Analysis of PLR for Shared Switch Fabrics

Abstract

A modular upgrade design for packet transport switch nodes is presented where packet loss is dramatically reduced by intra‐module and inter‐module buffer sharing. This modular design offers significant cost and power reduction in a high‐data‐rate system where buffers are highly costly and power‐greedy.

Add-drop Benes network for scalable optical packet networks

“…As described in [13], shared switch fabric enhances the PLR dramatically. However, the previous analytic model has assumed every switch fabric module as strictly nonblocking network.…”

“…A node architecture model of shared modular switch fabrics and its gain from the points of views of system cost, power consumption [1], and packet loss rate (PLR) [13] are presented. Figure 2 illustrates our shared ADBN node model.…”

“…where P(X) indicates the probability of X number of packet contentions [13] , and j x is the number of contended packets at the j-th ADBN. Sub-index (ik) is wrapped around by modulo of N. ( , ) ( , , )…”

“…For this example, we consider three different policies where the neighbor transfer cut-off is set at the first, second, or third neighbor (k = 1, 2, or 3) in a system with W>3. When we limit neighbor transfer to the first, second, and third, Table 2 shows numbers of required buffer-drop ports equivalent to those of 2, 3, and 4-shared switch fabrics of the analytic model in [13], respectively. As observed by the comparisons with the previous analytic model, it is clear that the connection capability limitations of the ADBN have negligible effect on buffer counts by proposed algorithm and architectural rules.…”