In this paper, we demonstrate the feasibility and scalability of building a considerably reliable storage system with multi-terabyte capacity without any additional hardware cost in a cluster while maintaining high I/O performance to alleviate the I/O bottleneck for parallel applications.
Four different duplication protocols are designed to implement the
fault tolerance in the parallel file system. Our experimental
results and analytical analysis lead us to conclude that these
protocols can improve the reliability over the original PVFS 40
times while degrading the peak write performance only around
in the best case, and around 
in the worst case when
compared with PVFS with the same total number of servers. In
addition, these duplication protocols strike different balances
between reliability and write performance. A protocol that has
higher bandwidth is most likely to be inferior in reliability.
Between the server-driven protocols, the asynchronous one achieves
a write performance that is 
higher than the synchronous
one, which comes at the expense of an average 
reliability
degradation. Similarly, between the client-driven protocols, the
asynchronous one has a write performance that is 
higher
than the synchronous one, while paying a premium of an average
reduction in reliability. We also proposed a hybrid
protocol that optimizes the tradeoff between the reliability and
write performance. In this hybrid protocol, if the total number of
jobs of a data-intensive application is less than the server
number of one storage group, the synchronous server duplication is
used to mirror the data. Otherwise, the asynchronous client
duplication is preferred.
None of the proposed protocols employs high-cost but more reliable techniques such as ``forced writes'' to the disks, and thus can potentially lose data if a disk or node fails while data is being copied from the I/O buffer (cache) on the processor to the disk. We will further investigate the tradeoff when considering ``forced writes''.