Buffering of multicast cells in switching networks

<u>Technical Field</u>

0001This invention relates to a method of and an apparatus for providing a multicast routing capability in a self-routing multi-stage network and, more particularly, for providing improved minimum-tree multicast routing.

<u>Background of the Invention</u>

0002An Asynchronous Transfer Mode (ATM) switching network has the goal of providing cell transport from an input port to one (unicast) or more (multicast) output ports. The capability for handling multicast connections is an essential requirement of any ATM switch architecture design. Some ATM switches use minimum-tree multicasting, where multicast cells are replicated as far downstream as possible in the switch. Specifically, a single copy of a multicast cell received in an input port is replicated at each tree vertex as many times as the number of different other vertices or outputs to which it connects. Minimum-tree multicasting minimizes the average queuing delay and the required buffer size of the switch. A continuing problem, however, in ATM switches providing minimum-tree multicasting capability is to avoid the congestion that occurs when delays in one output port affect the traffic destined to other output ports.

<u>Summary of the Invention</u>

0003The present invention is directed to apparatus for and a method of multicasting an ingress cell received at one of a plurality of input ports of a switch network to one or more of a plurality of output ports. An ingress cell is received including a data payload and a bitmap segment identifying the one or more output ports to which the data payload is to be outputted. The data payload is stored in an available cell buffer; associated with each cell buffer is a reference counter for storing a number equal to the number of ports to which that ingress cell is to be outputted. The bitmap segment is used to identify output ports to which the data payload is to be delivered, and a plurality of linked list output queues, one for each of the output ports, is established. The bitmap segment is interpreted to identify output ports to which the data payload is to be delivered and an entry is made into the associated linked list output queue. One copy of the data payload is transmitted to each of the identified output ports, according to a sequence determined by the entries in the corresponding linked list queue. At each of the singly linked list queues, an entry is deleted therefrom when the data payload is outputted to an identified output port and the reference counter is decremented by one. When the reference counter equals zero, the cell buffer is made available to another received ingress cell.

<u>Brief Description of the Drawing</u>

0004In the drawing, <ul id="ul0001" list-style="none"><li>FIG. 1 is an illustrative single-stage ATM switch in accordance with the present invention;</li><li>FIG. 2 is a block diagram of a port card;</li><li>FIG. 3 is a block diagram of a switch module ASX;</li><li>FIG. 4 is a block diagram of a switch unit AXB;</li><li>FIG. 5 shows the standard and modified ATM cell formats used by the present invention;</li><li>FIG. 6 shows a header translation table;</li><li>FIG. 7 shows a header translation lookup table;</li><li>FIG. 8 shows an example of two-level indexing using look-up tables;</li><li>FIG. 9 shows a virtual connection (translation) table used in the implementation of the present invention;</li><li>FIG. 10 shows an ABM memory;</li><li>FIG. 11 shows a multi-stage ATM switch utilizing three types of backpressure mechanisms;</li><li>FIG. 12 shows the format of a backpressure bitmap;</li><li>FIG. 13 shows an illustrative backpressure bitmap; and</li><li>FIG. 14 shows a flowchart of typical operating sequences of the present invention.</li></ul>

<u>General Description</u>

0005In the following description, each item or block of each figure has a reference designation associated therewith, the first number of which refers to the figure in which that item is first located (e.g., 110 is located in FIG. 1).

0006With reference to FIG. 1, we provide an overview description of the operation of the present switch node. The invention provides a technique for multicasting in Asynchronous Transfer Mode (ATM) or similar switches.

0007FIG. 1 illustrates a representative implementation of the invention in a three-stage switching node or system. In this system, a multiplicity of input port cards of input port unit 110 are each connected by a single internal link 120 to a central interconnecting switch unit (also referred to herein as a switch fabric) 130 that is, in turn, connected by a single internal link 140 to each of a multiplicity of output port cards of output port unit 150. The input port unit 110 illustratively includes six input port cards (0 - 5) and the output port unit 150 includes six output port cards (0 - 5). In one embodiment, input port 0 and output port 0 share the same port card 0. Six port cards are shown, each providing interconnection to 32 inputs and 32 outputs. Each port card includes ATM Layer Manager (ALM) and ATM Buffer Manager (ABM) integrated circuits. In one embodiment, the lookup (LUT1, LUT2), translation (TR), policing (POL), and header translation (HT) tables as well as the ingress counter 903 and egress counter 904 (all discussed later) are parts of a single ALM memory unit. External input 100 and output 160 links to the input and output port units are numbered top to bottom, with the 100's digit identifying the input (output) port unit and the lower two digits identifying the subport position within that input (output) port unit. For example, an input subport 1,02 is subport 2 of input port unit 1.

0008With reference to FIG. 2, there is shown a block diagram of an illustrative port card, including several (typically 30) input subport circuits (210) and output subport circuits (211), and a plurality of control logic and RAM memory blocks (201 - 208) interconnecting those subport circuits with common connecting circuits (212 and 213) to and from the switching fabric (FIG. 1 or, alternatively, FIG. 11). The control logic blocks in the current implementation comprise two integrated circuits, the ATM Layer Manager (ALM) (201 and 206) with its associated RAM memory (204), and the ATM Buffer Manager (ABM) (202, 203, 207 and 208) with its associated RAM memory (205). The input and output subports are normally connected for service to physical layer transmission interface devices (not shown). For purposes of system configuration and administration purposes and for call setup, both the ALM and the ABM are provided with microprocessor interfaces (not shown). The ALM provides ingress policing of virtual connections, routing lookups for the switching fabric, egress virtual circuit address translations, performance statistics, and other functions described in greater detail below and in the state machine descriptions 1 and 11 of the Appendix. The ABM provides ingress and egress buffering, egress multicast routing and rate control to the output subports, congestion management and flow control, additional performance statistics, and other functions described in greater detail below and in the state machine descriptions 2, 3, 9 and 10 of the Appendix.

0009With reference to FIG. 3, there is shown a block diagram of a switch module ASX, which comprises the switch fabric 130 of FIG. 1, for interconnecting six input ports IP0 - IP5 to six output OP0 - OP5 ports. The ASX module is a 6 x 6 port shared memory switching device which routes any input port to one or more output ports. It supports output queuing, multicasting, single-stage and multi-stage configurations. It also enables a non-blocking lossless switch node operation. It is configured through a microprocessor interface 301. The switch module ASX includes six input port circuits (302) and six output port circuits (303) which connect to the cell buffer unit 304. The cell buffer unit 304 connects to a routing and arbitration circuit 305, configurable linked lists 306, a queue processor 307, backpressure processor 308 and microprocessor interface 301. The routing and arbitration circuit 305 and backpressure processor 308 also interconnect to input and output ports. Six output queues OP0 -OP5 of the ASX circuit shown in FIG. 1 reside in cell buffer 304 and linked list 306 and operate in a manner that is similar to that of the ABM circuits of the port card.

0010With reference to FIG. 11, there is shown an illustrative switch node including an input port unit 1110, a three-stage switch unit (switch fabric) 1130, and an output port unit 1150. The switch node shown in FIG. 11 provides an 18 x 18 port switch fabric. The input and output port units, respectively 1110 and 1150, are implemented using the same circuitry as the input and output port units, respectively 110 and 150, shown in FIG. 1. The three-stage switch fabric 1130 is implemented having a first (input) stage 1131 including six switch modules ASX, a second (crossbar) stage 1132 including six switch modules AXB, and a third (output) stage 1133 including six switch modules ASX. Each of the six ASX modules of the first stage is configured as a 3-to-6 expander which can receive inputs from any of three port cards and which can access any of the six second stage AXB modules. Each of the six second stage AXB modules is configured as a 6 x 6 port crossbar which can access any of the six third stage ASX modules. Each of the six third stage ASX modules is configured as a 6-to-3 concentrator which can output to any of three port cards. The ASX modules are the same as those previously described with reference to FIG. 3.

0011The AXB modules shown in FIG. 4 are processor based and include six input ports 401 which interconnect to six output ports 402 and interconnected microprocessor interface 403, backpressure processor 404 and routing and arbitration circuit 405. The backpressure processor 404 and routing and arbitration circuit 405 also connect via paths 410 and 420, respectively, to the switch modules of the first 1131 and third 1133 stages. The AXB module is a 6 x 6 port crossbar interconnect device. Much of the functionality of the AXB module is similar to that of the ASX module. However, the only buffering utilized is in the input 401 and output 402 port units. The AXB module, illustratively, does not require internal RAM storage.

0012By means of our invention, an ATM cell (or similar packet) associated with a multicast request, received at any input port of a single-stage (FIG. 1) or multi-stage ATM switch system (FIG. 11), may be delivered to any combination of output ports of that switch, as defined by that multicast request, in such a way as to make use of the least possible numbers of storage sites and interconnection links within that switch to arrive at all of the desired outlets (output ports) in a single pass through the switch. To achieve this optimal use of resources, the original multicast request, D, defining a set of desired outlets, is resolved into one or more parts of an explicit set of routes constituting a minimal tree path through the interconnection links of the switch. In accordance with our invention, the representation of these optimum route elements is such that the control components of the switching system, including its storage control mechanisms and its link assignment mechanisms, can interpret them directly to achieve the desired optimal use of the system resources.

0013Our method for representing each part of the optimum route is to modify or augment the standard ATM cell (500 of FIG. 5) inputted to the switch fabric by modifying the header to include an explicit output port bitmap (e.g., 521 of FIG. 5) of the links to be traversed from its current position, through the one or more stages of the switch fabric, to the desired output ports. The pattern of the output port bitmap to be included is extracted from a translation table (FIG. 9) location identified with the multicast request - typically through reference to a virtual circuit (VC) address (VPI, VCI) within the cell header (501 of FIG. 5). In the case of a multi-stage switch fabric (e.g., FIG. 11), the output port bitmap may identify the tree path through only part of the total number of switch stages. At the end of each partial path, the part of the bitmap already traversed is discarded and a new part of the bitmap is extracted (or generated) and attached, again by reference to the VC address or other system-specific connection information provided in the implementation. This new partial path bitmap represents only elements of the subtree forward of the present switch fabric location toward the output port(s) destinations. With this hierarchical or modular technique of partial path decomposition, each bitmap segment or subtree representation may be held to a small number of bits, while the total path traversed may consist of many more links and ports than could be attached to the cell at any one time or place without compromising the storage and transmission efficiency of the switch node or system.

0014At each node of the switching system, i.e., at each point of link and/or port interconnection within the switch, a cell may have to be stored to await its turn for access to each forward link in its path. In our invention, a cell is stored once at each switch node, independent of the number of links it must traverse forward from this node. The technique for accomplishing this is, first, that the cell is stored with an added reference count field, where that reference count is automatically derived from the bitmap representation of the number of links forward from the node at which the cell is being stored. Second, as the cell is stored, the information needed to queue it for delivery to each of the links by which it is to pass forward from this node is also stored within the time necessary to serve one cell from each of the incoming links or ports. This concurrent storage of the multicast queuing information assures that if there is room for a cell to enter a node of the switch, its delivery to all forward links and/or ports from that node in its multicast tree is assured. In the implementation described, the concurrent storage of the multicast queuing information is achieved in two different ways: <ul id="ul0002" list-style="none"><li><b>1</b>. In the highest bandwidth system elements, a fully parallel enqueuing mechanism establishes entries for all of the cell's multicast destinations concurrently with the cell within the cell service time of those elements.</li><li><b>2</b>. In the lower bandwidth elements, and especially in those with large numbers of outputs, the bitmap is stored in a special multicast staging queue concurrently with the cell storage and within the cell service time. Additional timeslots are then provided within the cell service time so that a multiplicity of queue entries can be made from the multicast staging to the output queues, in accordance with the multicast destination bitmap stored with each staging queue entry. By providing a sufficient number of such extra timeslots within each cell service time, the multicast entries can be made at a rate that exceeds the rate at which all cells can physically depart the system. And with such time sharing of the queuing hardware, an overall economy of implementation can be achieved for the lower bandwidth elements.</li></ul>

<u>Detailed Description</u>

<u>Virtual Connection Information</u>

0015With joint reference to FIGS. 9 and 10, we describe the setup of a virtual connection in the switch system of FIG. 1. In response to a caller's multicast request (step 1401 of FIG. 14), the switch system establishes a multicast virtual connection (VC) identifier and initializes a Virtual Connection (or translation) table, 900 of FIG. 9, to store and update information associated with each caller's VC. Each VC table entry, illustratively, consists of three consecutive words 905-907, and has the format shown in FIG. 9. For a specific port card, the size of the Virtual Connection table is a function of the maximum number of incoming virtual connections that the port is designed to support.

0016The address 901 into the VC table 900 is calculated using the VC_index and a pointer to the base location of the table in memory, as follows:<maths id="math0001" num=""><math display="block"><mrow><mtext>VC_table_address = VC_table_base + VC_index</mtext></mrow></math><img file="EP0785698A2_D0001.tif" /></maths> This VC table address 901 is used to store information associated with this multicast virtual connection, including: <ul id="ul0003" list-style="bullet"><li>The Output Port bitmap (e.g., 36 bits wide) 905 indicates the output port(s) of the switch system to which this cell must be routed. For a unicast cell, only a single bit in the field is set to '1'; for a multicast cell, multiple bits are set to '1'. Individual bits 0 to 31 correspond to output ports 0 to 31 respectively of an output card while bits 32-35 identify which output part card 0-15.</li><li>The Delay Priority (DP) field 911 provides one of four delay priorities (0 - 3) for this virtual connection, where a value of 0 has the lowest relative latency.</li><li>The Queue of Routing field 912 is associated with an input port and is used by the ingress queuing logic to identify an output queue (Out0 - Out5 on the port card) in which to store this cell. On the port card, the port bitmap is not used to determine the output queue.</li><li>The Queue_of_Backpressure (QB) 913 is used by the ingress queuing logic to determine traffic congestion in the switch fabric and when to send multicast cells to the switch fabric.</li><li>The Connection_Tag 908 is used to calculate an address (HXLU) into a Header Translation Lookup (HTL) table, FIG. 7, for the port card to obtain an htm_index. From the htm_index, an address into the Header Translation (HT) table, FIG. 6, can be obtained. The HT table is where the outgoing or new VPI / VCI for each connection is stored and accessed. The 16-bit Connection_Tag supports up to 64K outgoing VCs. The HXLU data, HXLU format, HT table data, and the HT table format are described in a later section of this specification.</li></ul>

0017With the VC table established, the switch system (of FIG. 1 or FIG. 11) is ready to receive standard ATM cells of the requested multicast connection.

0018In step 1402, the system establishes the lookup tables LUT1, LUT2 of FIG. 8, HT table of FIG. 6, and HTL table of FIG. 7. In step 1403, the multicast cells are received by the system in the standard ATM format.

0019With reference to FIG. 5, we describe the format of the standard ATM cell 500 received at input port units 110 of FIG. 1. The ATM cell format includes a standard ATM header 501 which is five bytes (40 bits) long and a payload which is 40 bytes long. The ATM header has four bits of Generic Flow Control (GFC), eight bits of Virtual Path Identifier (VPI) 502, 16 bits of Virtual Channel Identifier (VCI) 503, three bits of Payload Type Indicator (PTI), one bit of Cell Loss Priority (CLP), and eight bits of Header Error Control (HEC).

0020The various steps 1404 through 1410 of FIG. 4 are performed by finite state machines in accordance with the present invention. The particular operating details of these various machines are described in more detail in the Appendix of this specification. The operation of a single-stage ATM switch is described by the steps 1401 - 1411 and utilizes machines 1-3 and 7-11. The operation of a multistage ATM switch involves steps 1401 - 1406, 1412, 1409 - 1411 and utilizes all of the machines 1 - 11 described in the Appendix.

0021In step 1404 (performed by machine 1), the header of the ATM cell is modified by the input port card. The modified format of the egress cell outputted (over links 120 of FIG. 1) from input port units 110 and inputted to switch unit 130 is shown by 520 to include a local header 523, a reduced-standard header 524, and a payload. In accordance with the invention, the VPI and VCI segments of the standard header 501 are used to access lookup tables LUTI and LUT2 of FIG. 8, the VC (translation) table of FIG. 9, the HTL table of FIG. 7, and HT table of FIG. 6 to generate the local header. The illustrative local header includes a plurality of segments including a priority bit (P), connection tag 522, output port map 521, queue of routing (QR), service feature information (NI,CI), policing information (POL), and an S-bit. The reduced-standard header 524 includes the GFC, VCI, PTI, and CLP segments of the standard header which are obtained during a header translation process. Header translation is described in detail in the following paragraphs.

<u>Header Translation</u>

0022Our illustrative switch node of FIG. 1 can operate with standard UNI and NNI interfaces. For a UNI interface, the five-bit subport number, the eight least significant bits (LSBs) of the VPI 502, and the 14 LSBs of the VCI 503 create a potential address space of 27 bits. For an NNI interface, an additional four VPI bits (the GFC bits of the header of ATM cell 500) are used to create a potential address space of 31 bits. Each port card maps all possible virtual connections to a maximum of 64K usable addresses per port card.

0023A Virtual Connection Index (VC_Index) is used to calculate a VC_Address into a VC table 900 (also referred to as a translation table) (FIG. 9) to obtain an output port bitmap 905 (also referred to as a multicast bitmap) associated with a particular virtual connection. Two levels of indexing are used. For a UNI interface, the first level maps each subport into a space of up to 8K unique VPI connections. For an NNI interface, the first level maps each subport into a space of up to 64K unique VPI connections. For both UNI and NNI interfaces, the second level maps each VPI connection into a space of up to 64K virtual connections. FIG. 8 shows an example of two-level indexing using lookup tables LUTI and LUT2. The subport number 801 is used as the access LUT1_Address to obtain a VP_Base and MAX_VPI numbers from the table LUT1. If the VPI 502 (received in the ATM cell 500) falls between the number VP_Base and VP_Base + MAX_VPI, the ATM cell is accepted. Otherwise, it is rejected. The VP_Base + VPI 502 is used as the access address to table LUT2 (LUT2_Address) and is used to obtain a VP_Base, MAX_VCI, and VCI_Index. The ATM cell 500 is accepted if the VPI 502 received is less than VP_Base + MAX_VCI. Otherwise, it is rejected.

0024As shown in FIG. 8, the virtual connection VC table address 901 obtained from table LUT2 is VP_Base + VC_Index which is used to access the VC (translation) table 900, policing table 902, and the ingress counters 903 of FIG. 9.

0025With reference to FIG. 9, the policing table 902 is accessed in the same manner as translation table 900 and provides memory and control information for standard "UPC" operation. That is, it is used to mark or re-mark the cells.

0026The ingress counters 903 and egress counters 904 are used to gather statistics on the result of policing.

0027The VC translation table 900 provides access for each virtual connection, using the VC table address 901, to the output port bit map 905, connection tag 908, delay priority 911, and Queue of Routing (QR) field 912.

0028The output port bitmap field 905 provides, for our exemplary multicast connection, a map from the input port 102 through the interconnecting switch unit 130.

0029Once the translation table 900 entry for a multicast connection has been accessed, header translation can occur.

0030Header translation is used to generate the local header 523 and reduced standard header 524 for egress cells read from the subport queues of input port units 110. The header translation occurs by accessing a Header Translation (HT) table 600, shown in FIG. 6, which consists of the following fields: P (odd parity over word); R (reserved); S (VP Switch Flag); vpi4 (control bit to indicate whether to replace the GFC/VPI header bits); new VPI_4 (4 most significant bits MSB of VPI); new VPI_8 (8 least significant bits LSB of VPI); and new VCI (16 bits).

0031As previously noted, new VPI and VCI values 601 were stored in the Header Translation (HT) table of FIG. 6 by the microprocessor for each outgoing VC during call setup. The control bit stored with the new VPI / VCI values, called "vpi4 bit", determines how the 4 MSBs of the new VPI are handled. These values, together with the VP switch flag S, and the ATM header 501 are used to generate the new, modified header (local 523 and reduced-standard 524 header) for the egress cell 520 of FIG. 5.

0032The PTI and CLP bits are read from the ATM header 501 and inserted into the reduced-standard header 524 of the egress cell 520. The S bit determines whether the original VCI (stored in the ATM header 501) or the new VCI (stored in the HT table 600) is inserted into the outgoing reduced-standard header 524. The vpi4 bit determines whether the GFC field (stored in the ATM header 501) or the new VPI (4 MSB bits of the new VPI stored in the HT table 600) is inserted into field 515 of reduced-standard header 524. The remaining bits of the new VPI (8 LSB bits of new VPI stored in the HT table 600) are also inserted into field 525 of reduced-standard header 524.

0033To generate modified header segments 523 and 524, access must be made to the appropriate HT table 600 entry, via a two-level lookup. The connection tag of word 907 of VC table 900 is used as an index into the Header Translation Lookup (HTL) table of FIG. 7. The address of a word in the HTL table, hxlu, is found from:<maths id="math0002" num=""><math display="block"><mrow><mtext>hxlu_address = hxlu_base + connection tag.</mtext></mrow></math><img file="EP0785698A2_D0002.tif" /></maths>

0034The hxlu_address of the HTL table stores an index (htm_index 701) that points to htt_address in HT table 600, FIG. 6, associated with the connection tag. The htt_address is used to obtain the new VPI and VCI values, as well as the S and vpi4 bits for a given egress cell which is stored in local header 523 and reduced-standard header 524.

0035The address of a word in HT table 600 is found from:<maths id="math0003" num=""><math display="block"><mrow><mtext>htt_address = htt_base + htm_index + htm_offset.</mtext></mrow></math><img file="EP0785698A2_D0003.tif" /></maths>

0036For a unicast cell, the htm_offset (also referred to herein as a multicast address offset) is defined equal to zero. For a multicast call, multiple headers, one for each multicast output port, are stored in consecutive memory locations. For each output port connection of the multicast call, the appropriate htm_offset is used. The htm_offset is used to address each of the multicast headers.

0037To support 64K egress VCs, 64K words of memory are required for the HTL table 700.

0038The format of the cell which is outputted (over links 140 of FIG. 1) from switch 130 and inputted to output port units 150 is shown as 540 in FIG. 5. The format of cell 540 includes a modified local header 542 (with the output port map segment having been replaced by an output port backpressure map), the reduced-standard header, and the payload.

0039The format of the cell which is outputted from output port units 150 is shown by 560 in FIG. 5. The cell 560 is the standard ATM cell shown by 500 except that it has a new VPI segment and may or may not have a new VCI segment.

0040According to the disclosed embodiment, all of the tables of FIGS. 6, 7, 8 and 9 are stored in an ALM memory of the ALM integrated circuit of a port card.

<u>Multicast Connection Processing</u>

0041The following example will describe multicast connection processing provided at the input port unit 110, switch unit 130 and output port unit 150.

0042As a processing example, consider a cell arriving at input port 102 (i.e., subport 1,02), associated with a multicast connection request for delivery to output ports 101 (i.e., subport 1 or port card 1, output port 01), 104, 203, 204, and 402. The delivery of the multicast cells occurs through a virtual connection (VC) established between input port 102 and the requested output ports. As previously discussed, in steps 1401 and 1402, when the multicast connection is first established, the switch system of FIG. 1, in response to the caller's multicast request, establishes the LUT1, LUT2, VC, HTL and HT tables including entries such as an output port bitmap PBM, queue number QN, delay priority DLP, and connection tag and associates VPI and VCI identifiers with all cells sent during the multicast connection. In step 1403, the VPI and VCI identifiers are then included as part of the header of each ATM cell (see 500 of FIG. 5) sent by the caller during that multicast connection. The VPI and VCI identifiers enable the input port 102 to generate, in step 1404, the modified header 523 and 524. The output port bitmap 521, identified in the modified header 523 and 524, enables the switch node to establish the desired virtual connection between the input port 102 and the identified output ports.

0043In our example, the output port bitmap for output port 101 (i.e., output card 1, output port 01) would have a '1' in bit 1 and all '0's in bits 0, 2-31 in the Output Port bitmap 905 of VC table 900 of FIG. 9 and bits 32-35 of Output Port bitmap 905 would be binary encoded to "0001" to represent output card 1.

0044The ATM cell 500 is processed by ATM Layer Manager (ALM) and ATM Buffer Manager (ABM) circuitry located in a port card. In our example, since the ATM cells arrive at input port 102, the ALM1 and ABM1 of port card 1 would process the ATM cells

0045For our exemplary multicast request, each ATM cell is to be delivered through the interconnecting switch unit 130 to output port cards 1, 2 and 4, with a representative payload bit pattern "1101".

<u>Ingress Queuing</u>

0046In step 1405 (performed by machine 2), when the cells are inputted to the port card 1 from input subports (e.g., input port 102), they are stored in ABM memory (FIG. 10). The ABM memory (FIG. 10) has two separate memory sections to store information associated with a cell: the pointer memory 1010 and cell memory 1020. Each is subdivided into areas corresponding to ingress 1011, 1021 and egress 1012, 1025 sections. The storage address for the cells is obtained from a linked list 1001 of available memory addresses (i.e., a free list). The linked list 1001, arranged on a cell basis, is used to access cell buffers in the ingress section 1021 of cell memory 1020.

0047In a well-known manner, an indexing scheme is used to keep track of each cell on linked list 1001. An index is not a physical memory address, but rather is an offset into a table. These indices are chained to form a linked list, each index being a pointer to the next entry in the linked list. The pointers are allocated and stored in the queue pointer section of linked list 1001. There is a one-to-one correspondence between the pointers in ingress linked list memory 1001 and the cells stored in cell buffer 1023. This means that the head register of a given list stores the index of the cell stored in cell memory.

0048The ingress port bitmap (word 521 of local header 523) is stored in ingress port bitmap section 1003 of pointer memory 1010.

0049Cells leaving the input port card destined for the switch fabric 1130 are read from cell memory 1020 and taken off or dequeued from an ingress linked list 1001. In each timeslot, an incoming cell is stored in cell memory 1020 and a cell from memory is sent to the switch fabric 1130.

0050To dequeue a cell, step 1406 (performed by machine 3), it is removed from an ingress (input) queue 1001 and sent to the switch fabric 1130 once per timeslot in accordance with a read selection algorithm. According to our read selection algorithm, described as machine described as machine 3 in the Appendix, queues are visited in order on a round-robin basis. If a queue is empty or if all of the delay priorities (Pr0 - Pr3) in a queue are backpressured (described in detail in machine 3), then that queue loses its opportunity and is skipped by the selection logic.

0051After step 1406 (performed by machine 3) has outputted a cell, that cell is processed by the switching fabric as described in step 1407 (performed by machine 7). When a single-stage, single device is implemented, the operation of that device, steps 1407 and 1408, is essentially similar to that described for machines 7 and 8 (described in a later paragraph). The significant difference is that queue backpressure is signaled directly to the output port card rather than indirectly through the crossbar ??? process (described as part of machines 5, 6 and 7).

0052Because the routing bitmap already covers the routing through the interconnecting switch unit 1130, no VC-to-routing map lookup function is required. The operations within the interconnecting switch unit 1130 are relatively straightforward and are similar to the previously described operation of the ABM circuit memory of FIG. 10.

0053In step 1408 (performed by machine 8), in any of the multiplicity of output queues (Out0 - Out5) of the interconnecting switch unit 1130 to which the cell in question has been assigned, after all other entries ahead of that cell in question have been served, the multicast cell is extracted from the cell buffer 1023 of the input port unit, the stored reference count is decremented by one, and the cell is delivered across the link 1140 associated with that queue to the output port units 1150 as shown in FIG. 10. If the decrement operation on the reference count results in a zero value, this indicates that all of the queue entries for that cell have been served, and the associated cell buffer in the interconnecting switch unit is marked as free.

0054In the exemplary implementation of FIG. 10, the initial routing bitmap only extends through the input port and interconnect switching stages 1130. Therefore, on entry to the output port unit, a second reference must be made to that unit's VC- to-map lookup circuit (i.e., LUT1, LUT2 and TR tables).

<u>Egress queuing</u>

0055The following section provides an overview of egress queuing on the output port card.

0056In step 1409 (performed by machine 9), cells which enter the output port card from the switch fabric 1130 are stored in the egress section of cell memory (e.g., 1025 of FIG. 10) until they can be transmitted. The egress cells are organized into 32 output subport queues with four priority linked lists per queue. Each linked list corresponds to a different delay priority.

0057The queue number QN and delay priority DLP from the virtual connection (VC) table 900, FIG. 9, identify one of egress queues and one of four lists for a given cell and are used to enqueue the cell on the appropriate egress linked list 1004.

0058For each of the lists, the following values are maintained: <ul id="ul0004" list-style="bullet"><li>head    pointer to the first cell of the list</li><li>tail    pointer to the last cell of the list</li></ul>

0059An indirect-pointer scheme is used to keep track of each cell on a list. Small blocks of index values, called index pairs (IP), are chained to form a linked list 551. Each IP contains two fields: <ul id="ul0005" list-style="bullet"><li>cell index    index to the first word of the cell data</li><li>next index    pointer to the next entry in the linked list</li></ul>

0060The IPs are allocated and stored in egress IP memory 1051 separately from the cell buffer 1023. This means that the head of a given list points to an IP: one index gives the location of the cell and the other index gives the location of the next IP. This requires two free lists: one for the IPs (IP free list) and one for the cells (egress free list).

0061The egress queuing of the output port card uses the egress pointer memory 1012 and egress cell memory 1025 of FIG. 10

0062The egress pointer memory 1012 contains the following data for the egress side: <ul id="ul0006" list-style="bullet"><li>IP free list (stored in the egress IP memory 1051)</li><li>Egress free list and multicast lists (stored in the egress linked list memory 1004)</li><li>Egress lists (32 queues × 4 = 128 subport lists; stored in egress IP memory 1051)</li><li>Egress header translation offsets (stored in the egress header translation offset table 1052)</li><li>Subport bitmaps (one per incoming connection; stored in the egress subport bitmap 1053)</li><li>Subport multicast bitmaps/multicast counters (one per egress cell that is stored in cell memory; stored in subport multicast bitmaps area 1054)</li><li>Round-robin schedule (stored in 1005)</li></ul>

0063The egress index pair (IP) memory 1051 includes an IP free list and all active queue entries for the various egress subports. The index pair (IP) free list is a linked list of index pairs (cell index and link pointer) in 1051 of the pointer memory 1010 that is defined to store egress IPs. The list consists of head and tail registers which point to the first and last entries in the list. An egress IP consists of two words, where the first word is the cell index and the second is the next index. The cell index field is 15 bits, sufficient to address up to 32K egress cells in the cell memory 1020. The first free address is found from the value of the next index stored in the head register of the egress IP free list. IPs taken from the free list and linked to the various subport egress queues are modified with the cell index of the stored cell in egress cell memory 1025. Thus, several entries on various subport queues can all reference the same cell storage location.

0064The egress linked list memory 1004 contains both the egress free list and the multicast staging linked lists. The egress free list is a linked list of free blocks in the pointer memory used by both unicast and multicast cells. The free list consists of head and tail registers which point to the first and last entries in the list. A multicast staging linked list entry consists of one word - the next index pointer. The next index pointer is sufficient to address up to 32K multicast cells in the cell memory. Each multicast list, one for each priority, consists of head and tail registers which point to the first and last cells in the list.

0065When the multicast queue links a cell onto a subport queue (e.g., Sub0), the offset associated with the subport queue is stored in 1052. The offset will be used during header translation to find the header translation entry for the particular leg of the multicast. There is one header translation offset for each cell index pair (IP) in egress IP memory 1052.

0066The subport bitmap 1054 consists of 32 bits, each bit corresponding to a subport on an egress port card. Each incoming VC connection has an associated connection tag and subport bitmap 1054. The subport bitmap 1054 is written by the port card microprocessor at the time a VC connection is set up. It is accessed after a cell arrives from the switch fabric 1130 to determine the egress subport queue to which the received cell is written. The subport bitmap for a specific connection is located in 1054 using the connection tag data corresponding to the connection. For a unicast cell, only one bit of subport bitmap 1054 is set and the cell is appended to the corresponding subport list. Subport multicast is discussed in a later paragraph.

0067In step 1410 (performed by machine 10), when a multicast cell comes to the head of a multicast list, the subport multicast bitmap 1054 is retrieved in order to determine where copies of the cell are destined. Each multicast cell in cell memory 1020 has an associated subport multicast bitmap 1054 stored in the pointer memory 1005. Once the subport multicast bitmap has been read by the entry at the head of the multicast queue, a multicast address offset is stored in its place. Each egress cell in cell memory is associated with a unique 5-bit (32-count) multicast counter 1055 which determines how many copies of the cell are to be created. The multicast count is initialized to '1' when a unicast cell is received from the switch fabric and written into cell memory. When a multicast cell is received from the switch fabric, the multicast counter is not initialized until after the subport multicast bitmap has been read and scanned by the multicast queue. The multicast counter 1055 for a given cell is decremented each time a cell is read from a subport queue. When the multicast count equals 0, the multicast cell location is moved to the egress free list. The next multicast cell in the linked list 1004 comes to the head of the multicast list.

0068The round-robin schedule consists of a variable number of entries stored in 1005 in pointer memory 1010. One entry is read per timeslot.

0069One cell is removed from a subport queue and sent to the output subport 1160 during each timeslot. The read selection algorithm is similar to that of the previously discussed ingress side. However, instead of serving the queues in strict order, the queues are serviced using a weighted round-robin schedule 1005 of variable length. Each entry in the table specifies the subport number to read this timeslot, as well as a 'no cell read' value.

0070Subport output rate is determined by the frequency at which a particular subport appears in the schedule An illustrative schedule length ( <i>L</i> ) of 461 permits data rates from 1.5 Mbps to 622 Mbps in 1.5 Mbps steps. Using <i>N</i> as the number of entries referencing a particular subport queue and <i>L</i> as the total schedule length, the possible rates are calculated as:<maths id="math0004" num=""><math display="block"><mrow><mtext mathvariant="italic">Rate</mtext><mtext> = (</mtext><mtext mathvariant="italic">N/L</mtext><mtext>) × 622.08 </mtext><mtext mathvariant="italic">Mbs</mtext><mtext> (1 ≤ </mtext><mtext mathvariant="italic">N</mtext><mtext> ≤ </mtext><mtext mathvariant="italic">L</mtext><mtext>)</mtext></mrow></math><img file="EP0785698A2_D0004.tif" /></maths>

0071The delay priorities are serviced in the subport queues in exactly the same manner as in the ingress side, using 16-entry weighted round-robin schedules previously described.

0072The data format of an egress cell from a subport as shown by 560 is the same as that of the ingress cell except that the VPI is new and VCI may or may not be new.

0073Subport multicast operations proceed as follows:

0074For each cell received from the switch fabric 130, the connection tag is used to access the subport bitmap stored in pointer memory 510. If more than one bit in the subport bitmap is set to '1', then the cell received is a multicast cell. The cell is temporarily stored on a special multicast queue until it is sequentially transferred to all of the proper subport queues on an interleaved basis with any unicast cells of other virtual connections.

0075For example, it is assumed that the queue processor has sufficient capability to enqueue two cells and dequeue one cell each timeslot. If there are any cells waiting on the multicast queue, then one of the enqueued cells is taken from the multicast queue and sent to the proper subport queue. Under full load, when a new unicast cell arrives every timeslot, this results in multicast cells interleaved 1:1 with unicast cells on a particular subport list. Of course, if an idle cell is received from the switch fabric, it can be substituted for by one of the multicast cells. A given cell remains at the head of the multicast queue until it is written to each of the subports to which it is destined, as specified in the subport bitmap. Since a multicast cell can be written to up to 32 subport queues, there may be 32 different egress IPs for the cell, all containing the same value in their respective cell_index fields, i.e., each pointing to the same cell in cell memory. When the cell is written to each of the subport queues, it is removed from the head of the multicast list and the head register of the multicast list is updated.

0076For our exemplary multicast request, the output port unit at the upper right of FIG. 10 will yield a bitmap representation for ports 101 and 104, say "1001". We take this case to describe the second implementation, that for lower performance elements, as described in paragraph <b>2</b> above. As with the original translation (in the input port unit), the presence of a valid routing map entry causes the cell to be stored in a free place in the cell buffer of the output port unit, but in this case without the multicast routing bitmap and with a reference count equal to the number of '1's in the multicast bitmap.

0077Concurrently with the storage of the cell, an entry containing the multicast routing bitmap and a pointer to the cell storage buffer is entered into the multicast staging queue of the output port unit. Typically this staging queue would also be a FIFO queue. As multicast entries come to the head of the staging queue, the output port unit logic distributes them, one at a time, to the final egress queues (lines 101 and 104 in the example), each with a copy of the pointer to the cell storage buffer. By providing a sufficiently high service rate at the output of the multicast staging queue, one copy of the control logic can serve all copies of the multicast at a rate greater than the delivery rates to the final output ports. A speedup factor of only two is sufficient to meet this objective for cell inputs in any combination of multicast requests that can possibly be delivered through the final output ports.

0078Finally, as prior queued entries are serviced at the output ports, the two copies (at 101 and 104 in the example) will be emitted and the cell buffer reference count will be decremented, and <i>as</i> above when the cell buffer reference count reaches zero, the cell buffer will be marked free.

0079Backpressure is utilized to ensure lossless operation of the switch fabric (i.e., switch units 130 and 1130) and to prevent any one output port from seizing a disproportionate share of the cell memory due to an overload condition. Backpressure is applied selectively to each delay priority of each queue. During each timeslot, the maximum number of memory locations allocated to any queue list is dynamically updated. If the size of any list exceeds that calculated threshold, a backpressure status bit corresponding to that queue list is asserted. Collectively, the status bits are referred to as the backpressure bitmap. The backpressure status for each queue list indicates whether or not a cell from a particular queue list should be transmitted to the next stage of the switch fabric or output port. Backpressure bitmaps provide congestion feedback from the egress side of output port card 1130 to third stage modules 1133; third stage modules 1133 to second stage crossbar modules 1132; second stage modules 1132 to first stage modules 1131; first stage modules 1131 to the ingress side of the output port card 1110.

0080In step 1411, performed by machine 11, output cell headers are translated based on connection tag and output offset as described in tables 6 and 7.

<u>Multi-stage switch unit configuration</u>

0081With reference to FIG. 11, we describe the three types of backpressure mechanisms utilized in the multi-stage switch unit 1150 configuration. These backpressure mechanisms represent logical paths rather than actual physical connectivity paths. For example, there is no actual physical connection from a first stage module to the ingress side of a port card.

Third stage to second stage

0082Prevention of buffer overflow on the third stage, caused by the second stage, is effected in two parts. During the first part, backpressure information is transferred from the third stage modules 1133 to each second stage crossbar module 1132. During the second part, the second stage communicates with the first stage modules 1131, as discussed in a later paragraph.

0083Every timeslot, each third stage module 1133 checks to see which Queue of Record (QR.) counters have exceeded the current dynamic threshold. The bits corresponding to lists which have exceeded the given threshold are set in a backpressure bitmap. The backpressure bitmap from each module 1133 is directly conveyed to each crossbar module 1132 by dedicated paths. Consider the 18 x 18 port switch fabric 1130. The backpressure bitmap format for modules configured to support three output ports is shown in FIG. 12. Hence, each output module transmits 12 bits of information to each crossbar module 1132 of FIG. 11. As a result, each crossbar receives a complete backpressure bitmap corresponding to up to 18 output ports once per timeslot.

Second stage to first stage

0084Returning to FIG. 11, the second stage modules 1132 to first stage modules 1131 backpressure is asserted to prevent buffer overflow on the third stage modules 1133 or to avoid output line contention on a crossbar module 1132. The second-to-first stage feedback is conveyed via a dedicated path. The second stage provides the following feedback to the first stage modules: accept; buffer full in third stage module; output line not available on second stage module.

First stage to ingress side of the port card

0085Backpressure from first stage modules 1131 to input port cards 1110 is asserted to prevent buffer overflow on the first stage modules 1131 of the switch fabric 1130. Application of backpressure is based on whether the size of a linked list (AL counter) exceeds a calculated threshold. Feedback is conveyed to the ingress side of the line cards directly connected to the given input stage module.

0086The backpressure signaling occurs in two parts. First, the backpressure bitmap is transmitted from a first stage module to the corresponding third stage module out of band. Each input stage module keeps a backpressure bitmap for the 18 queues which it handles, as shown in FIG. 13. It takes two timeslots to transfer backpressure bitmaps corresponding to each plane from the input stage to the output stage. In single-plane fabrics, each first stage module 1131 communicates backpressure information to one third stage module 1133 via a dedicated path.

0087The output stage module of an 18 x 18 port switch fabric takes the backpressure maps from each plane and multiplexes the information onto the port bitmap of outgoing ATM cells. The output module outputs bitmaps corresponding to ports 0 - 8 and 9 - 17 sequentially every two timeslots. Hence a complete 72-bit bitmap is transferred once every two timeslots. When the bitmap corresponding to ports 0 - 8 is transmitted to an outgoing ATM cell, the backpressure map sync bit is set to '1'; otherwise, the bit is set to '0'.

0088In the event that no cell is destined to a given port card, an idle cell is generated to carry the backpressure bitmap to that port card.

Egress side of port card to third stage

0089Each output port card also provides four bits of backpressure status, one for each delay priority, to its corresponding third stage module 1133 of the switch fabric 1130 in-band via cells being transmitted from the ingress side of the port card 1110 to the switch fabric. The four bits of information are transmitted serially over the course of four timeslots, starting with delay priority 0 and ending with delay priority 3. An egress backpressure sync bit (3) accompanies the information. It is set to '1' when delay priority '0' is sent. Otherwise, it is set to '0'. First stage modules strip off the two bits of backpressure information from incoming ATM cells and convey that information along with the first-stage egress side of the output port card backpressure status to the corresponding third stage modules 1133 via a dedicated path. In the event that an input port card has no cells targeted for the switch fabric, the ABM generates an idle cell to carry the delay priority backpressure information to the switch fabric. When a backpressure bit for a given delay priority and output port is asserted, the given third stage module 1133 stops selecting cells for transmission to that output port.

0090According to another embodiment of the invention, a high-performance implementation alternative, previously referred to in paragraph <b>1</b> of the General Description, is described. The multicast cell, with its attached routing bitmap, appears on the input bus at the specific time (one or more clock intervals, not necessarily contiguous) scheduled for serving the interconnecting link from its input port unit. During this reserved time, the cell is copied to a free location in the cell buffer of the interconnecting switch unit, without its multicast routing bitmap but with a reference count of the number of bits set in that bitmap. Concurrently with the copy, the bits of the bitmap are used to initiate parallel queuing operations in those queues identified by the bitmap and associated with the internal links and output port units through which the cell will ultimately be delivered. By thus providing parallel hardware and concurrent operation, the clock rates of highest throughput elements of the system can be reduced.

<u>Illustrative Embodiment of the Invention</u>

0091An illustrative, more specific implementation of the multicast routing techniques of the present invention is represented in the attached Appendix. The Appendix describes the operating sequences (in a high-level pseudo-code format) for each of eleven processing machines which together perform the functions of the present invention. For purposes of this description, a machine is defined as a controller that runs from an input to an output in a well-defined program sequence. A machine is implemented as a well-known finite state machine using state registers and combinatorial logic. When data are stored in buffers for various lengths of time, the buffers serve as outputs and then as inputs to machines, and the "sequencers" on either side are most easily understood as separate machines.

0092Before we describe the operating sequences for each of the machines, a few general statements are made which pertain to all of the machines described. The operating sequences utilize well-known high-level program control statements (e.g., if, then, else, etc.). Comments which explain particular functions or applications of the program statements are preceded by "//", e.g., //this is a comment.

0093The operation of the various machines and the data flow between these machines will be described with particular reference to the multi-stage ATM switch embodiment shown in FIG. 11 along with FIGS. 1, 3, 4 and others.

0094Machine 1 describes the operating sequence for ALM ingress operations (for ALM 0 - ALM 17 of input port unit 1110 of FIG. 11) which is performed at all input subports.

0095Machine 2 describes the operating sequence for ABM ingress buffer insertion (for ABM 0 - ABM 17 of input port unit 1110 of FIG. 11) for each cell received from an ALM.

0096Machine 3 describes the operating sequence for ABM ingress buffer unloading (for ABM 0 - ABM 17 of input port unit 1110 of FIG. 11) during each cell period.

0097Machine 4 describes the operating sequence for the first-stage ASX switching input buffer insertion (for each ASX input module 0 - 5 of FIG. 11, with joint reference to FIG. 3) for each input port.

0098Machine 5 describes the operating sequence for the first-stage ASX switching output buffer unloading (for each ASX of FIG. 11, with joint reference to FIG. 3) for all of the parallel queue lists (over their respective four priority levels).

0099Machine 6 describes the operating sequence for center-stage AXB crossbar switching device (for each AXB crossbar 0 - 5 of FIG. 11, with joint reference to FIG. 4) for each input port, cyclically from a "random" starting point and in "random" order from that point.

0100Machine 7 describes the operating sequence for third-stage ASX switching input buffer insertion (for each ASX output module 0 - 5 of FIG. 11, with joint reference to FIG. 3) for each input port to an ASX.

0101Machine 8 describes the operating sequence for third-stage ASX switching output buffer unloading (for each ASX output module 0 - 5 of FIG. 11, with joint reference to FIG. 3) for each output port of an ASX.

0102Machine 9 describes the operating sequence for ABM egress buffer insertion (for ABM 0 - ABM 17 of output port unit 1150 of FIG. 11, with joint reference to FIG. 3) for each cell received from an ASX output module.

0103Machine 10 describes the operating sequence for ABM egress buffer unloading (for ABM 0 - ABM 17 of output port unit 1150 of FIG. 11, with joint reference to FIG. 3) for each output subport, in a round-robin sequence.

0104Machine 11 describes the operating sequence for ALM egress (for ABM 0 - ABM 17 of output port unit 1150 of FIG. 11, with joint reference to FIG. 3) for each cell available from the associated ABM.

0105What has been described is merely illustrative of the application of the principles of the present invention. Other arrangements and methods can be implemented by those skilled in the art without departing from the scope of the present invention. In particular, the description in terms of a specific number of sequential machines partitioned over a specific set of four integrated circuit devices is for clarity, and the principles of operation described could be realized in other configurations using fixed or programmable logic or processors of suitable performance.<img file="EP0785698A2_D0005.tif" /><img file="EP0785698A2_D0006.tif" /><img file="EP0785698A2_D0007.tif" /><img file="EP0785698A2_D0008.tif" /><img file="EP0785698A2_D0009.tif" /><img file="EP0785698A2_D0010.tif" /><img file="EP0785698A2_D0011.tif" /><img file="EP0785698A2_D0012.tif" /><img file="EP0785698A2_D0013.tif" /><img file="EP0785698A2_D0014.tif" /><img file="EP0785698A2_D0015.tif" /><img file="EP0785698A2_D0016.tif" /><img file="EP0785698A2_D0017.tif" /><img file="EP0785698A2_D0018.tif" /><img file="EP0785698A2_D0019.tif" /><img file="EP0785698A2_D0020.tif" /><img file="EP0785698A2_D0021.tif" /><img file="EP0785698A2_D0022.tif" /><img file="EP0785698A2_D0023.tif" /><img file="EP0785698A2_D0024.tif" /><img file="EP0785698A2_D0025.tif" /><img file="EP0785698A2_D0026.tif" /><img file="EP0785698A2_D0027.tif" /><img file="EP0785698A2_D0028.tif" /><img file="EP0785698A2_D0029.tif" /><img file="EP0785698A2_D0030.tif" /><img file="EP0785698A2_D0031.tif" /><img file="EP0785698A2_D0032.tif" /><img file="EP0785698A2_D0033.tif" /><img file="EP0785698A2_D0034.tif" /><img file="EP0785698A2_D0035.tif" /><img file="EP0785698A2_D0036.tif" /><img file="EP0785698A2_D0037.tif" /><img file="EP0785698A2_D0038.tif" /><img file="EP0785698A2_D0039.tif" /><img file="EP0785698A2_D0040.tif" /><img file="EP0785698A2_D0041.tif" /><img file="EP0785698A2_D0042.tif" /><img file="EP0785698A2_D0043.tif" /><img file="EP0785698A2_D0044.tif" />