Files in this item
|(no description provided)|
|Title:||Checkpointing in distributed virtual memory by utilizing local virtual memory|
|Author(s):||Hadi, F.X. Nursalim|
|Doctoral Committee Chair(s):||Campbell, Roy H.|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||This study explores a recovery strategy using checkpointing in a distributed shared virtual memory (DVM) system. DVM shares virtual memory in a loosely-coupled multi-computer system and is implemented at the software-level. The goal of this recovery strategy is to obtain a consistent recovery line that is close to the time of failure. Therefore the system could be rolled back from the time of failure to the closest possible state of normal execution.
In order to achieve the objective, this thesis proposes a checkpointing strategy that utilizes virtual memory (VM) as transient checkpoint storage in addition to commonly-used stable storage. In controllable checkpoint intervals, these additional checkpoints make checkpoint intervals shorter; in turn making the recovery line closer to the time of failure. Compared to the cost of taking checkpoints to stable storage, taking these additional checkpoints does not cost much since they are saved to virtual memory. This thesis will show that the additional cost of these transient storage checkpoints is very low, while the benefit of reducing the rollback cost is very high.
The utilization of VM will be applied to commonly-used independent checkpointing and coordinated checkpointing strategies. The checkpointing protocols of both strategies are changed to accommodate additional checkpointing to VM. This thesis will show that the modified protocols still guarantee state consistency after recovery.
Simulations on trace data and experiments on the Choices operating system are conducted to measure the performance of the proposed checkpointing strategies. We compare independent checkpointing strategies with and without VM utilization; we also compare coordinated checkpointing strategies with and without VM utilization. The simulations and experiments demonstrate that in the independent checkpointing strategy, utilizing of VM reduces rollback costs with only a small fraction of additional checkpoint costs. The same result also applies to the coordinated checkpointing strategy utilizing VM.
|Rights Information:||Copyright 1995 Hadi, F. X. Nursalim|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9543601|