The efficiency of NUMAchine's interconnection network can be shown using a number of performance metrics. Figure 17 depicts the utilization of the station buses, local rings and central ring. It indicates that none of these components is likely to become a performance bottleneck. Figure 8 shows the delays in ring interfaces. Each vertical bar shows two components of the delay. The lower portion of each bar corresponds to the minimum delay (in the absence of network traffic) and the upper portion indicates additional delay due to traffic contention. The average packet delays in the upward and downward paths in the local ring interfaces are shown in Figure 4.5. The upward path delay is small for all applications. The larger delays for the downward paths are due to the way in which we have implemented the packet handler and the queues, which we are currently redesigning to reduce these delays. Packet delays from the central ring to a local ring have the same small delays as for the upward path in Figure 4.5. The average packet delays from a local ring to the central ring are only slightly larger, as shown in Figure 4.5. This indicates that for our targeted system size the rings are expected to perform well.