Network interconnect as a switch

Abstract

An interconnect as a switch module (“ICAS” module) comprising n port groups, each port group comprising n-1 interfaces, and an interconnecting network implementing a full mesh topology where each port group comprising a plurality of interfaces each connects an interface of one of the other port groups, respectively. The ICAS module may be optically or electrically implemented. According to the embodiments, the ICAS module may be used to construct a stackable switching device and a multi-unit switching device, to replace a data center fabric switch, and to build a new, high-efficient, and cost-effective data center.

Claims

1. A data center network, comprising: a plurality of network pods, a plurality of first spine planes, and a plurality of second spine planes; wherein each of the plurality of network pods comprises a first ICAS module and a plurality of first layer switches; wherein either the plurality of first spine planes or the plurality of second spine planes is deployed as a data center design option; wherein the first ICAS module comprises: n port groups, each port group comprising n-1 interfaces, wherein n is an integer equal or larger than 3; an interconnecting network implementing a full mesh topology, wherein each of the n port groups connects one of the n-1 interfaces to another of the n port groups statically, respectively; wherein the n port groups are indexed with an integer from 0 to n-1; wherein the n-1 interfaces of the n port groups are labeled with the same indexes as those of connected n port groups; wherein an interface with index j of one of the n port groups with index i is connected to an interface with index i of one of the n port groups with index j, where i is in the range of 0 to n-1, j is in the range 0 to n-1, wherein i does not equal to j, and wherein the interconnecting network comprises all connections between the n port groups; interfaces of each of the plurality of first layer switches are configured to be grouped into n port groups one of intralink and interlink respectively; wherein a number of the plurality of first layer switches is n, and wherein the plurality of first layer switches is indexed with an integer from 0 to n-1; wherein the plurality of first spine planes and network pods are interconnected through interlinks; wherein interlinks k of the network pods each connects one of the interlinks of kth of the plurality of first spine plane, respectively, and wherein the interlinks p of the plurality of first spine planes each connects one of the interlinks of the pth network pod, respectively; a plurality of downlinks to receive and transmit data signals to and from a plurality of servers; wherein the plurality of second spine planes and the plurality of network pods are interconnected through interlinks; wherein interlinks k of the network pods each connects one of the interlinks of the kth second spine plane, respectively, and wherein the interlinks p of the second spine planes each connects one of the interlinks of the pth network pod, respectively; and wherein a plurality of downlinks to receive and transmit data signals to and from a plurality of servers.

2. The data center network of claim 1, wherein the network pod further comprises: a plurality of first layer switches whose interfaces are divided into downlink interfaces, interlink interfaces and intralink interfaces, wherein the downlink interfaces are configured to receive and transmit data signals to and from a plurality of servers, wherein the interlink interfaces of each of the plurality of first layer switches are configured into n port group of interlink, wherein the intralink interfaces of each of the plurality of first layer switches is configured into n port group of intralinks.

3. The data center network of claim 1, wherein interfaces of second layer devices of the first ICAS module are divided into intralink interfaces and uplink interfaces, wherein the intralink interfaces of the first ICAS module are grouped into the n port groups to connect to the intralink interfaces of the corresponding n port groups of the plurality of first layer switches, and wherein the uplink interfaces are configured to connect to an external network.

4. The data center network of claim 1, wherein each of the plurality of the first spine planes comprises a fanout cable transpose rack, wherein the fanout cable transpose rack comprises: k groups of first fiber adapters, each adapter of the k groups of the first fiber adapters comprising m interfaces, wherein the k groups of the first fiber adapters connect to corresponding ones of k switches through k groups of first fiber cables, wherein the k groups of the first fiber adapters also connect to an fiber adapter mounting panel by k groups of first fanout fiber cables, wherein each of a group of ┌p/m┐ first fiber adapters connect to a corresponding group of ┌p/m┐ fiber adapters of each switch by ┌p/m┐ first fiber cables, wherein each of the group of ┌p/m┐ first fiber adapters connect to the fiber adapter mounting panel by a group of ┌p/m┐ first fanout fiber cables, wherein ┌┐ is a ceiling function; and p groups of second fiber adapters, each adapter of p groups of the second fiber adapters comprising m interfaces, wherein the p groups of the second fiber adapters connect p groups of second fiber cables to form p groups of interlinks, wherein the p groups of the second fiber adapters also connect to the fiber adapter mounting panel by p groups of second fanout fiber cables, wherein each of a group of ┌k/m┐ second fiber adapters connects to a group of ┌k/m┐ second fiber cables to form an interlink, wherein each of the groups of the ┌k/m┐ second fiber adapters connects to the fiber adapter mounting panel by a group of ┌k/m┐ second fanout fiber cables, wherein ┌┐ is a ceiling function; wherein the fiber adapter mounting panel, the k groups of the first fanout fiber cables and the p groups of the second fanout fiber cables are cross-connected on the fiber adapter mounting panel, through cross-connection, wherein connections from the k switches are grouped into p interlinks, each interlink containing one connection from each of the k switches, with a total of k connections per interlink; and a plurality of third layer switches.

5. The data center network of claim 4, wherein the fanout cable transpose rack connects to the plurality of third layer switching devices through a plurality of fiber cables; and wherein the fanout cable transpose rack comprises a plurality of interlinks in the fanout cable transpose rack to connect to a plurality of network pods; wherein connections from the plurality of third layer switching devices are grouped into the plurality of interlinks in the fanout cable transpose rack through the fanout cable transpose rack.

6. The data center network of claim 5, wherein each of the plurality of interlinks in the fanout cable transpose rack contains one connection from one of the third layer switching devices, with each interlink in the fanout cable transpose rack having a number of connections, wherein the number equals to a number of the third layer switching devices.

7. The data center network of claim 4, wherein the kth spine plane interconnects the plurality of first layer switches of each of the network pods through the kth interlinks working under a full mesh connection has a characteristics from the network with an (n-1, n) bipartite graph, wherein the network with (n-1, n) bipartite graph comprises the predetermined number of the plurality of first layer switches and the predetermined number minus one n-1 of the third layer spine switches, and wherein, the plurality of first layer switches and the third layer spine switches interconnect in a Clos topology.

8. The data center network of claim 1, wherein each of the plurality of second spine planes further comprise a second ICAS module; wherein the second ICAS module comprises: n port groups, each port group comprising n-1 interfaces, wherein n is an integer equal or larger than 3; an interconnecting network implementing a full mesh topology, wherein each of the n port groups connects one of the n-1 interfaces to another of the n port groups statically, respectively; wherein the n port groups are indexed with an integer from 0 to n-1; and wherein the n-1 interfaces of the n port groups are labeled with the same indexes as those of connected n port groups; wherein an interface with index j of one of the n port groups with index i is connected to an interface with index i of one of the n port groups with index j, where i is in the range of 0 to n-1, j is in the range 0 to n-1, wherein i does not equal to j, and wherein the interconnecting network comprises all connections between the n port groups.

9. The data center network of claim 8, wherein the second ICAS modules comprises interfaces divided into interlink interfaces and uplink interfaces, wherein the interlink interfaces of the second ICAS modules is grouped into a plurality of n port groups, wherein each of the plurality of n port groups of the second ICAS modules connects to the corresponding one of the n port groups of the first layer switches of each of the plurality of the network pods; and wherein the uplink interfaces are configured to connect to an external network.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1a illustrates congestion due to hash collision in a fat tree network under ECMP.

(2) FIG. 1b illustrates aggregation congestion in a fat tree network topology.

(3) FIG. 1c illustrates congestion due to blocking condition in a fat tree network.

(4) FIG. 2a shows the architecture of a state-of-the-art data center network.

(5) FIG. 2b shows in detail an implementation of a spine plane of the data center network of FIG. 2a.

(6) FIG. 2c shows in detail an implementation of a server pod of FIG. 2a using four fabric switches to distribute machine-to-machine traffic across 48 top-of-rack switches.

(7) FIG. 2d shows in detail an implementation of an edge pod of FIG. 2a using four edge switches to provide interfaces for uplink to connect to external network.

(8) FIG. 3 illustrates a “full mesh” topology in a network of 9 nodes.

(9) FIG. 4a shows ICAS module 400, which interconnects 9 nodes, according to the full mesh topology of FIG. 3.

(10) FIG. 4b illustrates the connectivity between the internal interfaces and the external interfaces of port group7 of the 9-node ICAS module 400, in accordance with one embodiment of the present invention.

(11) FIG. 5a shows ICAS module 500 connecting port group 2 of each of TOR switches 51-0 to 51-8 in a full mesh topology, in accordance with one embodiment of the present invention.

(12) FIG. 5b illustrates, in the full mesh topology network 500 of FIG. 5a, port group 2 of TOR switch 51-1 routing a data packet to port group 2 of TOR switch 51-7 through internal interface 52-1-7 of port group 50-1 and internal interface 52-7-1 of port group 50-7 of ICAS2 module 500, in accordance with one embodiment of the present invention.

(13) FIG. 6a shows network 600, which is a more compact representation of the network of FIG. 5a.

(14) FIG. 6b shows network 620, after additional ICAS modules are added to network 600 of FIG. 6a, so as to provide greater bandwidth and path diversity.

(15) FIG. 7a shows that, in the architecture of the data center of FIG. 2a, the topology of a server pod may be reduced to a (4, 48) bipartite graph.

(16) FIG. 7b shows, as an example, network 720 represented as a (5, 6) bipartite graph.

(17) FIG. 7c shows the 6-node full mesh graph embedded in the (5, 6) bipartite graph of FIG. 7b.

(18) FIG. 8a shows an improved data center network 800, in accordance with one embodiment of the present invention; data center networks 800 includes 20 spine planes, providing optional uplinks 801, and 188 server pods, providing optional uplinks 802, uplinks 801 and 802 connecting to one or more external networks.

(19) FIG. 8b shows in detail an implementation of modified spine plane 820, having 20 spine switches, providing optional uplink 821 for connecting to an external network.

(20) FIG. 8c shows in detail an implementation of modified server pod 830 in a (20, 21) fabric/TOR topology, having 20 fabric switches for distributing machine-to-machine traffic across 20 top-of-rack switches, in accordance with one embodiment of the present invention; the 21.sup.st TOR switch is removed from the modified server pod 830 so that the connections are provided as optional uplink 831 for connecting the fabric switches to an external network.

(21) FIG. 9a shows ICAS-based data center network 900, achieved by replacing the server pods of network 800 of FIG. 8a (e.g., server pod 830 of FIG. 8c) with ICAS pods 91-0 to 91-197, each ICAS pod being shown in greater detail in FIG. 9c, according to one embodiment of the present invention; in FIG. 9a, optional uplinks 901, shared by 20 spine planes, and optional uplinks 902, shared by 188 ICAS pods are provided for connecting to an external network.

(22) FIG. 9b shows in detail spine plane 920, which implements one of the spine planes in data center network 900 and which is achieved by integrating a fanout cable transpose rack into spine plane 820 of FIG. 8b, according to one embodiment of the present invention; the spine switches in spine plane 920 provide optional uplink 921 for connecting to an external network.

(23) FIG. 9c shows in detail an implementation of ICAS pod 930, which is achieved by replacing fabric switches 83-0 to 83-19 in server pod 830 of FIG. 8c, according to one embodiment of the present invention; each ICAS pod provides 20×10G uplinks 932 for connecting to an external network.

(24) FIG. 9d illustrates a spine switch implemented with a single chip high-radix (i.e., a high port count) switching integrated circuit; such a spine switch makes use of the highest port count switching integrated circuit available at present time.

(25) FIG. 9e shows a spine switch formed by stacking together 4 switch boxes implemented with a Trident-II ASICs (96×10G configuration each) and 1 ICAS box 953. ICAS box 953 combines 4 ICAS modules 95-0 to 95-3 in one 1U chassis. Each ICAS module contains 3 copies of ICAS1X5 configuration. Together the ICAS box 953 provides non-blocking 1:1 subscription ratio to each of the 4 switches 96-0 to 96-3.

(26) FIG. 9f shows a spine switch of an ICAS-based multi-unit switching device where 4 ICAS-based fabric cards 97-0 to 97-3 get connected in a full mesh topology to switching ASIC's 98-0 to 98-3. Switching ASIC 98-0 and 98-1 are housed in line card 973, and switching ASIC's 98-2 and 98-3 are housed in line card 974.

(27) To facilitate cross-referencing among the figures and to simplify the detailed description, like elements are assigned like reference numerals.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

(28) The present invention simplifies the network architecture by eliminating the switches in the fabric layer based on a new fabric topology, referred herein as the “interconnect-as-a-switch” (ICAS) topology. The ICAS topology of the present invention is based on the “full mesh” topology. In a full mesh topology, each node is connected to all other nodes. The example of a 9-node full mesh topology is illustrated in FIG. 3. The inherent connectivity of a full mesh network can be exploited to provide fabric layer switching.

(29) As discussed in further detail below, the ICAS topology enables a data center network that is far superior to a network of the fat tree topology used in prior art data center networks. Unlike other network topologies, the ICAS topology imposes a structure on the network which reduces congestion in a large extent. According to one embodiment, the present invention provides an ICAS module as a component for interconnecting communicating devices. FIG. 4a shows ICAS module 400, which interconnects 9 nodes according to the full mesh topology of FIG. 3.

(30) FIG. 4a shows ICAS module 400 having port groups 40-0 to 40-8 and each port group providing 8 external interfaces and 8 internal interfaces. In ICAS module 400, each of the internal interfaces of a port group connects an internal interface of one of the other port groups, respectively. In fact, each port group is connected to every one of the other port groups through exactly one internal interface. In this context, each “interface ” includes a receive-transmit pair of optical fibers capable of, for example, a 10 Gbits per second data rate. In FIG. 4a, the port groups are indexed as 0-8. Indexes can also be arbitrary unequal values (For example, these 9 port groups can also be indexed as 5, 100, 77, 9, 11, 8, 13, 50, and 64). The 8 internal interfaces for these 9 port groups are indexed according to the indexes of the connected port groups (For example, the internal interfaces for the 7-th port group are 0, 1, 2, 3, 4, 5, 6 and 8 in the first example; and are 5, 100, 77, 9, 11, 8, 13 and 64 in the second example). Furthermore, internal interface j of port group i is connected to internal interface i of port group j. The external interfaces for each port group of ICAS module 400 are indexed sequentially as 0-7.

(31) FIG. 4b illustrates in detail the connectivity between the internal interfaces and the external interfaces of a port group 7 in ICAS module 400, in accordance with the present invention. As shown in FIG. 4b, in one embodiment, the external interfaces are connected one-to-one to the internal interfaces sequentially in the index order (For example, for port group 7, external interfaces 42-0 to 42-7 are sequentially connected to internal interfaces 41-0 to 41-6 and 41-8). Therefore, for port group i, external interfaces 0-7 are connected to internal interfaces 0, . . . , i-1, i+1, . . . , and 8 respectively. Therefore, it can be easily seen that any pair of port groups x and y are connected through internal interface x of port group y and internal interface y of port group x. This indexing scheme allows an external switching device to assign routes for data packets using the internal interface indices of the source port group and destination port group. No congestion condition (e.g., due to hash collision, aggregation model, or strict-sense blocking) can occur between any pair of port groups.

(32) The internal interconnection between the port groups of the ICAS module can be realized via an optical media to achieve a full mesh structure. The optical media may be an optical fiber and/or 3D MEMS. The 3D MEMS uses a controllable micro-mirror to create an optical path to achieve a full mesh structure. In both of these implementations MPO connectors are used. Alternatively, the ICAS module may also be electrically implemented using circuits. In this manner, the port groups of the ICAS module are soldered or crimped onto a PCB using connectors that support high-speed differential signals and impedance matching. The interconnection between the port groups is implemented using a copper differential pair on the PCB. Since signal losses significantly vary between different grades of high-speed differential connectors and between copper differential pairs on different grades of PCBs, an active chip is usually added at the back end of the connector to restore and enhance the signal to increase the signal transmission distance on the PCB. Housing the ICAS module in a 1U to multi-U rackmount chassis will form a 1U to multi-U interconnection device. The ICAS-based interconnection devices are then interconnected with switching devices to form a full mesh non-blocking network. This novel network will be explained in detail hereunder in a plurality of embodiments. When the ICAS module of the 1U to multi-U interconnection device is optically implemented (based on optical fiber and 3D MEMS), MPO-MPO cables are used to connect the ICAS-based interconnection devices and the switching devices. When the ICAS module of the 1U to multi-U interconnection device is electrically implemented as circuits (based on PCB+chip), DAC direct connection cables or AOC active optical cables are used to connect the ICAS-based interconnection devices and the switching devices.

(33) As switching in ICAS module 400 is achieved passively by its connectivity, no power is dissipated in performing the switching function. Typical port group-to-port group delay through an ICAS passive switch is around 10 ns (e.g., 5 ns/meter, for an optical fiber), making it very desirable for a data center application, or for big data, AI and HPC environments.

(34) The indexing scheme of external-to-internal connectivity in ICAS module 400 of FIG. 4a is summarized in Table 2 below:

(35) TABLE-US-00002 TABLE 2 Index of External Interface ICAS Port Group 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 8 1 0 2 3 4 5 6 7 8 2 0 1 3 4 5 6 7 8 3 0 1 2 4 5 6 7 8 4 0 1 2 3 5 6 7 8 5 0 1 2 3 4 6 7 8 6 0 1 2 3 4 5 7 8 7 0 1 2 3 4 5 6 8 8 0 1 2 3 4 5 6 7

(36) FIG. 5a shows network 500, in which ICAS module 510 and port group 2 of each of TOR switches 51-0 to 51-8 interconnects in a full mesh topology, in accordance with one embodiment of the present invention.

(37) As illustrated in FIG. 5b, in an ICAS module 510 in the full mesh topology network 500 of FIG. 5a, port group 51-1 of TOR switch 1 routes a data packet to port group 51-7 of TOR switch 7 through external interface 53-1-6 and internal interface 52-1-7 of port group 50-1 of ICAS module 510, and internal interface 52-7-1 and external interface 53-7-1 of port group 50-7 of ICAS module 510, in accordance with one embodiment of the present invention. As shown in FIG. 5b, TOR switch 51-1, which is connected to port group 50-1 of ICAS module 510, receives a data packet with a destination reachable through internal port group 52-1-7 of ICAS module 510. TOR switch 51-1 has a port group that includes 8 interfaces 54-1-0 to 54-1-7 (provided as two QSFP ports) mapping one-to-one to external interfaces 53-1-0 to 53-1-7 of port group 50-1 of ICAS module 510, which in turn maps one-to-one to internal interfaces 52-1-0, 52-1-2 to 52-1-8 in sequential order of port group 50-1 of ICAS module 510. TOR switch 51-7 has a port group that includes 8 interfaces 54-7-0 to 54-7-7 (provided as two QSFP ports) mapping one-to-one to external interfaces 53-7-0 to 53-7-7 of port group 50-7 of ICAS module 510, which in turn maps one-to-one to internal interfaces 52-7-0 to 52-7-6 and 52-7-8 in sequential order of port group 50-7 of ICAS module 510. Each interface in a TOR switch port may be a 10G interface, for example. As port groups 50-1 and 50-7 of ICAS module 510 are connected through the port groups' corresponding internal interfaces 52-1-7 and 52-7-1, TOR switch 51-1 sends the data packet through its interface 54-1-6 to external interface 53-1-6 of ICAS module 510. Since the connectivity in the ICAS module 510 adopts a full mesh topology, the data packet is routed to external interface 53-7-1 of ICAS module 510.

(38) In full mesh topology network 500, the interfaces of each TOR switch is regrouped into port groups, such that each port group contains 8 interfaces. To illustrate this arrangement, port group 2 from each TOR switch connects to ICAS module 510. As each TOR switch has a dedicated path through ICAS module 510 to each of the other TOR switches, no congestion can result from two or more flows from different source switches being routed to the same port of destination switch (the “Single-Destination-Multiple-Source Traffic Aggregation” case). In that case, for example, when TOR switches 51-0 to 51-8 each have a 10-G data flow that has TOR switch 51-0 as destination, all the flows would be routed on paths through respective interfaces. Table 3 summarizes the separate designated paths:

(39) TABLE-US-00003 TABLE 3 ICAS ICAS Source Destination Source Internal Internal Destination T1.p2.c0 .Math. ICAS2.pl.c0 .Math. ICAS2.p0.c1 .Math. T0.p2.c0 T2.p2.c0 .Math. ICAS2.p2.c0 .Math. ICAS2.p0.c2 .Math. T0.p2.c1 T3.p2.c0 .Math. ICAS2.p3.c0 .Math. ICAS2.p0.c3 .Math. T0.p2.c2 T4.p2.c0 .Math. ICAS2.p4.c0 .Math. ICAS2.p0.c4 .Math. T0.p2.c3 T5.p2.c0 .Math. ICAS2.p5.c0 .Math. ICAS2.p0.c5 .Math. T0.p2.c4 T6.p2.c0 .Math. ICAS2.p6.c0 .Math. ICAS2.p0.c6 .Math. T0.p2.c5 T7.p2.c0 .Math. ICAS2.p7.c0 .Math. ICAS2.p0.c7 .Math. T0.p2.c6 T8.p2.c0 .Math. ICAS2.p8.c0 .Math. ICAS2.p0.c8 .Math. T0.p2.c7

(40) In other words, in Table 3, the single-connection data between first layer switch i connected to the port group with index i and first layer switch j connected to the port group with index j is directly transmitted through the interface with index j of the port group with index i and the interface with index i of the port group with index j.

(41) In Table 3 (as well as in all Tables herein), the switch source and the switch destination are each specified by 3 values: Ti.p.sub.j.c.sub.k, where T.sub.i is the TOR switch with index i, p.sub.j is the port group with index j and c.sub.k is the interface with index k. Likewise, the source interface and destination interface in ICAS module 500 are also each specified by 3 values: ICASj.p.sub.i.c.sub.k, where ICASj is the ICAS module with index j, p.sub.i is the port group with index i and c.sub.k is the internal or external interface with index k.

(42) An ICAS-based network is customarily allocated so that when its port groups are connected to port group i from all TOR switches the ICAS will be labeled as ICASi with index i.

(43) Congestion can also be avoided in full mesh topology network 500 with a suitable routing method, even when a source switch receives a large burst of aggregated data (e.g., 80 Gbits per second) from all its connected servers to be routed to the same destination switch (the “Port-to-Port Traffic Aggregation” case). In this case, it is helpful to imagine the TOR switches as consisting of two groups: the source switch i and the rest of the switches 0 to i−1, i+1 to 8. The rest of the switches are herein collectively referred to as the “fabric group”. Suppose TOR switch 51-1 receives 80 Gbits per second (e.g., 8 10G flows) from all its connected servers all designating to destination TOR switch 51-0. The routing method for the Port-to-Port Traffic Aggregation case allocates the aggregated traffic to its 8 10G interfaces with port group 51-1 as in FIG. 5a, such that the data packets in each 10G interface is routed to a separate TOR switch in the fabric group (Table 4A):

(44) TABLE-US-00004 TABLE 4A ICAS ICAS Source Destination Source Internal Internal Destination T1.p2.c0 .Math. ICAS2.p1.c0 .Math. ICAS2.p0.c1 .Math. T0.p2.c0 T1.p2.c1 .Math. ICAS2.p1.c2 .Math. ICAS2.p2.c1 .Math. T2.p2.c1 T1.p2.c2 .Math. ICAS2.p1.c3 .Math. ICAS2.p3.c1 .Math. T3.p2.c1 T1.p2.c3 .Math. ICAS2.p1.c4 .Math. ICAS2.p4.c1 .Math. T4.p2.c1 T1.p2.c4 .Math. ICAS2.p1.c5 .Math. ICAS2.p5.c1 .Math. T5.p2.c1 T1.p2.c5 .Math. ICAS2.p1.c6 .Math. ICAS2.p6.c1 .Math. T6.p2.c1 T1.p2.c6 .Math. ICAS2.p1.c7 .Math. ICAS2.p7.c1 .Math. T7.p2.c1 T1.p2.c7 .Math. ICAS2.p1.c8 .Math. ICAS2.p8.c1 .Math. T8.p2.c1

(45) Note that the data routed to TOR switch 51-0 has arrived at its designation and therefore would not be routed further. Each TOR switch in the fabric group, other than TOR switch 51-0, then allocates its interface 0 for forwarding its received data to TOR switch 51-0 (Table 4B):

(46) TABLE-US-00005 TABLE 4B ICAS ICAS Source Destination Source Internal Internal Destination — .Math. — .Math. — .Math. — T2.p2.c0 .Math. ICAS2.p2.c0 .Math. ICAS2.p0.c2 .Math. T0.p2.c1 T3.p2.c0 .Math. ICAS2.p3.c0 .Math. ICAS2.p0.c3 .Math. T0.p2.c2 T4.p2.c0 .Math. ICAS2.p4.c0 .Math. ICAS2.p0.c4 .Math. T0.p2.c3 T5.p2.c0 .Math. ICAS2.p5.c0 .Math. ICAS2.p0.c5 .Math. T0.p2.c4 T6.p2.c0 .Math. ICAS2.p6.c0 .Math. ICAS2.p0.c6 .Math. T0.p2.c5 T7.p2.c0 .Math. ICAS2.p7.c0 .Math. ICAS2.p0.c7 .Math. T0.p2.c6 T8.p2.c0 .Math. ICAS2.p8.c0 .Math. ICAS2.p0.c8 .Math. T0.p2.c7

(47) In other words, at least one multi-connection data between the first layer switch i connected to the port group indexed i and the first layer switch j connected to the port group indexed j is transmitted through the first layer switches connected to at least one of the port groups other than the port group with source index. The multi-connection data arriving at the destination switch will cease to be further routed and transmitted.

(48) To put it more precisely, the multi-connection data transmission occurring between first layer switch i connected to the port group with index i and first layer switch j connected to the port group with index j includes the transmissions includes: as in Table 4A, the first layer switch i is connected, via a plurality of interfaces of the port group with a plurality of index i, to a plurality of first layer switches with a plurality of corresponding indexes for transmission; as in Table 4B, a plurality of the first layer switches with the indexes as shown are connected, via interfaces with index j of the port groups, to the interfaces with the indexes as shown of the port groups with index j of the first layer switches for transmission; those transmissions that arrive at a destination switch will stop routing.

(49) Thus, the full mesh topology network of the present invention provides performance that is in stark contrast to prior art network topologies (e.g., fat tree), in which congestions in the fabric switch cannot be avoided under Single-Destination-Multiple-Source Traffic Aggregation and Port-to-Port Traffic Aggregation cases.

(50) Also, as discussed above, when TOR switches 51-0 to 51-8 abide by the rule m≥2n-2, where m is the number of network-side interfaces (e.g., the interfaces with a port group in ICAS module 500) and n is the number of the TOR switch's input interfaces (e.g., interfaces to the servers within the data center), a strict blocking condition is avoided. In other words, a static path is available between any pair of input interfaces under any traffic condition. Avoiding such a blocking condition is essential in a circuit-switched network, but is not necessarily significant in a flow-based switched network.

(51) In the full mesh topology network 500 of FIG. 5a, each port group with 8 interfaces of ICAS module 500 connects to a port group with 8 interfaces (e.g., 8 10-G interfaces) of a corresponding TOR switch. Full mesh topology network 500 of FIG. 5a may be redrawn in a more compact form in FIG. 6a, with a slight modification. FIG. 6a illustrates ICAS2 module 60-2 interconnecting to port group 2 of each of TOR switches 61-0 to 61-8. In FIG. 6a, the interfaces between port group 2 of TOR switch 61-0 and port group 0 of ICAS module 60-2 (now labeled ‘ICAS2’) are represented as a single line (e.g., the single line between port group 2 of TOR switch 61-0 and port group 0 of ICAS module 60-2). Such a line, of course, represents all 8 eight interfaces between the TOR switch and a corresponding port group in ICAS module 60-2. This is exactly the case in FIG. 6b where each TOR switch 63-0 to 63-8 is shown also to have 4 port groups, to allow configuring network 620 of FIG. 6b, where three additional ICAS modules 62-0, 62-1 and 62-3 in addition to 62-2 and corresponding interfaces are added to network 600 of FIG. 6a.

(52) In full mesh topology network 500, uniform traffic may be spread out to the fabric group and then forwarded to its destination. In network 620 of FIG. 6b, the additional ICAS modules may be used to provide greater bandwidth. So long as the additional port groups are available in the TOR switches, additional ICAS modules may be added to the network to increase path diversity and bandwidth.

(53) The inventor of the present invention investigated in detail the similarities and the differences between the full mesh topology of the present invention and other network topologies, such as the fat tree topology in the data center network of FIGS. 2a. The inventor first observes that, in the architecture of the data center network of FIG. 2a, the fat tree network represented in a server pod (the “fabric/TOR topology”) can be reduced to a (4, 48) bipartite graph, so long as the fabric switches merely perform an interconnect function for traffic originated among the TOR switches. This (4, 48) bipartite graph is shown in FIG. 7a. In FIG. 7a, the upper set of nodes, nodes 0-3 (“fabric nodes”) 70-0 to 70-3, represent the four fabric switches in the server pod of FIG. 2a and the lower set of 48 nodes (i.e., leaf 0-47), labeled 71-0 to 71-47, represent the 48 TOR switches in a server pod of FIG. 2a.

(54) The inventor discovered that an n-node full mesh graph is embedded in a fabric-leaf network represented by a bipartite graph with (n-1, n) nodes (i.e., a network with n-1 fabric nodes and n TOR switch leaves). FIG. 7b shows, as an example, a (5, 6) bipartite graph with 5 nodes 72-0 to 72-4 and 6 leaves 73-0 to 73-5. FIG. 7c shows the 6-node full mesh graph 740 with 6 nodes 74-0 to 74-5 embedded in the (5, 6) bipartite graph of FIG. 7b.

(55) This discovery leads to the following rather profound results: (a) An n-node full mesh graph is embedded in an (n-1, n)-bipartite graph; and the (n-1, n) bipartite graph and the data center Fabric/TOR topology have similar connectivity characteristics; (b) A network in the (n-1, n) Fabric/TOR topology (i.e., with n-1 fabric switches and n TOR switches) can operate in same connectivity characteristics as a network with full mesh topology (e.g., network 500 of FIG. 5a); (c) Fabric switches are unnecessary in an (n-1, n) Fabric/TOR topology network, as the fabric switches merely performs interconnecting function among the TOR switches (i.e., these fabric switches can be replaced by direct connectivity among TOR switches); (d) A data center network based on a fat tree topology (e.g., the Fabric/TOR topology) can be improved significantly using ICAS modules.

(56) In the following, a data center network that incorporates ICAS modules in place of fabric switches may be referred to as an “ICAS-based” data center network. An ICAS-based data center network has the following advantages: (a) less costly, as fabric switches are not used; (b) lower power consumption, as ICAS modules are passive; (c) less congestion; (d) lower latency; (e) effectively less network layers (2 hops less for inter-pod traffic; 1 hop less for intra-pod traffic); (f) greater scalability as a data center network.

(57) These results may be advantageously used to improve typical state-of-the-art data center networks. FIG. 8a shows an improved data center network 800, in accordance with one embodiment of the present invention. Data center network 800 uses the same types of components as the data center network of FIG. 2a (i.e., spine switches, fabric switches and TOR switches), except that the number of fabric switches are increased to one less than the number of TOR switches (FIG. 8c shows equal number of fabric switches and TOR switches because one of the TOR switch, the 21.sup.st TOR switch, is removed so that the 20 interfaces connected to it from the 20 fabric switches are provided as uplink to connect to external network).

(58) FIG. 8a shows the architecture of an improved data center network, organized by three layers of switching devices—i.e., “top-of-rack” (TOR) switches and fabric switches implemented in 188 server pods 81-0 to 81-187 and spine switches implemented in 20 spine planes 80-0 to 80-19—interconnected by interlinks in a fat tree topology. An interlink refers to the network connections between a server pod and a spine plane. For example, interlink k of each of the 188 server pods is connected to spine plane k; interlink p of each of the 20 spine planes is connected to server pod p. The 20 spine planes each provide an optional uplink (e.g. uplink 801) and the 188 server pods each provide an optional uplink (e.g., uplink 802) for connection to one or more external networks. In this example, to allow comparison, the numbers of server pods and spine plane are chosen so that the improved data center network 800 and the state-of-the-art data center network 200 have the same network characteristics (2.2 Pbps total server-side bandwidth; 3:1 oversubscription ratio—server-side to network-side bandwidth ratio; Trident-II ASIC). Other configurations of the improved data center network are also possible, for instance, 32-TOR server pod or 48-TOR server pod but with higher radix switching silicon than the Trident-II ASIC.

(59) Details of a spine plane of FIG. 8a are shown in FIG. 8b. In FIG. 8b, spine plane 820 consists of 20 spine switches 82-0 to 82-19 each connecting to 188 server pods. The connections from all 20 spine switches are grouped into 188 interlinks, with each interlink including a connection from each spine switch 82-0 to 82-19, for a total of 20 connections per interlink.

(60) Details of a server pod of FIG. 8a are shown in FIG. 8c. In FIG. 8c, the network-side connection (as opposed to the server-side connection) of the server pod is separated into intra-pod links and inter-pod links (i.e., the interlinks). The two types of links are made independent from each other. The intra-pod region 832 consists of the intra-pod links, the 20 TOR switches 84-0 to 84-19 and the 20 fabric switches 83-0 to 83-19 interconnected by the intra-pod links in a fat tree topology. For example, connection k in each of the 20 TOR switches is connected to fabric switch k; connection p of each of the 20 fabric switches is connected to TOR switch p. 20 fabric switches each provide an optional uplink (e.g., uplink 831) to connect to an external network. The inter-pod region consists of the inter-pod links (i.e., the interlinks) and 20 TOR switches 84-0 to 84-19 on the interlink side. Each interlink provides 20 10G connections to connect to all 20 spine switches on the same spine plane. Each server pod includes a total of 20 links. For example, interlink k of each of the 188 TOR switches across the 188 server pods are connected to spine plane k; interlink p of each of the 20 spine planes are connected to server pod p. Each TOR switch provides 48×10G connections in 12×QSFP ports as downlink to connect to servers.

(61) The data traffic through the fabric switches is primarily limited to intra-pod. The TOR switches now route both the intra-pod traffic as well as inter-pod traffic and are more complex. The independent link types achieve massive scalability in data center network implementations. (Additional independent links provided from higher radix switching ASIC may be created to achieve larger scale of connectivity objectives). Additionally, data center network 800 incorporates the full mesh topology concept (without physically incorporating an ICAS module) to remove redundant network devices and allow the use of innovative switching methods, in order to achieve a “lean and mean” data center fabric with improved data traffic characteristics.

(62) As shown in FIG. 8c, FIG. 8b and FIG. 8a, data center network 800 includes 20×188 TOR switches and 20×188 fabric switches equally distributed over 188 server pods, and 20×20 spine switches equally distributed over 20 spine planes. In FIG. 8a, each TOR switch has 100 10G-connections (i.e., 25 QSFPs of bandwidth in 10G mode), of which 60 10G-connections are provided server-side and 40 10G-connections are provided network-side. (Among the network-side connections 20 10G-connections are used for intra-pod traffic and 20 10G-connections are used for inter-pod traffic). In each server pod, fabric switches 83-0 to 83-19 each include 21 10G-connections, of which 20 10G-connections are allocated to connect with a 10G-connection in each of TOR switches 84-0 to 84-19, and the rest being converted to provide as uplink to connect to external network. In this manner, fabric switches 83-0 to 83-19 support the intra-pod region data traffic and the uplinks in the server pod by a 21-node full mesh topology (with the uplinks of fabrics switches 0-19 collectively seen as one node). Using a suitable routing algorithm, such as any of those described above in conjunction with Single-Source-Multiple-Destination Traffic Aggregation and Port-to-Port Traffic Aggregation, network congestion can be eliminated from all fabric switches.

(63) As the network in the intra-pod region of each server pod can operate in the same connectivity characteristics as a full mesh topology network, all the 20 fabric switches of the server pod may be replaced by an ICAS module. ICAS-based data center network 900, resulting from substituting fabric switches 83-0 to 83-19 of data center network 800, is shown in FIG. 9a. To distinguish from the server pod of data center network 800, a server pod with its fabric switches replaced by an ICAS module is referred to as an “ICAS pod.”

(64) FIG. 9a shows the architecture of an ICAS-based data center network, organized by three layers of devices—i.e., “top-of-rack” (TOR) switches, ICAS module implemented in 188 server pods 91-0 to 91-187 and spine switches implemented in 20 spine planes for 90-0 to 90-19—interconnected by interlinks in a fat tree topology. 20 spine planes provide optional uplinks 901 and 188 ICAS pods provide optional 188×20×10G uplinks 902 for connecting to an external network. The number of network devices in the data center network should be interpreted as illustrative only.

(65) Details of a spine plane of FIG. 9a are shown in FIG. 9b according to one embodiment. In FIG. 9b, spine plane 920 includes 20 spine switches 92-0 to 92-19 and a fanout cable transpose rack 921. The fanout cable transpose rack contains: k first port groups 923 are connected to corresponding port groups of k spine switches through a plurality of first MPO-MPO fiber cables, where each first port group including ┌p/m┐ first MPO adapters, and each first MPO adapter including m interfaces (where each interface includes one transmit fiber channel and one receive fiber channel), and a plurality of first MPO fiber adapters from the k port groups 923 are connect to LC optical fiber adapter mounting panel 922 through a plurality of first MPO-LC fanout fiber cables , where k=20, p=188, m=4, and ┌┐ is a ceiling function; the fanout cable transpose rack 921 includes p second port groups 924 that are connected to a plurality of second MPO-MPO fiber cables to form interlinks 99-0 to 99-187, each second port group contains ┌k/m┐ second MPO fiber adapters, each of which includes m interfaces (where each interface includes one transmit fiber channel and one receive fiber channel), and a plurality of second MPO fiber adapters from the p port groups 924 are connected to LC optical fiber adapter mounting panel 922 through a plurality of the second MPO-LC fanout cables; a plurality of first MPO-LC fanout fiber cables cross-connect a plurality of second MPO-LC fanout fiber cables on the LC fiber adapter mounting panel 922, through cross-connection, all connections from k spine switches 92-0 to 92-19 are reorganized into p interlinks 99-0 to 99-187, each interlink includes one connection from each of the spine switches 92-0 to 92-19, each interlink contains k connections in total.

(66) That is, on one side of the fanout cable transpose rack 921 is k first port groups 923, each first port group has ┌p/m┐ of first MPO adapters, where ┌┐ is a ceiling function, each port groups connects to a corresponding port group of a spine switch through the ┌p/m┐ first MPO-MPO cables. On the other side of the fanout cable transpose rack 921 is p second port groups 924, each second port group has ┌k/m┐ of second MPO adapters, where ┌┐ is an ceiling function, each port group connects to 5 second MPO-MPO cables to form an interlink to the ICAS pod.

(67) As pointed out earlier in this detailed description, the state-of-the-art data centers and switch silicon are designed with 4 interfaces (TX, RX) at 10 Gb/s or 25 Gb/s each per port in mind. Switching devices are interconnected at the connection level in ICAS-based data center. In such a configuration, a QSFP cable coming out from a QSFP transceiver is separated into 4 interfaces, and 4 interfaces from different QSFP transceivers are combined in a QSFP cable for connecting to another QSFP transceiver. Also, a spine plane may interconnect a large and varying number of ICAS pods (e.g., in the hundreds) because of the scalability of an ICAS-based data center network. Such a cabling scheme is more suitable to be organized in a fanout cable transpose rack (e.g., fanout cable transpose rack 921), which may be one or multiple racks and be integrated into the spine planes. Specifically, the spine switches and the TOR switches may each connect to the fanout cable transpose rack with QSFP straight cables. Such an arrangement simplifies the cabling in a data center. FIG. 9b illustrates such an arrangement for data center network 900 of FIG. 9a.

(68) In the embodiment shown in FIG. 9b, the first and the second optical fiber adapters are MPO adapters, the first and the second cables are MPO-MPO cables, the first and the second fanout cables are MPO-LC fanout cables, the mounting panel is LC optical fiber adapter mounting panel. One skilled in the art would understand that different types of optical fiber adapters/cable/optical fiber adapter mounting panel may also be used, such as FC, SC, LC, and MU.

(69) Details of an ICAS pod of FIG. 9a are shown in FIG. 9c. In FIG. 9c, the network side interface (as opposed to the server-side interface) of an ICAS pod is divided into intra-pod links (i.e. intralinks) and inter-pod links (i.e., interlinks) and the two types of links are made independent from each other. The intra-pod region consists of intralinks between the 20 TOR switches 93-0 to 93-10 19 and ICAS module 931, interconnected by 10G connections in a full mesh topology. Each ICAS module may provide 20 10G uplinks 932 to connect to one or more external networks. The inter-pod region consists of interlinks. ICAS pod may comprise 20 TOR switches 93-0 to 93-19 each connects one of 20 spine planes by interlinks, respectively; each interline comprising 20 connections each connects one of 20 spine switches in a spine plane, respectively. For example, interlink k of each of 188 TOR switches across the 188 ICAS pods is connected to spine plane k; interlink p of each of the 20 spine planes is connected to server pod p. Each TOR switch provides 60×10G interfaces in 15×QSFP ports as a downlink for connecting to servers.

(70) The data traffic through the ICAS module is primarily limited to intra-pod. The TOR switches now perform routing for the intra-pod traffic as well as inter-pod traffic and are more complex. The independent link types achieve massive scalability in data center network implementations. (Additional independent link provided from higher radix switching ASIC may be created to achieve a larger scale of connectivity objectives).

(71) As shown in FIG. 9c, FIG. 9b and FIG. 9a, each TOR switch allocates 20×10G-interfaces (5×QSFPs in 10G mode) to connect to its associated ICAS module (e.g., ICAS module 931) to support intra-pod traffic, and 5 QSFPs in 10G mode (20 10G-interfaces) to connect to the fiber transpose rack to support inter-pod traffic. As shown in FIG. 9c, each ICAS pod includes 20×5 QSFP transceivers for intra-pod traffic, connected by 100 QSFP straight cables, and 20×15 QSFP (10G mode) transceivers for server traffic, for a total 500 QSFP transceivers. The 20 TOR switches in an ICAS pod may be implemented by 20 Trident II ASICs. Although 20 TOR switches are shown in each ICAS pod in FIG. 9c, the ICAS module is scalable to connect up to 48 TOR switches in an ICAS pod (based on 32×QSFP Trident-II+switch ASIC).

(72) Together, the ICAS pods and the spine planes form a modular network topology capable of accommodating hundreds of thousands of 10G-connected servers, scaling to multi-petabit bisection bandwidth, and covering a data center with congestion improved and non-oversubscribed rack-to-rack performance.

(73) According to one embodiment of the present invention, a spine switch can be implemented using a high-radix (e.g., 240×10G) single chip switching device, as shown in FIG. 9d. Single-chip implementation saves the cost of extra transceivers, cables, rack space, latency and power consumption than multi-unit (rack unit) chassis-based switching device and stackable switching device implementations. The disadvantage of the single-chip spine switch approach is its network scalability, which limits the system to 240 ICAS pods at this time. As mentioned above, the semiconductor implementation limits the scale of a high-radix switching integrated circuit.

(74) To overcome the limitation on the port count of the silicon chip, one or more 1U to multi-U rackmount chassis each packaged with one or more ICAS modules, and a plurality of 1U rackmount chassis each packaged with one or more switching devices, can be stacked up in one or more racks, interconnected, to form a higher-radix (i.e. high network port count) stackable spine switching device (e.g., ICAS-based stackable switching device). Each ICAS module is connected to the plurality of switching devices, such that the ICAS module interconnects at least some interfaces of at least some port groups of different switching devices to form a full mesh non-blocking interconnection. The interfaces of the rest of the at least some port groups for interconnecting different switching devices are configured as an uplink. When the ICAS-module-based 1U to multi-U rackmount chassis are optically implemented (based on optical fiber and 3D MEMS), MPO-MPO cables may be used to connect the ICAS-based interconnection devices and the switching devices. When the ICAS-module-based 1U to multi-U rackmount chassis are electrically implemented as circuits (based on PCB+chip), DAC direct connection cables or AOC active optical cables may be used to connect the ICAS-based interconnection devices and the switching devices.

(75) Details of an ICAS-based stackable switching device 950 are shown in FIG. 9e. FIG. 9e shows ICAS modules 95-0 to 95-3 each connected in a full mesh topology to switches 96-0 to 96-3. In one embodiment, 4 Trident-II ASIC-based switches 96-0 to 96-3, each having a switching bandwidth of 24 QSFPs in 10G mode provided in 1:1 subscription ratio, and an ICAS box 953 integrating 4 ICAS modules 95-0 to 95-3 in one 1U chassis and each ICAS module containing 3 duplicate copies of ICAS1X5 sub-modules and each sub-module providing 4×10G of uplink 951 may be used to builds a stackable spine switch, as shown in FIG. 9e. The 4 switches 96-0 to 96-3 provide ports 952 of 1.92 Tbps of bandwidth to connect to servers. The ICAS-based stackable switching device 950 provides total uplink bandwidth of 480 Gb/s (4×3×40 Gb/s) to connect to external network, facilitates non-blocking 1:1 subscription ratio and provides full mesh non-blocking interconnect with a total of 1.92 Tbps of switching bandwidth.

(76) ICAS-based stackable switching device has the benefits of improved network congestion, saving the costs, power consumption and space savings than the switching devices implemented in the state of the art data center. As shown in the “ ICAS+Stackable Chassis” column of Table 5, data center with ICAS and ICAS-based stackable switching device performs remarkably on data center network with total switching ASIC saving by 53.5%, total power consumption saving by 26.0%, total space saving by 25.6% and much improved network congestion. However total QSFP transceiver usage is increased by 2.3%.

(77) The above stackable switching device is for illustrative purpose. A person experienced in the art can easily expand the scalability of the stackable switching device and should not be limited as in the illustration.

(78) The stackable switching device addresses the insufficiency in the number of ports of network switching chip, thus making possible a flexible network configuration. However, a considerable number of connecting cables and conversion modules have to be used to interconnect the ICAS-based interconnection devices and the switching devices. To further reduce the use of cables and conversion modules, ICAS modules and switch chips can be electronically interconnected using a PCB and connectors, which is exactly how the multi-unit switching device is structured. Specifically, the ICAS module of the ICAS-based multi-unit switching device is electrically implemented as circuits, and the port groups of the ICAS module are soldered or crimped onto a PCB using connectors that support high-speed differential signals and impedance matching. The interconnection between the internal port groups is realized using a copper differential pair on the PCB. Since signal losses vary significantly between different grades of high-speed differential connectors and between copper differential pairs on different grades of PCBs, an active chip can be added at the back end of the connector to restore and enhance the signal to increase the signal transmission distance on the PCB. The ICAS module of the ICAS-based multi-unit switching device may be implemented on a PCB called a fabric card, or on a PCB called a backplane. The copper differential pair on the PCB interconnects the high-speed differential connector on the PCB to form a full mesh connectivity in the ICAS architecture. The switch chips and related circuits are soldered onto a PCB called a line card, which is equipped with a high-speed differential connector docking to the adapter on the fabric card. A multi-U chassis of the ICAS-based multi-unit switching device includes a plurality of ICAS fabric cards, a plurality of line cards, and one or two MCU- or CPU-based control cards, one or more power modules and cooling fan modules. “Rack unit” (“RU” or “U” for short) measures the height of a data center chassis, equal to 1.75 inches. A complete rack is 48U (48 rack units) in height.

(79) One embodiment of the present invention also provides a chassis-based multi-unit (rack unit) switching device. A multi-unit chassis switching device groups multiple switch ICs onto multiple line cards. Chassis-based multi-unit switching equipment interconnects with line cards, control cards, and CPU cards via PCB-based network cards or backplanes, which saves the cost of transceivers, fiber optic cable and rack space required for interconnection.

(80) Details of an ICAS-based multi-unit chassis switching device 970 are shown in FIG. 9f. FIG. 9f shows 4 ICAS-based fabric cards 97-0 to 97-3 interconnected in a full mesh topology to switching ASIC's 98-0 to 98- 3. Switching ASIC 98-0 and 98-1 are housed in line card 973, and switching ASIC's 98-2 and 98-3 are housed in line card 974. Line cards 973 and 974 are connected through high speed PCB (printed circuit board) connectors to fabric cards 97-0 to 97-3. In one embodiment, 4 Trident-II ASIC-based switches 98-0 to 98-3, each having a switching bandwidth of 24 QSFPs in 10G mode provided in 1:1 subscription ratio, and 4 ICAS-based fabric cards 97-0 to 97-3 containing 3 duplicate copies of ICAS1X5 sub-modules and each sub-module providing 4×10G of uplink 971 may be used to builds a multi-unit chassis switch, as shown in FIG. 9f Two line cards provide data ports 972 of total 1.92 Tbps of bandwidth to connect to servers. ICAS-based multi-unit chassis switching device 970 provides total uplink bandwidth of 480 Gb/s (4×3×40 Gb/s) to connect to external network, facilitates full mesh non-blocking 1:1 subscription ratio interconnect with a total of 1.92 Tbps of switching bandwidth.

(81) Multi-unit chassis-based switching device with fabric cards that are ICAS-based full mesh topology has the benefits of improved network congestion, saving the costs and power consumption than that of ASIC-based fabric cards implementation with fat tree topology. As shown in the “ICAS+Multi-unit Chassis” column of Table 5, data center with ICAS and ICAS-based multi-unit chassis-based switching device performs remarkably on data center network with total QSFP transceiver saving by 12.6%, total switching ASIC saving by 53.5%, total power consumption saving by 32.7%, total space saving by 29.95% and much improved network congestion.

(82) The above multi-unit chassis switching device is for illustrative purpose. A person experienced in the art can easily expand the scalability of the multi-unit chassis switching device and should not be limited as in the illustration.

(83) The multi-unit chassis-based switching device has the disadvantage of a much longer development time and a higher cost to manufacture due to its system complexity, and is also limited overall by the form factor of the multi-unit chassis. The multi-unit chassis-based switching device, though provides a much larger port count than the single-chip switching device. Although the stackable switching device requires additional transceivers and cables than that of the multi-unit chassis-based approach, the stackable switching device approach has the advantage of greater manageability in the internal network interconnection, virtually unlimited scalability, and requires significantly less time for assembling a much larger switching device.

(84) The material required for (i) the data center networks of FIG. 2a, using state of the art multi-unit switching device (“Fat tree+Multi-unit Chassis”), (ii) an implementation of data center network 900 of FIG. 9a, using ICAS-based multi-unit switching device “ICAS+Multi-unit Chassis”, and (iii) an implementation of data center network 900 of FIG. 9a, using ICAS-based stackable switching device “ICAS+Stackable Chassis” are summarized and compared in Table 5.

(85) TABLE-US-00006 TABLE 5 Fat tree + ICAS + ICAS + Multi-unit Multi-unit Stackable Chassis Chassis Chassis Intralink (within Pod) N/A 5 5 Interlink (Across Pod) 4 5 5 Downlink (to Server) 12 15 15 Total 16 25 25 D:U ratio 3 3 3 D:I ratio N/A 3 3 Number of 10 G Interface (for comparison) 96 184.3 184.3 QSFP XCVR Module (Watt) 4 4 4 TOR Switch (Watt) 150 200 200 Multi-unit Chassis (Watt) 1660 0 0 Spine-side Interlink QSFP XCVR 18432 18800 38000 TOR-side Interlink QSFP XCVR 18432 18800 18800 Fabric/TOR-side Intralink QSFP XCVR 36864 18800 18800 Server-side QSFP XCVR 55296 56400 56400 Total QSFP XCVR 129024 112800 (12.6%) 132000 (−2.3%) ASIC in Spine Switch 2304 1600 1600 ASIC in Fabric Switch 4608 0 0 ASIC in TOR Switch 4608 3760 3760 Total Switching ASIC 11520 5360 (53.5%) 5360 (53.5%) Spine Switch (KW) 392.448 327.2 472.0 Fabric Switch (KW) 784.896 0 0 TOR Switch (KW) 986.112 1128.0 1128.0 Total Power Consumption (KW) 2163.456 1455.2 (32.7%) 1600 (26.0%) 96 × QSFP Spine Switch (8U) 1536 0 0 96 × QSFP Fabric Switch (8U) 3072 0 0 48 × QSFP Spine Switch (4U) 0 1600 1600 TOR Switch (1U) 4608 3760 3760 ICAS1X5TRIPLE (1U) 0 0 400 ICAS5X21 (2U) 0 376 376 Transpose Rack (36U) 0 720 720 ICAS2X9 (1U) 0 0 0 ICAS8X33 (4U) 0 0 0 ICAS10X41 (6U) 0 0 0 ICAS16X65 (16U) 0 0 0 Total Rack Unit (U) 9216 6456 (29.95%) 6856 (25.6%) Pod Interlink Bandwidth (Tbps) 7.7 4.0 4.0 Pod Intralink Bandwidth (Tbps) 7.7 4.0 4.0 Total Data Link Bandwidth (Pbps) 2.2 2.2 2.2 Per Plane Uplink Bandwidth (Tbps) 7.7/plane 0 0 Total Spine Uplink Bandwidth (Tbps) 0 150.4 601.6 Total ICAS Uplink Bandwidth (Tbps) 0 37.6 37.6 Spine-side Interlink QSFP Cable 18432 18800 18800 QSFP Fanout Cable (Transpose Rack) 0 37600 37600 QSFP Fanout Cable (ICAS5X21) 0 19740 19740 TOR-side Interlink QSFP Cable 0 18800 18800 TOR-side Intralink QSFP Cable 18432 18800 18800 Spine Switch QSFP Cable 0 0 19200 QSFP Fanout Cable (ICAS1X5TRIPLE) 0 0 19200 Total QSFP Cable 36864 56400 75600 Total QSFP Fanout Cable 0 57340 76540

(86) As shown in Table 5, the ICAS-based systems require significantly less power dissipation, ASICs and space, resulting in reduced material costs and energy.

(87) The above detailed description is provided to illustrate specific embodiments of the present invention and is not intended to be limiting. Numerous modifications and variations within the scope of the present invention are possible. The present invention is set forth in the accompanying claims.