Performance and network economy at a large scale

The previous section demonstrated Pangaea's ability to fetch data from a nearby source and distribute updates through fast links, yet only at a small scale. This section investigates whether these benefits still hold at a truly large scale, by using a discrete event simulator that runs Pangaea's graph-maintenance and update-distribution algorithms. We extracted performance parameters from the real testbed we used in the previous section, and ran essentially the same workload as before. We test two network configurations. The first configuration, called HP, is the same as Figure 10, but the number of nodes in each LAN is increased eighty-fold, to a total of 3000 nodes. The second configuration, called U, keeps the size of each LAN at six nodes, but it increases the number of regions to 500 and connects regions using 200ms RTT, 5Mb/s links.

Figures 14 and 15 show average file-read latency and network bandwidth usage in these configurations. These figures show the same trend as before, but the differences between the configurations are more pronounced. In particular, in the HP configuration, Pangaea propagates updates almost entirely using local-area network for popular files, since it crosses over wide-area links only a fixed number of times, regardless of the number of replicas. In the U configuration, Pangaea still saves bandwidth, more visibly when many replicas exist. The systems cannot improve read latency much in U, because most of the accesses are forced to go over wide area links, but Pangaea still shows improvement with many replicas.

Figure 14: File-reading latency in a simulated 3000-node system. The meaning of the numbers is the same as in Figure 11.

Figure 15: Wide-area network bandwidth usage during file updates in simulated 3000-node systems. The meaning of the numbers is the same as in Figure 12.