Switch matrix

BACKGROUND

0001<patcit id="pcit0001" dnum="US2007121499A"><text>US 2007121499</text></patcit> discloses a data networking technique and apparatus using a physically distributed but logically shared and data-sliced synchronized shared memory switching data path architecture integrated with a distributed data control path architecture. This provides an output-buffered switching of data in networking systems, such as routers and switches, to support the increasing port densities and line rates with maximized network utilization and with per flow bit-rate latency and jitter guarantees. This can be achieved while maintaining optimal throughput and quality of services under all data traffic scenarios, and with features of scalability in terms of number of data queues, ports and line rates, particularly for requirements ranging from network edge routers to the core of the network. The technique results in complete non-blocking data switching between ingress and egress ports, under all circumstances and scenarios.

0002The invention relates to a switch system for connecting a configurable number of ports together.

0003A switch has a number of resources, such as ports, buffers and internal busses. These resources will often depend on each other (e.g., the usage of some resources typically requires the availability of others). At a certain time, a buffer at each input port typically holds packets to several destinations, while packets to a certain destination are held in different input buffers. If an input buffer has a single read port, and this read port is busy feeding data to one output port, another output port might have to wait idle until the read port of the input buffer becomes available.

0004Different switch implementations can use different protocols ((e.g., PCI Express, Ethernet, InfiniBand, etc.). It is desirable to manage packet transfers such that it is possible to transfer the highest number of packets at the same time irrespective of the protocol being used.

0005The present invention seeks to provide an adaptable switch system that is configurable to different protocols.

SUMMARY

0006The invention is defined in the claims

0007An embodiment of the invention can provide a switch matrix with a configurable, protocol agnostic, packet switching core, the switch matrix providing a non-blocking switch connecting a configurable number of ports together.

0008An embodiment of the invention can provide a switch system for connecting a configurable number of ports together, the switch comprising a plurality of switch matrix units and a switch interconnect interconnecting the switch matrix units, at least one said switch matrix unit providing a port connection to a port and providing a plurality of virtual output queues for respective destination ports, each virtual output queue being configured to buffer received packets to be transmitted to a respective destination port.

0009An embodiment of the invention can provide a method of operation a switch system for connecting a configurable number of ports together, the method comprising: providing a plurality of switch matrix units and a switch interconnect interconnecting the switch matrix units; providing port connections to the ports via respective switch matrix units; and buffering packets from the ports to be transmitted to destination ports in respective virtual output queues in the switch matrix unit, each virtual output queue being configured to buffer received packets to be transmitted to a respective destination port.

0010Although various aspects of the invention are set out in the accompanying independent and dependent claims, other aspects of the invention include any combination of features from the described embodiments and/or the accompanying dependent claims, possibly with the features of the independent claims, and not solely the combinations explicitly set out in the accompanying claims.

BRIEF DESCRIPTION OF THE DRAWINGS

0011Specific embodiments are described by way of example only with reference to the accompanying Figures in which: <ul id="ul0001" list-style="none" compact="compact"><li><figref idref="f0001">Figure 1</figref> is a schematic representation of the rear of an example switch chassis;</li><li><figref idref="f0002">Figure 2</figref> is a schematic representation of the front of the example switch chassis;</li><li><figref idref="f0003">Figure 3</figref> is a schematic representation of a midplane illustrating the logical connectivity through the midplane between cards at the rear and cards at the front orientated orthogonally with respect to each other;</li><li><figref idref="f0004 f0005">Figure 4</figref>, formed of <figref idref="f0004">Figures 4A</figref> and <figref idref="f0005">4B</figref>, is a schematic diagram of an example management infrastructure;</li><li><figref idref="f0006 f0007 f0008 f0009 f0010 f0011 f0012">Figures 5 to 11</figref> are views of an example of a switch chassis;</li><li><figref idref="f0013">Figure 12</figref> is a first isometric view of an example of a midplane;</li><li><figref idref="f0014">Figure 13</figref> is a further isometric view of an example of a midplane;</li><li><figref idref="f0015">Figure 14</figref> is an isometric view of an example of a line card;</li><li><figref idref="f0015">Figure 15</figref> is an isometric view of an example of a fabric card;</li><li><figref idref="f0016">Figure 16</figref> is schematic representations of part of a switch chassis;</li><li><figref idref="f0016">Figure 17</figref> is a further schematic representation of part of a switch chassis;</li><li><figref idref="f0017">Figure 18</figref> is a schematic representation of the connections of two cards orthogonally with respect to each other;</li><li><figref idref="f0017">Figure 19</figref> is a schematic representation of an example of orthogonally arranged connectors;</li><li><figref idref="f0018">Figure 20</figref> is a schematic side view of one of the connectors of <figref idref="f0017">Figure 19</figref>;</li><li><figref idref="f0018">Figure 21</figref> is a plan view of an example configuration of vias for the orthogonal connector pairing of <figref idref="f0017">Figure 19</figref>;</li><li><figref idref="f0019">Figure 22</figref> is a cross-section through of a via;</li><li><figref idref="f0019">Figure 23</figref> is a schematic side view of example of an alternative to the connector of <figref idref="f0018">Figure 20</figref>;</li><li><figref idref="f0020">Figure 24</figref> is a schematic end view of an example cable connector;</li><li><figref idref="f0020">Figure 25</figref> is a schematic side view of the example cable connector;</li><li><figref idref="f0020">Figure 26</figref> represents a footprint of the cable connector;</li><li><figref idref="f0021">Figures 27 and 28</figref> illustrates example of signal routing for a cable connector;</li><li><figref idref="f0022">Figure 29</figref> illustrates an example of a power supply for the cable connector;</li><li><figref idref="f0022">Figure 30</figref> illustrates an example of cable status sense detection circuitry;</li><li><figref idref="f0023">Figure 31</figref> illustrates an example of hot plug control circuitry;</li><li><figref idref="f0024">Figure 32</figref> is a schematic representation of airflow though a switch chassis;</li><li><figref idref="f0025">Figure 33</figref> is a schematic block diagram of an example 6 port switch matrix;</li><li><figref idref="f0026">Figure 34</figref> is a schematic representation of a switch matrix unit of the switch matrix of <figref idref="f0025">Figure 33</figref>;</li><li><figref idref="f0027">Figure 35</figref> is a schematic block diagram of an example virtual output queue of the switch matrix unit of <figref idref="f0026">Figure 34</figref>;</li><li><figref idref="f0028">Figure 36</figref> is a schematic representation of an example organization for packet category storage in a virtual output queue;</li><li><figref idref="f0029">Figure 37</figref> is a schematic representation of a state following a reset;</li><li><figref idref="f0030">Figure 38</figref> is a schematic representation of an example of a state after one packet is received for a destination;</li><li><figref idref="f0031">Figure 39</figref> is a schematic representation of an example of a state where packets for two lists are interspersed with free lists entries;</li><li><figref idref="f0032">Figure 40</figref> is a schematic representation of congestion at a single port;</li><li><figref idref="f0033">Figure 41</figref> is a schematic representation of the result of continued congestion;</li><li><figref idref="f0034">Figure 42</figref> is a schematic representation of where a maximal set of connections is reached;</li><li><figref idref="f0035">Figure 43</figref> is a schematic representation of an implementation of a TDM interface layer between local ports and a switch matrix unit;</li><li><figref idref="f0036">Figure 44</figref> is a schematic representation of a structure of a TDM storage primitive</li><li><figref idref="f0037">Figure 45</figref> is a schematic representation of the use of TDM primitives to allow sharing of logic between multiple independent pipelines; and</li><li><figref idref="f0037">Figure 46</figref> is a schematic representation of a 64 bit data path.</li></ul>

DETAILED DESCRIPTION

0012An example embodiment of a 3456-port InfiniBand 4x DDR switch in a custom rack chassis is described, with the switch architecture being based upon a 5-stage CLOS fabric. The rack chassis can form a switch enclosure.

0013The CLOS network, first described by Charles Clos in 1954, is a multi-stage fabric built from smaller individual switch elements that provides full-bisectional bandwidth for all end points, assuming effective dispersive routing.

0014Given that an external connection (copper or fiber) costs several times more per port than the silicon cost, the key to make large CLOS networks practical is to minimize the number of external cables required and to maximize the number of internal interconnections. This reduces the cost and increases the reliability. For example, a 5-stage fabric constructed with switching elements of size (n) ports supports (n*n/2*n/2) edge points, using (5*n/2*n/2) switch elements with a total of (3*n*n/2*n/2) connections. The ratio of total to external connections is 5:1, i.e. 80% of all connections can be kept internal. The switch elements (switch chips) in the described example can be implemented using a device with 24 4x DDR ports.

0015An example switch uses a connector that support 3 4x ports per connector, which can further to minimize a number of cables needed. This can provides a further 3:1 reduction in the number of cables. In a described example, only 1152 cables (1/3*n*n/2*n/2) are required.

0016In contrast if prior commercially available 288-port switches and 24-port switches were used to create a 3456-port fabric a total of 6912 cables (2*n*n/2*n/2) would be required.

0017The example switch can provide a single chassis that can implement a 5-stage CLOS fabric with 3456 4x DDR ports. High density external interfaces can be provided, including fiber, shielded copper, fiber and twisted pair copper. The amount of cabling can be reduced by 84.4% when compared to building a 3456-port fabric with commercially available 24-port and 288-port switches. In the example embodiment, an orthogonal midplane design can be provided that is capable of DDR data rates.

0018An example switch can address a full range of HPC cluster computing from a few hundred to many thousand of nodes with a reliable and cost-effective solution that uses fewer chassis and cables than prior solutions.

0019<figref idref="f0001">Figures 1</figref> and <figref idref="f0002">2</figref> are schematic diagrams of an example of a switch chassis as viewed from the rear (<figref idref="f0001">Figure 1</figref>) and front (<figref idref="f0002">Figure 2</figref>), respectively. This example comprises a custom rack chassis 10 that is 60" high, 47" wide, and 36" deep, not including a cable management system. The example embodiment provides a passive orthogonal midplane design (not shown in <figref idref="f0001">Figures 1</figref> and <figref idref="f0002">2</figref>) that provides a direct interface between Line Cards (LC) 12 and Fabric Cards (FC) 14. The line cards provide connections to external lines and the fabric card form switch fabric cards for providing switching functions.

0020In the present example, up to 18 fabric cards (FCO to FC17) 12, <figref idref="f0001">Figure 1</figref> are provided. Each fabric card 12 plugs vertically into the midplane from the rear.

0021In the present example, up to 24 line cards (LCO to LC23) 14, <figref idref="f0002">Figure 2</figref> can be provided. Each line card provides 144 4x ports (24 stacked 168-circuit cable connectors). Each line card plugs horizontally into the midplane from the front.

0022Up to 16 hot-pluggable power supply units (PSO - PS16) 16, <figref idref="f0001">Figure 1</figref> are each plugged into the chassis 10 from the rear. Each power supply unit 16 has an alternating current (AC) power supply inlet (not shown). The power supply units 16 plug into a power distribution board (PDB), which is not shown in <figref idref="f0001">Figures 1</figref> and <figref idref="f0002">2</figref>. Two busbars (not shown in <figref idref="f0001">Figures 1</figref> and <figref idref="f0002">2</figref>), one per group of 8 power supply units, distribute direct current (DC) supply to the line cards 12 and the fabric cards 14.

0023Two hot-pluggable Chassis Management Controllers (CMCs) 18, <figref idref="f0002">Figure 2</figref> plug into the power distribution board from the front. Each chassis management controller 18 comprises a mezzanine card.

0024The power distribution board is a passive power distribution board that supports up to 16 power supply units DC connectors and 2 chassis management controller slot connectors. The power distribution board connects to the midplane through ribbon cables that carry low-speed signals.

0025In the present example, up to 144 fan modules (Fan#0 - Fan#143) 20 are provided, with 8 fan modules per fabric card 12 in the present instance. Cooling airflow is controlled to be from the front to the rear, using redundant fans on the fabric cards to pull the air from the line cards 14 through openings (not shown in <figref idref="f0001">Figures 1</figref> and <figref idref="f0002">2</figref>), in the midplane. The power supply units 16 have their own fans for cooling with the air exiting through the rear of the chassis. The power supply units 18 are also used to cool the chassis management controllers 18.

0026<figref idref="f0003">Figure 3</figref> is a schematic representation of a printed circuit board 30, which is configured as a midplane 30 in the switch chassis 10. The midplane 30 is configured in an orthogonal manner such that each fabric card 12 can connect to each of the line cards 14 without requiring any signal traces on the midplane 30. The orthogonal midplane design can provide excellent signal integrity in excess of 10 Gbps per differential pair.

0027Also, although the fabric cards and the switch cards are described as being orthogonal to each other, they do not need to be exactly orthogonal to each other. Indeed, it should be appreciated that in an alternative embodiment they could be angled with respect to each other but need not be exactly orthogonal to each other.

0028The midplane 30 is represented schematically to show an array of midplane connector pairs 32 as black squares with ventilation openings shown as white rectangles. Each midplane connector pair 32 comprises a pair of connectors (to be explained in more detail later) with one connector on a first face of the midplane and a second connector on the other face of the midplane, the first and second connectors being electrically interconnected by way of pass-through vias (not shown in <figref idref="f0003">Figure 3</figref>) formed in the midplane 30. As will be explained later, the first and second connectors of a midplane connector pair 32 are each multipath connectors. They are arranged orthogonally with respect to one another such that a first midplane connector of a midplane connector pair 32 is connectable to a fabric card 12 on a first side of the plane 30 in a first orientation and a second midplane connector of the midplane connector pair 32 is connectable to a line card on a second side of the plane 30 in a second orientation substantially orthogonally to the first orientation.

0029In an example described herein, each of the first connectors of the respective midplane connector pairs 32 of a column 31 of midplane connector pairs 32 can be connected to one fabric card 12. This can be repeated column by column for successive fabric cards 12. In an example described herein, each of the second connectors of the respective midplane connector pairs 32 of a row 33 of midplane connector pairs 32 can be connected to one line card 14. This can be repeated row by row for successive line cards 14. As a result, the midplane can be populated by vertically oriented fabric cards 12 on the first side of the midplane and horizontally orientated line cards 12 on the second side of the midplane 30.

0030In the present example the midplane 30 provides orthogonal connectivity between fabric cards 12 and the line cards 14 using orthogonal connector pairs. Each orthogonal connector pair provides 64 differential signal pairs, which is sufficient to carry the high-speed signals needed as well as a number of low-speed signals. The orthogonal connector pairs are not shown in <figref idref="f0003">Figure 3</figref>, but are described later.

0031The midplane 30 is also configured to provide 3.3VDC standby power distribution to all cards and to provide I2C/System Management Bus connections for all fabric cards 12 and line cards 14.

0032Another function of the midplane 30 is to provide thermal openings for a front-to-rear airflow. The white holes in <figref idref="f0003">Figure 3</figref> (e.g., hole 34) form openings 34 in the midplane for airflow. In this example the midplane is approximately 50% open for airflow.

0033The fabric cards 12 each support 24 connectors and the line cards 14 each support 18 connectors.

0034<figref idref="f0003">Figure 3</figref> also illustrates an example of how the fabric cards 12, the midplane 20 and the line cards 14 interconnect. In this example there are 24 switch chips on a line card 14 and 8 chips on each of the 18 fabric cards 12.

0035As previously mentioned a 5-stage Clos fabric has a size n*n/2*n/2 in which n is the size of the switch element. The example switch element in <figref idref="f0003">Figure 3</figref> has n equal to 24 ports. Each line card 14 has 24 chips in 2 rows with 12 chips in each row. Each of 12 ports of each switch chip 35 in a first row 36 of the line card 14 is connected to 2 cable connectors 42, with 6 ports per cable connector. There are a total of 24 cable connectors per line card 14. Each cable connector can accommodate two physical independent cables that each carries 3 ports (links). Each cable connector 42 can accommodate 6 ports. The remaining 12 ports of each switch chip 35 in the first row 26 is connected to one chip 35 each in a second row 38 of chips 35.

0036There are 18 midplane connectors 32 per line card 14. Each midplane connector 32 provides one physical connection to one fabric card 14. Each midplane connector 32 can accommodate 8 4x links (there are 8 differential pairs per 4x link and a total of 64 differential pairs provided by the orthogonal connector)

003712 ports of each of the switch chips 35 in the second row 38 of the line card 14 are connected to 2 line card connectors 40 that are used to connect the line card 14 to the midplane connectors 32 and thereby with the fabric cards 12 through the orthogonally oriented midplane connector pair. Of the 12 ports per switch chip 35, eight ports are connected to one line card connector 40, and the remaining four ports are connected to another line card connector 40 as represented by the numbers 8 and 4 adjacent the two left hand switch chips 35 in the second row 38. 2 switch chips are thereby connected to a group of 3 line card connectors 40 and hence to a group of three midplane connectors pairs 32.

0038The remaining 12 ports of each switch chip 35 in the second row 38 of the line card 14 are connected to each of the 12 switch chips 35 in the first row 36 of the line card 14.

0039At the fabric card 12 all links through an orthogonally oriented midplane connector pair 32 are connected to one line card 14. A single orthogonal connector 46 carries 8 links. These links are connected to one switch element 44 each at the fabric card 12.

0040Also shown in <figref idref="f0003">Figure 3</figref> are power connectors 37 on the midplane and power connectors 39 on the fabric cards 12.

0041There has been described a system with 24 line cards with 144 ports each, realized through 48 physical cable connectors that each carry 3 links. The switch fabric structure of each line card 14 is fully connected, so the line card 14 itself can be viewed upon as a fully non-blocking 144 port switch. In addition each line card 14 has 144 links that are connected to 18 fabric cards. The 18 fabric cards then connect all the line cards 14 together in a 5-stage non-blocking Clos topology.

0042<figref idref="f0004 f0005">Figure 4</figref>, formed of <figref idref="f0004">Figures 4A</figref> and <figref idref="f0005">4B</figref>, is a schematic diagram of an example management infrastructure. This example provides redundant chassis management controllers 18. In addition each fabric card 12 and line card 14 supports an management controller. There are redundant management connections from each chassis management controller 18 to each of the fabric card and line card management controllers. In addition there are I2C connections to each of the power supply units 16. The management connections pass between the fabric cards 12, the line cards 14, the power supply units 16 and the chassis management cards 18 via the midplane and the power distribution board 22 in the present example.

0043<figref idref="f0006 f0007 f0008 f0009 f0010 f0011 f0012">Figures 5 to 11</figref> provide various schematic views of an example of a switch chassis.

0044<figref idref="f0006">Figure 5</figref> is a front view of the switch chassis 10 showing cable management structures 50. <figref idref="f0007">Figure 6</figref> is a rear view of the switch chassis 10 showing the fabric cards 12, the power supply units 16 and cable management structures 50. <figref idref="f0008">Figure 7</figref> is a side view of the switch chassis 10 further showing the cable management structures 50. <figref idref="f0009">Figure 8</figref> is a side view of the switch chassis 10 further showing the cable management structures 50. <figref idref="f0010">Figure 9</figref> is an isometric view of the switch chassis 10 from the line card 14 (front) side further showing the cable management structures 50. <figref idref="f0011">Figure 10</figref> is an isometric view of the switch chassis 10 from the line card 14 (front) side showing four line cards 12 installed horizontally in the chassis 10 and part of the cable management structures 50. <figref idref="f0012">Figure 11</figref> is an isometric view of the switch chassis 10 from the fabric card 12 (rear) side showing four fabric cards 12 installed vertically in the chassis 10 and part of the cable management structures 50.

0045<figref idref="f0013">Figures 12</figref> and <figref idref="f0014">13</figref> provide various schematic views of an example of a midplane 30. <figref idref="f0013">Figure 12</figref> is an isometric view of the midplane 30 from the line card 14 (front) side and <figref idref="f0014">Figure 13</figref> is an isometric view of the midplane 30 from the fabric card 12 (rear) side. <figref idref="f0013">Figure 12</figref> shows the array formed from rows and columns of the second connectors 64 of the midplane connectors pairs 32 described with reference to <figref idref="f0003">Figure 3</figref>. <figref idref="f0014">Figure 13</figref> shows the array formed from rows and columns of the first connectors 62 of the midplane connectors pairs 32 described with reference to <figref idref="f0003">Figure 3</figref>.

0046<figref idref="f0015">Figure 14</figref> is an isometric view of an example of a line card 14. This shows the first and second rows 36 and 38 of switch chips 35, the line board connectors 40 and the cable connectors 42. As can be seen in <figref idref="f0015">Figure 14</figref>, the cable connectors 42 are stacked double connectors such each cable connector can connect to two cables 52 and 54.

0047<figref idref="f0015">Figure 15</figref> is an isometric view of an example of a fabric card 12. This shows the fabric card connectors 46 and the switch elements 44.

0048<figref idref="f0016">Figure 16</figref> is a schematic representation of an example of two chassis management controllers 18 plugged into one side of a power distribution board 22 and 16 power supply units 16 plugged into the other side of the power distribution board 22. In the present example, the chassis management controllers 18 are plugged into the front side of the power distribution board 22 and the power supply units 16 are plugged into the rear side of the power distribution board 22 as mounted in the switch chassis. <figref idref="f0016">Figure 17</figref> illustrates bus bars 24 for a 3.3V standby supply.

0049In the present example the midplane 30 is a passive printed circuit board that has dimensions of 1066.8mm (42") x 908.05mm (35.75") x 7.1mm (.280"). The active area is 40" X 34". 864 8x8 midplane connectors (432 midplane connectors per side) are provided. There is a ribbon cable connection the power distribution board 22 and a 3.3V standby copper bar to the power distribution board 22.

0050In the present example a fabric card 12 comprises a printed circuit board with dimensions of 254mm (10") X 1016mm (40") X 4.5mm (177"). It comprises 24 8x8 fabric card connectors 46, one power connector 39, 8 fan module connectors and 8 switch chips 44.

0051In the present example a line card 14 comprises a printed circuit board with dimensions of 317.5mm (12.5") X 965.2mm (38") X 4.5mm (177"). It comprises 24 stacked cable 168-circuit connectors 42, 18 8x8 card connectors 40, 1 busbar connector and 24 switch chips 35.

0052In the present example a power distribution board 22 comprises a printed circuit board, 16 power supply DC connectors, 14 6x6 card connectors (7 connectors per chassis management card 18, ribbon cable connectors for low-speed connectivity to the midplane 30, and a 3.3V standby copper bar to the midplane 30.

0053In the present example a chassis management card 18 comprises 14 6x6 card connectors (7 connectors per chassis management card, two Rj45 connectors with magnetics for Ethernet available on a chassis management card panel, two Rj45 connectors for serial available at the chassis management card panel, three Rj45 for line card/fabric card debug console access at the chassis management card panel, three HEX rotary switches used to select between which line card/fabric card debug console is connected to the three Rj45s above, and a 220-pin connector for the mezzanine.

0054In the present example a mezzanine has dimensions: 92.0 mm x 50.8 mm and comprises 4 mounting holes for screws with either 5mm or 8mm standoff from the chassis management card board, a 220-pin connector for connectivity to chassis management board.

0055<figref idref="f0017">Figure 18</figref> is a schematic isometric view of an example of a midplane connector pair 32. As can be seen in <figref idref="f0017">Figure 18</figref>, the connector comprises a first, fabric side, connector 62 and a second, line card side, connector 64. In this example, each of the connector 62 and 64 is substantially U-shaped and comprises an 8x8 array of contact pins.

0056It will be noted that the second connector 64 of the midplane connector pair 32 is rotated through substantially 90 degrees with respect to the first connector 62. The first connector 62 is configured to connect to a corresponding fabric card connector 46 of a fabric card 12. The second connector 62 is configured to connect to a corresponding fabric card connector 46 of a line card 14. Through the orientation of the second connector 64 of the midplane connector pair 32 substantially orthogonally to the orientation of the first connector 62, it can be seen that the line card 14 is mounted substantially orthogonally to the fabric card 12. In the present example the line card 14 is mounted substantially horizontally and the fabric card is mounted substantially vertically 12.

0057Each of the contact pins on the connector 62 is electrically connectable to a corresponding contact of the fabric card connector 46. Each of the contact pins on the connector 64 is electrically connectable to a corresponding contact of the line card connector 40. The connector pins of the respective connectors 62 and 64 are connected by means of pass-through vias in the midplane 30 as will now be described in more detail.

0058<figref idref="f0017">Figure 19</figref> illustrates an example of the configuration of a first midplane connector 62 and a second midplane connector 64 of a midplane connector pair 32 in more detail. In the example shown in <figref idref="f0017">Figure 19</figref> that second connector 64 (the line card side connector) comprises a substantially U-shaped frame 70 including a substantially planar base 71 and first and second substantially planar walls 72 and 74 that extend at substantially at 90 degrees from the base 71. The inside edges of the first and second substantially planar sides 72 and 74 are provided with ridges 76 and grooves 78 that provide guides for the line card connector 40.

0059As can be seen in <figref idref="f0017">Figure 18</figref>, the line card connector 40 has a structure that comprises a plurality of contact planes 63 that are aligned side by side, such that it has a generally planar construction that extends up from the line card 14. Line card connector planes comprise printed circuit boards carrying traces leading to contacts. The traces and contacts can be provided on both sides of the printed circuit boards of the line card connector planes.

0060By comparing <figref idref="f0017">Figures 18 and 19</figref>, it can be seen that each contact plane 63 of the line card connector 40 can be entered into a respective one of the grooves 78 so that connectors of the line card connector 40 can then engage with contact pins 80 of the second connector 64. In the case of the line card side connector portion 64, the orientation of second connector 64 and the grooves 78 therein means that the line card 12 is supported in a substantially horizontal orientation. In the example shown in <figref idref="f0017">Figure 19</figref>, an 8x8 array of connector pins 80 is provided.

0061The first midplane connector 62 (fabric card side connector) of the midplane connector pair 32 has substantially the same form as the second midplane connector 62 of the midplane connector pair 32, except that it is oriented at substantially 90 degrees to the second midplane connector 64. In this example the second midplane connector 62 comprises a substantially U-shaped support frame 75 including a substantially planar base and first and second substantially walls and that extend at substantially at 90 degrees from the base. The inside edges of the first and second substantially planar sides are provided with ridges and grooves that provide guides for the fabric card connector 46. The fabric card connector 46 has the same basic structure as that of the line card connector 40 in the present instance. Thus, in the same way as for the line card connector, each of a plurality of contact planes of the fabric card connector 46 can be entered into a respective one of the grooves so that connectors of the fabric card connector 46 can then engage with contact pins of the first connector 62. The orientation of the first connector 62 and the grooves therein means that the fabric card 12 is supported in a substantially vertical orientation.

0062In the example illustrated in <figref idref="f0017">Figure 19</figref>, the orthogonal connector 60 provides an 8x8 array of connector pins 80 is provided that can support supports 64 differential pairs or 32 bi-directional serial channels (two wires per direction) in a footprint of 32.2x32.2 mm.

0063As mentioned above, the contact pins of the first and second midplane connectors 62 and 64 of a midplane connector pair 32 are connected by means of pass through vias in the midplane.

0064<figref idref="f0018">Figure 20</figref> illustrates a side view of an example of a midplane connector, for example the midplane connector 62 mounted on the midplane. In the example shown in <figref idref="f0018">Figure 20</figref> the midplane connector 64 comprises a substantially U-shaped frame 70 including a substantially planar base 71 and first and second substantially planar walls 72 and 74 that extend at substantially at 90 degrees from the base 71. The contact pins 80 are each connected to pairs of contact tails 81 that are arranged in sprung pairs that are arranged to be push fitted into pass through vias 83 in the midplane 30.

0065In use, the other midplane connector (e.g., the first midplane 62) of the midplane connector pair would be inserted into the pass through vias in the other side of the midplane 30 in the orthogonal orientation as discussed previously.

0066<figref idref="f0018">Figure 21</figref> is a schematic representation of an area of the midplane for receiving the midplane connectors 62 and 64 of the midplane connector pair 32. This shows the array of vias 83. <figref idref="f0019">Figure 22</figref> is a schematic cross-section though such a via 83 in the showing the conductive wall 85 of the via 83. The conductive wall 85 can be formed by metal plating the wall of the via, for example.

0067The examples of the midplane connectors described with reference to <figref idref="f0017">Figures 18</figref> and <figref idref="f0018">20</figref> had a generally U-shape. However, other configurations for the midplane connectors are possible. For example <figref idref="f0019">Figure 23</figref> illustrates another example of a midplane connector pair 32', where the first and second midplane connectors 62' and 64'are generally the same as the first and second midplane connectors 62 and 64 described with reference to <figref idref="f0017">Figure 19</figref> except that, in addition to the first and second walls 72 and 74, third and fourth walls 73 and 75 are provided. The additional walls provide a generally box-shaped configuration that can facilitate the insertion and support for the cards to be connected thereto.

0068It will be appreciated that in other embodiments the first and second midplane connectors could have different shapes and/or configurations appropriate for the connections for the cards to be connected thereto.

0069Also, although in the above described embodiment the midplane connector pairs 32 are configured as first and second connectors 62 and 64, in another embodiment they could be configured as a single connector that is assembled in the midplane. For example, through connectors could be provided that extend through the midplane vias. The through connectors could be manufactured to be integral with a first connector frame (e.g., a U-shaped frame or a box-shaped frame as in <figref idref="f0017">Figures 19</figref> and <figref idref="f0019">23</figref>, respectively) and the contacts inserted through the vias from a first side o the midplane 30. Then a second connector frame could be inserted over the connectors on the second side of the midplane 30 in a mutually orthogonal orientation to the first connector frame.

0070The array of midplane connector pairs 32 as described above provides outstanding performance in excess of 10 Gbps over a conventional FR4 midplane because the orthogonal connector arrangements allow signals to pass directly from the line card to the fabric card without requiring any signal traces on the midplane itself. The orthogonal arrangements of the cards that can result from the use of the array of orthogonally arranged connector pairs also avoids the problem of needing to route a large number of signals on the midplane to interconnect line and fabric cards, minimizing the number of layers required. This provides a major simplification compared to existing fabric switches. Thus, by providing an array of such orthogonal connectors, each of a set of horizontally arranged line cards 14 can be connected to each of a set of vertically aligned fabric cards 12 without needing intermediate wiring.

0071<figref idref="f0020">Figures 24 and 25</figref> provide an end view and a side view, respectively, of an example of a cable connector 42 as mentioned with reference to <figref idref="f0003">Figures 3</figref> and <figref idref="f0015">14</figref>. As shown in <figref idref="f0020">Figure 24 and 25</figref>, the cable connectors 24 and 25 include first and second cable connections 92 and 94 stacked within a single housing 90. This provides for a very compact design. Board contacts 96 are provided for connecting the connector to a line card 14. <figref idref="f0020">Figure 26</figref> is a plan view of the connector footprint for the board contacts 96 of the cable connector 42. The stacked arrangement facilitates the providing of line cards that are high density line cards supporting a 12X cable providing 24 line pairs with 3 4X links aggregated into a single cable. The cable connectors provide 12X cable connectors that are smaller than a conventional 4X connector, 3X denser than a standard InfiniBand 4X connector and electrically and mechanically superior. Using 12X cable (24 pairs) can be almost 50% more area efficient than three 4X cables and requires three times fewer cables to install and manage.

0072<figref idref="f0021">Figures 27 and 28</figref> illustrate an example of the routing of signals from each of two 12x port sections 92 and 94 of a cable connector 42 to the equalizers and to a switch chip on a line card 14. <figref idref="f0021">Figure 27</figref> shown an example of routing from a first 12x port section. <figref idref="f0021">Figure 28</figref> shows an example of the routing from a second 12x port section. The transmit (Tx) lines are equalized, and can be connected directly from the switch chip to the cable connector. The lines can be routed on lower layers in order to minimize via stub effects.

0073<figref idref="f0022">Figure 29</figref> illustrates an example of a power supply for the cable connector and <figref idref="f0022">Figure 30</figref> illustrates an example of a cable status sense detection circuitry. The cable sense detection circuitry is operable to test from each end whether the other end is plugged or not, and, if plugged, to see if power from the power supply is on. Provisions are made such that "leaking" power from a powered to un-powered end is avoided. A valid status assumes that an active end is plugged. <figref idref="f0023">Figure 31</figref> is a schematic diagram of an example of a hot plug control circuit that enables hot plugging of cables. The switch chassis can thereby provide active cable support for providing active signal restoration at a cable connector. Active cable support can provides benefits of increased distances for copper cables as a result of active signal restoration at the connector, increased maximum cable distance by over 50%, using thinner and more flexible cables (e.g., reducing a cable diameter by up to 30%, which facilitates good cable management. A cable to connector interface can provide one, more or all of local and remote cable insertion detection, cable length indication, remote node power-on detection, remote power, a serial number and a management interface.

0074<figref idref="f0024">Figure 32</figref> is a schematic representation of the airflow through an example switch chassis. As illustrated by the arrows, the airflow is from the front to the rear, being drawn through by fans 20 in the fabric cards 12 and the power supplies 18.

0075The air inlet is via perforations at the line card 14 front panel. Fans 20 at the fabric cards 12 pull air across the line cards, though the openings 34 in the vertical midplane 30 and across the fabric cards 12.

0076Line card cooling is naturally redundant since the fabric cards are orientate orthogonally to the line cards. In other words, cooling air over each line card is as a result of the contribution of the effect of the fans of the fabric cards along the line card due to the respective orthogonal alignment. In the case that a fabric card fails or is removed, a portion of the cooling capacity is lost. However, as the cooling is naturally redundant the line cards will continue to operated and be cooled by the remaining fabric cards. Each fan is internally redundant and the fans on the fabric cards 12 can be individually hot swappable without removing the fabric card 12 itself. The fabric card 12 and line card 14 slots can be provided with blockers to inhibit reverse airflow when a card is removed. Empty line card 14 and fabric card 12 slots can be loaded with filler panels that prevent air bypass.

0077Each power supply has an internal fan that provides cooling for each power supply. Fans at the power supplies pull air through chassis perforations at the rear, across the chassis management cards 18, and through the power supply units 16. Chassis management card cooling is naturally redundant as multiple power supply units cool a single the chassis management card.

0078Although in the present example cooling is provided by drawing air from the front to the rear, in another example embodiment cooling could be from the rear to the front.

0079A switch system as described above can provide a switch matrix that includes a configurable, protocol agnostic, packet switching core. It can provide a non-blocking switch connecting a configurable number of ports together. This will be described in more detail in the following.

0080An example switch matrix can receive as its input data packets of any variety as a raw stream of data along with sideband control signals indicating start/stop, size and routing instructions for the packet. It can buffers these packets as needed in virtual output queues, schedules them for transmission using a distributed, iterative routing mechanism and, once the scheduling is complete, transmits the packets to their destination port. The data path width can be a configurable parameter and, in the present example can be any multiple of 8 (currently supported data widths are 64 and 128 bits).

0081The basic building block in the example switch matrix is referred to herein as Switch Matrix Unit (XMU). Each XMU implements one port connection. Optionally, an XMU can be configured in a Time Division Multiplex (TDM) mode, wherein it connects to a TDM port and can be configured to function as multiple ports simultaneously. Valid TDM port configurations are dependent on the data width being used. If it is intended to make use of the TDM concept then the data path width can be further constrained to a number of bytes that is a divisible by the number by the number of ports supported.

0082<figref idref="f0025">Figure 33</figref> is a schematic block diagram representing a 6 port Switch Matrix 100 as an illustrative example only. In this example, six XMUs 101-106 are instantiated and connected together by means of a switch interconnect 110. The switch interconnect could be implemented in the form of a fully connected point to point multiplexing structure. In this example, the switch interconnect 110 is operable to connect to treat XMUs 101, etc. as being synchronous and to meet timing (1 cycle) on all connections between XMUs.

0083Theoretically, the number of XMUs 101, etc. instantiated in one switch matrix is unlimited. However, practical implementations can determine how many XMUs can are in any given design. It will be appreciated that is a practical implementation, many more than six XMUs will typically be implemented, the choice of six XMUs in <figref idref="f0025">Figure 33</figref> merely being for illustrative convenience.

0084<figref idref="f0026">Figure 34</figref> is a schematic representation of an example configuration of an XMU 101. The other XMUs 102-106 can be similarly configured. The XMU 101 is connected to receive packets from a port 111. The packets are allocated by input logic 113 to a Virtual Output Queue (VOQ) mechanism 112 that is used to buffer the packets received at the XMU 101.

0085The concept behind a VOQ mechanism is that packets are sorted into separate virtual queues based on their destination, as will be described in more detail hereinafter. In this way, packets bound to different destinations do not block each other. The input data packets are received as a raw stream of data with sideband control signals indicating start/stop, size and routing instructions for the packets. The input logic 113 of the VOQ mechanism uses the packet contents and the sideband control signals to sort the packets into the separate virtual queues.

0086Arbiter logic is operable to communicate via the switch interconnect 110 with corresponding arbiter logic of the other XMUs 102 - 106 to schedule the sending of the packets buffered in the VOQ mechanism 112 to appropriate output ports. The arbiter logic is coupled to the VOQ mechanism 112 and also to an output multiplexer 114, which receives packets from the switch interconnect 110 to be passed to the port 111, to schedule and control the operation of the VOQ mechanism 112 and the multiplexer 114.

0087<figref idref="f0027">Figure 35</figref> is a schematic representation of an example of a VOQ mechanism 112 for use in an XMU (say XMU 101). In the example shown in <figref idref="f0027">Figure 35</figref>, all packets received, for example, from an input port 111 are placed in the virtual queue 120, 121, 122, 123 associated with their destination port 130, 131, 132, 133 (Packets destined for Port 0 (130) are placed in Virtual Queue 0 (120), etc). In this way, a stall on Port 0 (130) will not prevent packets in Virtual Queue 1 (121) from making progress to Port 1 (131) via the switch interconnect 110.

0088These Virtual Output Queues 120, 121, 122, 123 are not subject to independent flow control on the input to the XMU 101, etc. The queues exist as a single physical entity in the XMU 101, etc., and the separation into the separate queues is only a virtual separation, as represented by the gate 125.

0089The Virtual Output Queues form a first level of categorization for incoming packets. A further, second, level of categorization is in the form of packet categories. This second layer of categorization sits on top of the Virtual Output Queue mechanism. Packet Categories can be separately flow controlled on the input to the XMU 101, etc., using the sideband control signals. The number of Packet Categories can be a user defined value. It should be noted, that the XMU can also support a vanilla switch that contains no packet categories.

0090<figref idref="f0028">Figure 36</figref> is a schematic representation of an example organization for Packet Category storage. As shown in <figref idref="f0028">Figure 36</figref>, each Packet Category (PCO, PC1 ... PCn) has its own independent Virtual Output Queue structure. In other words, each Packet Category has a number of independent virtual output queues equal to the number of destination ports in the Switch Matrix network. In the example shown in <figref idref="f0028">Figure 36</figref>, there are four output ports 130, 131, 132, 133, and for each of the n Packet Categories, there are four virtual output queues 120, 121, 122, 123.

0091Accordingly, the total number (#) of queues in an example of an XMU is given by the formula below: <maths id="math0001"><math display="block"><mi>#Queues</mi><mo>=</mo><mfenced><mi># of Destinations</mi></mfenced><mo>*</mo><mfenced><mi># of</mi><mspace width="1em" /><mi>Packet Categories</mi></mfenced></math><img file="EP2018071B1_D0001.tif" /></maths>

0092In an example embodiment, the amount of physical storage assigned to each Packet Category can be run-time configurable, for example using control and status registers (CSRs).

0093In order to support the Virtual Output Queuing mechanism described above, a buffer management scheme can be implemented. The example scheme used in the present example is a full scatter/gather algorithm which operates on a memory size of 128 bits. On ingress, individual flits (parts) of a packet will be "scattered" (physically) throughout the RAM. On egress, the read state machine will "gather" the flits back together to reconstruct the packet.

0094In the present example, the scatter/gather algorithm is implemented by maintaining several linked lists per Packet Category: <ul id="ul0002" list-style="dash"><li>One Free List</li><li>One Packet List per Destination</li></ul>

0095The Free List contains the addresses of all of the unused buffer space and the Packet Lists contain the addresses of all of the packets currently in the buffer. Note that an address can only exist in one of the linked lists at a time.

0096<figref idref="f0029">Figure 37</figref> is a schematic representation of the state following a reset. In this state the buffers will all be empty, so the entire contents of the RAM will be contained in the free list and the packet lists will be empty. The free list 140 is ordered from "F0" to "F15", with a free head pointer 142 pointing to the head F0 of the free list 140 and the free tail pointer 144 pointing to the tail F15 of the free list 140.

0097When a packet arrives for the packet category in question, the linked lists will begin to change. The first flit of the packet will be written into the first entry contained in the free list 140 and so on. The free list 140 and the appropriate packet lists will then be updated to reflect the new state of the queue.

0098<figref idref="f0030">Figure 38</figref> is a schematic representation of an example of a state after one packet is received for Destination A. In this state, the free list 140 now only comprises entries F0-F9, with the free head pointer 142 pointing to the head F0 of the free list 140 and the free tail pointer 144 pointing to the tail F9 of the free list 140. The flits of the packet for destination A is then held in entries Pkt1.0 to Pkt1.5 with a List A head pointer 152 pointing to the head Pkt1.0 of the List A 150 packet and the List A tail pointer 154 pointing to the tail Pkt1.5 of the List A 150 packet.

0099After some time the lists can start to appear fragmented. Entries containing packets and free space will be interspersed. The lists might look something like the representation of <figref idref="f0031">Figure 39</figref> where packets for lists A 150 and B 160 are interspersed with free lists 140 entries. As represented in <figref idref="f0031">Figure 39</figref>, list A starts at the entry pointed to by List A head pointer 152 and then follows links within the entries to entry also pointed to by List A tail pointer 154. Also, list B starts at the entry pointed to by List B head pointer 162 and then follows links within the entries to entry also pointed to by List B tail pointer 164. Similarly the free list starts at the entry pointed to by the free head pointer 142 and then follows links within the entries to the entry also pointed to by the free tail pointer 144. As a result, individual packet flits are scattered throughout the RAM and the linked lists maintain the context.

0100Control of the linked lists is accomplished through two algorithms, one for packet reads (transmit), one for packet writes (receive). When a packet is received into the queue it is written into the unused portion of the buffer. On each cycle, the incoming flit is always written into the buffer entry indicated by the first entry in the free list. The free list and the appropriate packet list are then updated according to this algorithm: <ol id="ol0001" compact="compact"><li>1. If the packet list is empty then the head of the free list becomes the new head of the packet list.</li><li>2. If the packet list is not empty, then the head of the free list is added to the end of the packet list.</li><li>3. The next entry in the free list becomes the new head of the free list.</li></ol>

0101As a packet is transmitted, the buffer is cleared of that packet and the resulting free space is added to the free list. On each cycle, the outgoing flit is read from the buffer entry indicated by the first entry in the appropriate packet list. The free list and the appropriate packet list are then updated according to this algorithm: <ol id="ol0002" compact="compact"><li>1. If the free list is empty then the head of the packet list becomes the new head of the free list.</li><li>2. If the free list is not empty then the head of the packet list is added to the end of the free list.</li><li>3. The next entry in the packet list becomes the new head of the packet list.</li></ol>

0102In the present example, all packets transfers are aligned to 32 bit boundaries. If the data path width is 64 or 128 then this implies that packets can start and end in the same cycle. An alignment buffer can be used when writing the packet into the data buffer managed by the scatter/gather algorithm. With an alignment buffer is used, the first flit of packet can be written into the alignment buffer instead of to the regular buffer. The existence of this alignment buffer is hidden from the local port and is not included in the flow control reporting. In this way, the holes in the memory introduced by misaligned packets can be hidden.

0103In the interest of keeping latency as low as possible, the packet buffers can be designed to support a "cut through" mode (software configurable). In this mode, a packet can be eligible to be transmitted to the destination port before the entire contents of the packet is written into the input buffer.

0104To determine when a packet is eligible for transmission, the control logic can take into account the relative speed of the source and destination ports. Transmission of a packet does not start until a sufficient portion of the packet is present in the buffer to guarantee that the entire packet can be sent without stalling to wait for the packet to arrive from the input port.

0105If the input port is faster than the output port then the packet is eligible for transmission as soon as the first flit is received.

0106If the output port is fast than the input port, then an example formula for determining eligibility can be used as follows: <ul id="ul0003" list-style="none" compact="compact"><li>flits_in = pkt_length - pkt_length * (recv_spd / xmit_spd).</li><li>"flits_in" is the total number of flits that must be present in the buffer before the packet is eligible for transmission.</li><li>"pkt_length" is the length of the packet in flits,</li><li>"recv_spd" is the speed of the receive port in number of flits per cycle.</li><li>"xmit_spd" is the speed of the transmit port in number of flits per cycle.</li></ul>

0107Optionally, a 'minimum arbitration latency' could also be factored into this formula.

0108It is to be noted that packets that are eventually determined to be bad might already be in flight to the destination device before the error is detected.

0109When "cut through" mode is turned off, a store and forward scheme can be implemented. With a store and forward scheme, a packet would not be considered eligible for arbitration and transmission until the entire contents of the packet have been placed into the input buffer and the source port has confirmed that the packet is valid.

0110The virtual output queue structure can help to prevent Head of Line (HoL) blocking between packet categories. Since all packet categories have separate flow control and buffering resources, no amount of congestion occurring on one packet category will affect any other packet category.

0111The virtual output queue structure can also help to alleviate HoL within a packet category inside an XMU. <figref idref="f0032">Figure 40</figref> shows that congestion at a single port (port 1) will not prevent the queues for other destinations from making progress.

0112If the congestion continues, however, the VOQ will not be able to prevent HoL problems completely within a packet category. Eventually, continued congestion will result in the entire queue for that Packet Category being filled with packets that are bound for the congested port. This problem is illustrated in <figref idref="f0033">Figure 41</figref>. The buffer space allocated to this Packet Category is consumed entirely by packets bound for output port 1, which is congested. This congestion then propagates back to the external port 170. The external port 170 is now unable to send packets to destinations 0, 2 and 3, even though they are not congested.

0113Since it is not possible to entirely prevent this kind of head of line blocking in the XMU 121, etc., these conditions should be detected and removed to prevent them from affecting the network any more than necessary. To support this goal, the XMU 121, etc. has a number of built trigger conditions that can be used in HoL detection. These can be used to detect when forward progress has stopped within a Packet Category. The actual trigger conditions included can be configured by the user at compile time.

0114Once a HoL blocking condition is detected, an external entity 170 is responsible for controlling any recovery needed. To support this, the XMU 101, etc. includes hardware support 180 for flushing the contents of its queues. It can be programmed to flush a specified number of packets from any specified queue, or it can be programmed to flush an entire queue.

0115The XMCJ 121, etc. implements a distributed arbitration scheme. There is no central arbiter involved, instead each individual XMU 121, etc. implements input and output arbiters 182 and 184 (see <figref idref="f0034">Figure 42</figref>). These arbiters collaborate to arrive at scheduling decisions.

0116The distributed scheme is implemented as a three phase process: <ol id="ol0003" compact="compact"><li>1. Advertise - input arbiters 182 advertise available packets to output arbiters 184.</li><li>2. Request - Output arbiters 184 request a packet from an input arbiter 182 and provide an indication of how busy they currently are.</li><li>3. Acknowledge - Input arbiters 182 acknowledge one packet request and begin transmitting the selected packet.</li></ol>

0117This arbitration scheme is repeated every cycle in a pipelined fashion. In this way, the algorithm iterates over the scheduling problem. If a particular input arbiter 182 or output arbiter 184 failed to make a match during the previous iteration then it will continue to arbitrate on the following iterations. Decisions made on previous cycles are fed back into the arbitration algorithm and new connection matches are sought after. Eventually, a maximal set of connections will be reached. This is not an absolute maximal set, but rather a maximal set given all previous decisions (IOW, no backtracking is done). <figref idref="f0034">Figure 42</figref> illustrates this process visually.

0118In the advertise phase, each input arbiter (iarb) will notify each output arbiter (oarb) of any packets that it has received which are routed to that output arbiter. The iarb will send the following information to each oarb: <ul id="ul0004" list-style="dash" compact="compact"><li>Packet length - The length of the packet in bytes.</li><li>Local destination port - If the XMU services more than one destination port then it is necessary to specify for which port the packet is intended.</li><li>Packet category - The packet category with which the packet is associated.</li><li>Weight - An adjustable value used by the iarb to indicate to the oarb what priority the packet in question should be considered.</li></ul>

0119The weight mechanism is the primary means of arbitration in this phase. Each iarb is assigned a pool of 'weight credit'. The iarb then assigns this weight credit to the packets in his queue. Assigning more weight to a particular packet increases the likelihood that the packet will be requested by the output arbiter. Factors that are taken into account in assigning weight credit to a packet include: <ul id="ul0005" list-style="dash" compact="compact"><li>Time in queue - How long has the packet been waiting?</li><li>Number of packets in queue - The total weight will be shared amongst multiple packets.</li><li>Congestion at the destination port - The oarb will be providing congestion information back to the iarb, if an iarb notices a congested port, then it can choose to lower the priority on packets to that port and focus on servicing other ports.</li></ul>

0120The iarb is designed such that the exact algorithm used in this phase is programmed by the user at compile time. The user can write his own verilog module to define this arbitration algorithm, or he can instantiate one of the predefined arbitration algorithms.

0121The maximum amount of weight credit an iarb can have can be configured at compile time. The actual amount of weight credit an iarb has can be configured at run-time through state registers (CSRs).

0122During this phase, the oarb will examine all of the packets, as advertised by all of the iarbs in phase 1, that are available to transmit on his local ports. The oarb will first check if the local port has the necessary credit for transmitting the packets. Any packets for which there is not necessary credit will be eliminated from arbitration. The oarb will then use the advertised weights of the remaining packets to determine which packet it would like to have transmitted. A single packet (per local destination port) is chosen, and a packet request is sent to the corresponding iarb.

0123Like the iarb, the oarb is also designed such that the exact algorithm used in this phase is programmed by the user at compile time. The user can write his own verilog module to define this arbitration algorithm, or he can instantiate one of the predefined arbitration algorithms.

0124It is during this final phase of arbitration that connections are made and packets start being transmitted. The iarb gathers all of the requests for transmit from all of the oarbs in the matrix. The iarb then must choose which request to acknowledge. The factors used to make this selection are: <ul id="ul0006" list-style="dash" compact="compact"><li>Time in queue.</li><li>Congestion at destination port.</li></ul>

0125This arbitration algorithm can also defined by the user at compile time. The user can write his own verilog module to define this arbitration algorithm, or can instantiate one of a number of predefined arbitration algorithms.

0126Time division multiplexing (TDM) is a mechanism introduced to allow a pipeline to be run in various configurations without unnecessary duplication of logic. It is based around the idea that a single pipeline providing a set bandwidth, N, can be split into separate, independent, pipelines whose total bandwidth is still equal to N and all data processing logic can be shared by the individual pipelines. The pipeline can then be configured to work in the desired configuration within the current system.

0127TDM is implemented through the use of "TDM storage primitives" and the use of a system wide "phase" signal which will be described in subsequent sections.

0128TDM is an optional feature in the XMU that can be configured by the user at compile time. This allows a single XMU to service more than one local port in a TDM manner. A TDM interface layer is implemented between local ports 192 and 194 and an XMU 121 as shown in <figref idref="f0035">Figure 43</figref>.

0129If TDM is enabled, then the number of local ports supported can also be configured. The configurations depend on the data path width being used. If the data path is configured to be 64 bits then the configurations are: <ul id="ul0007" list-style="dash" compact="compact"><li>One 64 bit device.</li><li>Two 32 bit devices.</li></ul>

0130If the data path is configured to be 128 bits then the configurations are: <ul id="ul0008" list-style="dash" compact="compact"><li>One 128 bit device.</li><li>Two 64 bit device.</li><li>Four 32 bit device.</li><li>One 64 bit device + two 32 bit device.</li></ul>

0131Furthermore, each of these devices can actually be run at slower speeds through the use of the phase enable signal (see below).

0132A fundamental building block of an example TDM embodiment is the TDM storage primitive. The storage primitive consists of a duplicate register bit for each supported pipeline in the design. In one example, 4 register bits are provided per bit of storage needed. These four registers are then controlled by a phase signal. <figref idref="f0036">Figure 44</figref> shows an example structure of a TDM storage primitive 200. Illustrated in <figref idref="f0036">Figure 44</figref> are the four registers 212, which receive data in, and are selectively enables 206 in response to the phase signal 202. The data is selectively output 210 by a multiplexer 208, which also operates in response to the phase signal 202.

0133Thus, control of the TDM storage primitive 200 is implemented with the 'phase' signal 202. The phase signal 202 is a globally synchronous signal that indicates which core is active for the current clock cycle. The meaning of the phases is dependant on the configuration of the pipeline. Table 1 summarizes and example indication of which core is active based on the phase signal value and the core configuration. <tables id="tabl0001" num="0001"><table frame="all"><title>TABLE 1 - ACTIVE CORE</title><tgroup cols="5"><colspec colnum="1" colname="col1" colwidth="32mm" /><colspec colnum="2" colname="col2" colwidth="14mm" /><colspec colnum="3" colname="col3" colwidth="14mm" /><colspec colnum="4" colname="col4" colwidth="14mm" /><colspec colnum="5" colname="col5" colwidth="14mm" /><thead><row><entry valign="top" /><entry namest="col2" nameend="col5" align="left" valign="top"><b>Phase</b></entry></row><row><entry valign="top">Core Configuration</entry><entry valign="top">0</entry><entry valign="top">1</entry><entry valign="top">2</entry><entry valign="top">3</entry></row></thead><tbody><row><entry>1 x 128</entry><entry>Core 0</entry><entry>Core 0</entry><entry>Core 0</entry><entry>Core 0</entry></row><row><entry>2 x 64</entry><entry>Core 0</entry><entry>Core 1</entry><entry>Core 0</entry><entry>Core 1</entry></row><row><entry>4 x 32</entry><entry>Core 0</entry><entry>Core 1</entry><entry>Core 2</entry><entry>Core 3</entry></row><row><entry>1 x 64, 2 x 32</entry><entry>Core 0</entry><entry>Core 2</entry><entry>Core 0</entry><entry>Core 3</entry></row></tbody></tgroup></table></tables>

0134An additional bit of information is also carried with the phase signal. This is the 'phase enable' signal. Normally, this signal will always be on, indicating that the phase is active. If the signal is toggled off then it indicated that the phase is inactive. In this way a particular core can be run at a fraction of its configured speed. For example, a 32 bit device can be run in 8 bit mode by toggling its phase enable signal on only once every 4 cycles.

0135<figref idref="f0037">Figure 45</figref> demonstrates an example of using TDM primitives in a system to allow sharing of logic between multiple independent pipelines.

0136This example shows how both data path 220 and state machines 222 can be implemented with TDM primitives 200. In this example, the combination logic is used and is active on each cycle. However, the storage elements are only active on those cycles on which the phase signal indicates that it is their active phase.

0137In this example Big Endian ordering is assumed on all data buses. The first byte in a packet is always contained in the highest position on the data bus with the rest of the packet following it. <figref idref="f0037">Figure 46</figref> illustrates this visually for a 64 bit data path.

0138The switch matrix can be used in a number of different configurations. Configuration possibilities can include one or more of the following:

0139Number of XMU blocks - Determines the number of XMUs instantiated in the complete switch matrix. In a non-TDM environment, on XMU equals one local port. Configuring an XMU as a TDM device allows that XMU to be shared amongst multiple ports (up to 4).

0140TDM ports per switch matrix - Each switch matrix can be configured to support full TDM with 1-4 ports connected, half TDM with 1-2 ports connected or TDM disabled allowing only a single port to be connected. In a TDM configured switch matrix, separate buffering resources are allocated to each TDM port.

0141Packet categories - Many protocols provide for various ways of categorizing packets in order to provide separation and differentiation of traffic. This configuration parameter allows each switch matrix to be configured to support any number of packet categories. Each category of packet can have its own buffering resources allocated to it. For example, this feature can support virtual lanes in InfiniBand or virtual channels and/or various packet types (Posted/Non-Posted/Completions) in PCI-Express.

0142Ordering Rules - Some protocols, like PCI Express, enforce a set of ordering rules on packets of different categories. The XMU will be designed such that these ordering rules are encapsulated in a way that can be easily modified.

0143Buffer sizes - The total amount of buffering present in a given switch matrix is set first, then it is broken down into min/max buffer space for each port and packet category. The actual amount of space allocated to any given port/packet category can be set with configuration status registers (CSRs).

0144Multiple read ports - Each input buffer can be configured to have multiple read ports. The switch interconnect can support multiple routing paths out of each XMU. In one example, up to 2 Read Ports can be supported per XMU. This feature is useful for congestion management.

0145Available contexts - The total number of packets that the switch matrix can maintain at any given time can be specified on a per port per traffic category basis. This can also be specified as 'infinite'. This means that an XMU will always accept new packets as long as there is data credit available for the packet. This feature may, for example, not be needed in PCI Express but could be used, for example in InfiniBand, which has only data credits (no header credits).

0146Credits per context - Some protocols (PCI Express) maintain separate credit for both header and data. The XMU however, can store the entire contents of the packet in the same buffer. To support header credit, the XMU can be configured to reserve sufficient data credit to support the number of contexts that it is advertising as available. This configuration parameter specifies how much space is to be reserved for this purpose (per context). For a PCIE device this should be set to 16 bytes (largest header). For InfiniBand, this would be set to 0.

0147Maximum Transfer Unit (MTU) - The switch matrix is intended to support packets up to 8k bytes in size. It may be useful to specify a smaller supported packet size. The switch matrix can also be configured to support larger packet sizes.

0148Arbitration Priority/Algorithm - The switch matrix can implement two types of arbitration. It can perform packet category arbitration and can perform port arbitration. These arbiter blocks can be written in a modular way.

0149HoL triggers - Several HoL/congested related trigger mechanisms can be available. One example is a simple timeout value indicating the number of cycles that a packet can be present in the XMU before it is considered to be blocked.

0150Priority levels - Within each packet category optionally a priority level can be set that can be used by the routing algorithm.

0151Fixed packet length and cell interleaving can optionally be used to reduce the complexity of routing in large switches. With this feature the packets sent to the XMU can be of a fixed size. If a supported protocol does not enforce a fixed size the local port can be operable to implement the required segmentation and reassembly.

0152Accordingly, there has been described, a switch system comprises a switch matrix providing a configurable, protocol agnostic, packet switching core, the switch matrix providing a non-blocking switch connecting a configurable number of ports together.

0153Although the embodiments above have been described in considerable detail, numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated.