Reliability Analysis

The numerical results, calculated according to the Markov chain model, show the significant impact of the mean-time-to-duplication on the whole system reliability, measured in terms of mean-time-to-data-loss, under different workload conditions. As the model indicates, the reliability of CEFT-PVFS depends on the write frequencies of the client nodes. The more frequently the client nodes write data into the storage nodes, the higher the probability that the primary storage group remains inconsistent with the backup group, thus giving rise to increased likelihood of data loss due to the failure of some nodes in the storage group. The write frequency, measured as mean-time-to-write, is highly dependent on the applications running on the client nodes.

To quantitatively compare the reliability of the four duplication protocols, we evaluate their reliability in the scenario of a simple benchmark presented in Section V. Although this simple benchmark does not reflect all applications that run on CEFT-PVFS, it gives a quantitative and fair comparison of these duplication protocols. We recorded the time instants of all the events on all server and client nodes and stored them into the files so that we could calculate the $MTTW$ and $MTTD$ of this simple benchmark. The $MTTD$ of Protocol 1 can be directly calculated from the trace files. The $MTTD$ of Protocol 2 and 4 can be regarded as 0 since the data is consistent as soon as the client node finishes the write process. To obtain the $MTTD$ of Protocol 3 is tricky because the duplication process is performed by the client nodes. In Protocol 3, we define $MTTD$ as the mean time difference between the arrivals of the acknowledgments from the primary node and the backup node.

Figure 15: Reliability comparison of CEFT-PVFS and PVFS

We assume that $MTTF = \mbox{1 year}$, $MTTF_{switch} = \mbox{3 years}$ and $MTTR = \mbox{2 days}$. In the simple benchmark, $MTTW = \mbox{1 minute}$. We calculate the $MTTDL$ curve as a function of the number of server nodes for the four protocols under the three server configurations. Figure 15 compares the reliability between CEFT-PVFS and PVFS and Compared with their $MTTDL$, on the average the four duplication protocols improve the reliability of PVFS by a factor of 41, 64 and 96 in the three server configurations, respectively. In addition, Protocol 1 is $93\%$, $93\%$ and $99\%$ of Protocol 2 and 4 under the three different server configurations, respectively, with an average degradation of $5\%$. Protocol 3 is $96\%$, $94\%$ and $99\%$ of Protocol 2 and 4, with an average degradation of $3.3\%$.