324 CHAPTER 13 Switching system is a scalable 256-by-256 switching system. Cisco Networks has recently announced its own version, the CRS-1 Router. 13.9.1 Measuring Switch Cost Before studying switch scaling, it helps to understand the most important cost metrics of a switch. In the early days of telephone switching, crosspoints were electromagnetic switches, and thus the N 2 crosspoints of a crossbar were a major cost. Even today this is a major cost for very large switches of size 1000. But because crosspoints can be thought of as just transistors, they take up very little space on a VLSI die. 4 The real limits for electronic switches are pin limits on ICs. For example, given current pin limits of around 1000, of which a large number of pins must be devoted to other factors, such as power and ground, even a single bit slice of a 500-by-500 switch is impossible to package in a single chip. Of course one could multiplex several crossbar inputs on a single pin, but that would slow down the speed of each input to half the I/O speed possible on a pin. Thus while the crossbar does indeed require N 2 crosspoints (and this indeed does matter for large enough N), for values of N up to 200, much of the crosspoint complexity is contained within chips. Thus one places the largest crossbar one can implement within a chip and then one interconnects these chips to form a larger switch. Thus the dominant cost of the composite switch is the cost of the pins and the number of links between chips. Since these last two are related (most of the pins are for input and output links), the total number of pins is a reasonable cost measure. More refined cost measures take into account the type of pins (backplane, chip, board, etc.) because they have different costs. Other factors that limit the building of large monolithic crossbar switches are the capacitive loading on the buses, scheduler complexity, and issues of rack space and power. First, if one tries to build a 256-by-256 switch using the crossbar approach of 256 input and output buses, the loading will probably result in not meeting the speed requirements for the buses. Second, note that centralized algorithms, such as iSLIP, that require N 2 bits of scheduling state will not scale well to large N. Third, many routers are limited by availability requirements to placing only a few (often one) ports in a line card. Similarly, for power and other reasons, there are often strict require- ments on the number of line cards that can be placed in a rack. Thus a router with a large port count is likely to be limited by packaging requirements to use a multirack, multichassis solution consisting of several smaller fabrics connected together to form a larger, composite router. The following subsections describe strategies for doing just this. 13.9.2 Clos Networks for Medium-Size Routers Despite the lack of current focus on crosspoints in VLSI technology, our survey of scalable fabrics for routers begins by looking at the historically earliest proposal for a scalable switch fabric. Charles Clos first proposed his idea in 1955 to reduce the expense of electromechanical switching in telephone switches. Fortunately, the design also reduces the number of compo- nents and links required to connect up a number of smaller switches. It is thus useful in a present-day context. Specifically, a Clos network appears to be used in the Juniper Networks T-series multichassis router product, introduced 47 years later, in 2002. 4 However, in the wheel of time, the number of crosspoints again may begin to matter for optical switches!