarantees that each checkpoint can be part of a global consistent checkpoint. It would be nice to define communication-induced checkpointing protocols that enforce NZC, adding a minimum number of checkpoints to remove all the Z-cycles from the distributed computation. In this paper, we prove that this is impossible by formulating the minimum Z-cycle removal (MinZCR) problem and showing that there are no online competitive protocols for it. Moreover, we prove that the problem of enforcing NZC with an optimal number of checkpoints is difficult even if the whole input instance is known because its decision version is NP-complete. Finally, we also prove that MinZCR is difficult to approximate: it is APX-hard and this implies that no polynomial time approximation scheme exists for the problem

Global checkpointing for a concurrent processing system

In global checkpointing, N (N  1) checkpoints are maintained since Faults might not be detected on occurrence owing to the error latency time. The time required for rollback and recovery is therefore dependent on error latency, the number of checkpoints maintained, the intercheckpoint interval and the way checkpoints are selected for rollback. This paper presents a model for describing the error latency time distribution, which is pivotal in approximately determining the point of occurrence of a fault, given that it is delected. Two rollback schemes, sequential and lookahead, are investigated. Although the sequential rollback is easier to implement, the lookahead rollback has a smaller error recovery cost. Numerical results are also given for a system with two processes. The optimal number of checkpoints to be maintained by the system to minimize the cost of rollback for each of the rollback strategies is given.

Selection of a checkpoint interval in a critical-task environment

The selection of an optimal checkpointing strategy has most often been considered in the transaction processing environment where systems are allowed unlimited repairs. In this environment an optimal strategy maximizes the time spent in the normal operating state and consequently the rate of transaction processing. This paper seeks a checkpoint strategy which maximizes the probability of critical-task completion on a system with limited repairs. These systems can undergo failure and repair only until a repair time exceeds a specified threshold, at which time the system is deemed to have failed completely. For such systems, a model is derived which yields the probability of completing the critical task when each checkpoint operation has fixed cost. The optimal number of checkpoints can increase as system reliability improves. The model is extended to include a constraint which enforces timely completion of the critical task. >