Execution-driven simulation of network storage systems

A number of new network storage architectures have emerged recently that provide shared, adaptable and high-performance storage systems for data-intensive applications. Three common storage networking architectures are direct-attached storage (DAS), network-attached storage (NAS), and storage area network (SAN). Efficient implementations of each of these classes of storage architecture can have a significant impact on overall system performance. To be able to tune both the performance of a network storage architecture and its underlying workload, an accurate simulation modeling environment can be very valuable. In this paper we present ParIOSim, a validated execution-driven simulator for network storage systems. This simulator can be used to accurately predict the performance of parallel I/O applications as a function of the underlying storage architecture. ParIOSim also provides a flexible simulation environment to guide system level storage optimizations. To evaluate the accuracy of our simulation environment, we compare the performance of ParIOsim to the performance obtained on an actual parallel system. To demonstrate the utility of ParIOSim, we report on the overall system performance obtained for a parallel I/O benchmark application as run on different storage architectures.