Scaling the Network

In this section we study CUP performance as we scale the size of the network.

Table 2 compares CUP and PCX for network sizes between $2^{7} = 128$ and $2^{14} = 16384$ nodes for a Poisson $\lambda$ rate of 1 query per second. The first row shows the CUP miss cost as a fraction of the PCX miss cost. The second and third rows show the average query latency in hops for PCX and CUP respectively. The number in parentheses is the standard deviation. As can be observed, CUP reduces average query latency respectively by 9.77, and 17.81, and 26.39 hops for the 4096, 8192, and 16384 node networks. This is a substantial reduction in average query latency that improves with increasing network size. Comparing the standard deviations of CUP and PCX we see that CUP also has less variability around its average query latency.

The fourth row in Table 2 shows the IR per overhead push performed by CUP. We observe a growth in the rate of return with 16.14, 24.85, and 35.98 for the last three network sizes. These numbers are quite strong, considering that the overhead is completely recovered.

Figure 3 shows the IR of CUP versus network size for Poisson with $\lambda$ = 1, 10, 100, and 1000 queries per second. From the figure we see that for a particular network size, if we increase the query rate the IR increases, and for a particular query rate, if we increase the network size, the IR also increases. This demonstrates that CUP scales to higher query rates and higher network sizes.

Figure 3: IR vs. net size. (Log-scale axes.)