Performance of STC104 vs. Telegraphos

We consider continuous routing on nn switches with various buuering schemes. By extending these models to networks, we analyze two 1024-node, unique-path, butterry-like networks conngured with input-output buuered STC104 32 32 switches, or central buuered Telegraphos 4 4 switches. 0 Introduction Parallel and distributed systems consist of processors communicating over an interconnection network conngured with switches and network interfaces. The n n input buuered switch (IDB) has n FIFO buuers. The packet to be transmitted through an output port is selected from all input buuers in round robin fashion. Any packet directed to a full buuer is stalled. The n n split input buuered switch (SIDB) has n 2 FIFO buuers, one for each input-output port combination. The SIDB switch can preroute a packet to determine if it is directed to a full buuer. If all packets corresponding to a given output port are directed to a full buuer, traac is temporarily stalled. Otherwise, the packet to be transmitted is selected using round robin. The nn output buuered switch (ODB) implements FIFO buuering at the output ports. All packets arriving simultaneously can be wired to the appropriate output buuer(s) within T s clock cycles. The ODB switch also uses prerouting to determine whether packets are directed to a full buuer. Such packets are temporarily stalled. The n n central buuer switch (CDB) provides one dedicated buuer for each source-destination pair in the network. Flow control is per virtual circuit (VC), where VC header information identiies uniquely the packet's source and destination. Round robin is implemented at each output port, for selecting the packet to be transmitted. We assume xed packet size, preventive ow control (no packet loss) and similar switch startup time (T s clock cycles), spent in buuering and prerouting before Supported by ANM and PRS of the French CNRS and EU ESPRIT 6253/SHIPS.