Roaming

Roaming, i.e., a single user moving between different nodes, is an important use of distributed file systems. We expect Pangaea to perform well in non-uniform networks in which nodes are connected with networks of different speeds. We simulated roaming using three nodes: $S$, which stores the files initially and is the server in the case of Coda, and two type-A nodes, $C_1$ and $C_2$. We first run the Andrew-Tcl benchmark to completion on node $C_1$, delete the *.o files, and then re-run only the compilation stage of the benchmark on node $C_2$. We vary two parameters: the link speed between $C_1$ and $C_2$, and the link speed between them and $S$. As seen from Figure 8, the performance depends, if at all, only on these two parameters.

Figure 9 shows the results. It shows that when the network is uniform, i.e., when the nodes are placed either all close by or all far apart, Pangaea and Coda perform comparably. However, in non-uniform networks, Pangaea achieves better performance than Coda by transferring data between nearby nodes. In contrast, Coda clients always fetch data from the server. (Pangaea actually performs slightly better in uniformly slow networks. We surmise that the reason is that Pangaea uses TCP for data transfer, whereas Coda uses its own UDP-based protocol.)

Figure 9: The result of recompiling the Tcl source code. 100Mb/s + 1Mb/s, for example, means that the link between the two client nodes (link (a) in the right-side picture) is 100Mb/s, and the link between the benchmark client and the server (link (b)) is 1Mb/s. The speed of other links is irrelevant in this experiment.