No abstract

Adda he first issue, L*., to guarantee recovery fine progre&- [7]. For applications which require process autonomy in taking checkpoints in order Abstract-Uncoordinated checkpointing allows process au-to exploit application-dependent information to checkpoint at tonomy and general nondeterministic execution, but suffers the "right time", e.g., when the process state is minimal, lazy from potential domino effects and the associated space overcheckpoint coordination [8] can be incorporated into an uncoor-_ head. Previous to this research, checkpoint space reclama-dinted of tion had been based on the notion of obsolete checkpoints; checkpointing protocol to provide a trade-off between as a result, a potentially unbounded number of nonobsolete coordination overhead and recovery efficiency. checkpoints may have to be retained on stable storage. In Another approach to eliminating the domino effect is to exthis paper, we derive a necessary and sufficient condition ploit the piecesise deterministic execution model [9][10][11], in which ____ for identifying all garbage checkpoints. By using the ap-each process execution is viewed as a number of deterministic proach of recovery line transformation and decomposition, state intervals bounded by noudeterministic events. It has been we develop an optimal checkpoint space reclamation algorithm and show that the space overhead for uncoordinated shown [12] that by considering each nondeterministic event log checkpointing is in fct bounded by N(N + 1)/2 checkpoints as a logical checkpoint [13] taken at the end of the ensuing state WM where N is the number of processes.interval, the same dependency model and hence the checkpoint -Keyword#--fault tolerance, message-pasing systems, unco-space reclamation algorithm developed in this paper can still be ordinated checkpointing, rollback recovery, garbage coilecapplied. tion.Traditionally, checkpoint space reclamation for uncoordinated checkpointing has been based on the notion of obsolete check-

No abstract

M a n y industrial applications consist of different components that were designed and implemented separately. Software reuse and component software architecture often dictate this style of building applications out of existing modules. In such a n application, some modules m a y use databases t o store critical data, while other components m a y use regular files or volatile m e m o r y either by design or t o improve performance. Therefore, global coordination is necessary f o r maintaining data consistency and recoverability across multiple datastores. W e describe a way of integrating checkpointing with transactions t o ensure consistency and recoverability of data stored outside the databases. Our approach transforms a n existing checkpoint library into a transactional resource manager that can participate in such a global coordination.

This paper describes an efficient approach for solving sparse lincar systems using direct methods on a shared-memory vector multiprocessor computer. Parallelism is accomplished by using a Ncsted Bordered Block Diagonal matrix partitioning technique. A Ncsted Block structure is used to represent the sparse matrix, enabling the use of vectorization to achieve high performance. This approach is suitable for many applications that require the repeated direct solution of sparse linear systems with identical matrix structure, such as circuit simulation. The approach has been implemented in a program that runs on an ALLIANT FW8 vector multiprocessor with shared memory. The performance of the program will also be given.