Optimization criteria for checkpoint placement

Abstract

Mean response time and availability as optimization criteria for checkpoint placement are better replaced by workable formulas that calculate the ratio between the marginal gain accrued to users who experience system failure and the (presumably slight) loss suffered on average by all users. OPTIMIZATION CRITERIA FOR CHECKPOINT PLACEMENT C. M. KRISHNA, KANG G. SHIN, and YANN-HANG LEECheckpointing is becoming increasingly popular in real-time and database systems as a means of mitigating the consequences of failure. Checkpointing involves storing authenticated process state information that can then be used in the event of failure to restart the affected computation from the latest checkpoint, instead of from the beginning. There are a number of articles in the literature dealing with the optimal number and placement of checkpoints [1-6, 9, 10]. Our purpose here is to address an important point that has been overlooked; that is, although the analyses presented are often ingenious and elegant, their practical validity is thrown into doubt by an inappropriate choice of optimization criteria. We propose alternative criteria that we believe to be of greater practical relevance.In every published analysis of optimal checkpointing that we have seen, the only optimization criteria used are mean response time and/or availability. No explicit justification for this is ever given. A carelessness in choosing optimization criteria can result in inadequate and misleading performance analyses, for, in a real sense, optimization criteria are relative, not absolute. The simple act of choosing a particular criterion imposes a bias on the results that follow. By definition, This work has been supported in part by NASA under Grant NAG 1-296. Any opinions here expressed are those of the authors and do not necessarily represent the views of NASA. All correspondence regarding this paper should he sent to Kang G. Shin. ©1984ACM0001-0782/84/lOO0-1008 75¢optimization criteria specify the commodity that is of importance and therefore to be optimized. Optimization criteria have a very subtle influence on the way systems are viewed. In a sense, they are languages through which we seek to convey system performance. We know from experience with natural languages that these affect not only the way in which ideas are expressed but also the very ideas themselves. Optimization criteria are no exception. They manipulate our view of system behavior, contorting it to fit a prefabricated mold. For this reason, the choice of optimization criteria determines the practical usefulness of the results that are then derived.Optimization criteria are performance measures. To choose them correctly, it is important to determine what it is we wish to express. In this case, where we wish to measure the advantages that accrue from checkpointing, it is best to trade the benefits derived from them against the overhead they impose: in other words, to carry out a value analysis. First, we shall show why mean response time and availability do not do so effectively and ...

Configuring Fault-Tolerant Servers for Best Performance

“…In line with earlier works [13,5,15,7,3] we assume that a failure is detected as soon as it occurs. We also assume that no failures occur during a failover [15,5,3].…”

“…The failure interarrival time distribution is also exponential (similar to [15,9,7,5]). The average failure rate is ¦ © .…”

“…Most results are obtained for settings where a longrunning job performing heavy calculations, is checkpointed from time-to-time [15,7,8,3,9]. Other models include request processing or message logging systems [5,13,10].…”

Performance and effectiveness analysis of checkpointing in mobile environments

“…These algorithms need some a priori knowledge about the processing time. In [12], Krishna et al point out that the optimization criteria for checkpoint placement is to trade the benefits derived from checkpointing with the overhead imposed. Ziv and Bruck [24] present an online algorithm for placement of checkpoints.…”

“…Note that the checkpoint creation time C is a random variable; however, it is fixed for a given repair and rollback process [16]. Following similar steps in Lemma 1 we can obtain the Equation (11) and (12).…”