This dissertation covers a broad scope on the studying of performance issues, however, certain issues have not been fully addressed, which become areas for our future investigations.
We believe that our performance model is a general model that covers a broad spectrum of parallel platforms. However, one of the important features of modeling is that models must not be made obsolete by advances in realizing technology. It is interest to see how our model fits to the next waves of technology achievements. For examples, on the networking aspect, we have 10 Gigabit Ethernet [5], WDM optical networks and the InfiniBand architecture [6]; on the system area, we have the massively produced 64-bit microprocessor, VIA communication infrastructure [107] and PCI-X [78]. All these technologies will increase the performance of data communication; however, we reckon that problems induced by congestion and buffering architecture will still have considerable impact on the performance.
Reliable support on lightweight messaging systems needs further studied. Our studies show that the reliable protocol interacts with the network buffering architecture on the congestion development; however, we still lack of information on determining which reliable mechanism is best suit for low-latency high-performance communication. For example, von Eicken et al. [108] has shown that the blame of poor TCP performance is not on the protocol, but on the particular implementation and its integration into the operating system. Their experiments only supported that we may have good TCP performance on light-load communications, but did not provided evidence that the TCP is still effective for handling congestion problem on low-latency communication under heavy congestion. Therefore, it is worthwhile to investigate on how TCP interacts with the network buffering architecture on the congestion development. The results could provide useful information to guide us to achieve better performance on situations under heavy congestion.
Furthermore, in the implementation of the reliable layer on top of DP, we find that one of the hindrances on the performance of user-level reliable support is the delay in delivering the control information. As the underlying network does not differentiate between control and data packets, it handles all packets with the same priority; this could delay some high priority events, e.g. Nack. We are investigating on using existing features of commodity networks, such as priority queues and virtual lanes, to improve the situation. This is importance to the development of congestion control scheme as having faster control information exchanges improves efficiency of the control scheme.
Our approach on tackling the congestion problem works the best for scheduling regular communication patterns on a well-structured, enclosed network. Two possible areas of extensions are the studies of irregular communication patterns and the uses of asymmetric/irregular networks. Another limitation of this thesis study is our optimization technique such as global congestion control is implemented in off-line approach. It is worthwhile to implement it in on-line mode, such as automation through compiler directives and runtime supports. For example, a large-scale cluster may be using in a space-shared mode by several concurrent parallel applications. If we assume that each parallel application is allocated with disjoint set of cluster nodes, how can the congestion control scheme adapt to the resulting irregular configuration and irregular traffic pattern?