4.5 Summary

In this chapter, we analyze and model the congestion loss problem experienced on today's high-speed commodity network. Two different analytical models for input-buffered and output-buffered architectures are constructed and analyzed. Through the modeling process, we study how different buffering architectures, as well as different settings of our Go-Back-N reliable protocol, affect on the congestion dynamic. To show our contributions, we organize our results around areas on:

1. Of the modeling aspect, we show that the deterministic model provide more accessible information than the stochastic model, but requires explicit delineation of relationships. However, stochastic model is more powerful and flexible, despite the needs of statistical expertise. Through the comparison on our two performance models, we point out that the behavioral different between the two buffering architectures on the congestion loss problem lies on how they interact with the reliable protocol.
2. In the analysis arena, we show that our study on the congestion loss behavior of the output-buffered case can explain other architectural situations and communication scenarios. In addition, we find that under the many-to-one traffic, input-buffered architecture has a higher threshold on the overflow problem; however, once the overflow situation occurs, the performance suffers significantly (both measured and predicted results showed that we could only achieve less than 50% of the bandwidth under congestion loss).
3. Of the contention studies, we find that buffering architecture has a significant impact on the congestion behavior, as a result, some measures on congestion control may find to be effective under one architecture but futile in others. In particular, the addition of the Stall state over classic GBN scheme has a positive impact on the congestion performance under the output-buffered case, but shows no effect on the input-buffered case. Besides, we find that the larger buffer capacity associated to a switch port, the better congestive loss performance we have on the output-buffered case.
4. Of the protocol design issue for cluster computing, we show that with many-to-one bulk data transfers, setting a longer timeout duration (e.g. $B_{L}*\exp <\overline{TO}\leq B_{L}*\exp ^{2}$) help relieving the congestion problem with our GBN reliable protocol. Besides, we should take great care on the selection of the flow control window size, as its setting pays a critical role on the whole performance spectrum, i.e. over all traffic conditions. In particular, one should make use of the available information on the buffering architecture and capacity, the communication pattern and the number of participating nodes to derive the corresponding settings. This shows that we need to have a global prospective, rather than a simple end-to-end view when designing communication protocols.

The above results are valuable information for us to devise effective strategies to handle the congestion loss problem. However, the best answer to the congestion problem is we should try to prevent any packet loss. This is because, no matter how efficient the congestion recovery protocol is, the performance still suffers.