Frequency of Performance Monitoring

Figure 11 shows the impact of the measurement frequency on the aggregate performance for the passive measurement scheme (Figure 11(a)) and the FrequencyCounts active measurement scheme (Figure 11(b)). Each figure plots the results for the five client workloads.

From Figure 11(a) we notice that longer sampling intervals surprisingly offer slightly better performance for passive measurement. To understand this better, consider the curve for the client arrival rate of 10 requests/s. A client arrival rate of 10 requests/s implies that an average of $10T$ connections are made by the clients every $T$ seconds, where $T$ is the sampling interval. However, in order to obtain samples for a fraction $f$ of the 100 destinations over the three providers, the passive measurement scheme will have to force $300f$ connections across the provider links. This leaves a fraction $1-\frac{30f}{T}$ which are not employed for measurement, and could be routed along the optimal provider, assuming that the passive measurement yields reasonably accurate estimate of performance². As $T$ increases, the fraction of connections routed over the optimal path is likely to increase, resulting in a marginal improvement in performance. This explains the slight downward slopes in Figure 11(a).

At the same time, infrequent sampling (i.e., large values of $T$) can have a negative impact on the overall performance. This is not immediately clear from Figure 11(a). However, Figure 11(b), which plots the performance from the FrequencyCounts scheme as a function of the sampling interval, sheds more light on this effect. A sampling interval of 450s suffers a 5-8% performance penalty relative to a smaller interval such as 60s. Notice that in the case of FrequencyCounts too, aggressive sampling (e.g, an interval of 30s) could slightly impact overall performance on some occasions due to the increased software overheads at the proxy.

To summarize, both passive and active measurement offer similar improvements in performance. However, they both have inherent drawbacks. Passive measurement relies on employing client connections for obtaining measurement samples. As a result, if the sampling interval is too small it could result in a significant fraction of connections being routed over suboptimal providers. Active measurement on the other hand has the overhead of transmitting active probe traffic to a fairly large number of destinations. However, the sampling interval could still be small enough as long as the probe traffic does not constitute a significant fraction of all traffic (in terms of the number of connections or connection attempts) from the end network.