Network topology with asymmetric fabrics

This invention relates generally to digital networks for interconnecting multiple users, and more particularly the invention relates to a multi-fabric topology for interconnecting nuilti-port user nodes.

Fault tolerant computer systems typically run in a multiprocessor environment in which computers can operate in parallel with one or more levels of redundancy. Such a system is described in U.S. Pat. No. 5,751,932 for “Fail-Fast, Fail-Functional, Fault-Tolerant Multiprocessor System”, assigned to Tandem Computers Incorporated, now Compaq Computer Corporation. Central Processing Units (CPUs) in this system operate in pairs with each user node having an X CPU with an X port and a Y CPU with a Y port. The X ports of all X processors are interconnected by an X fabric comprising a first topology of multi-port switches, and the Y ports of all Y processors are interconnected by a Y fabric comprising a second topology of multi-port switches. The patent describes the use of “TNet” Links comprising two uni-directional 10-bit sub-link busses connecting the X port and the Y port to the multi-port switches.

Tandem has introduced also a ServerNet-II cluster computing system consisting of network interface cards (NICs), 12-port crossbar routers (switches), and interconnecting links.

The switches and a switch topology for cluster computing are described in “ServerNet-II: a reliable interconnect for scalable high performance cluster computing”, Heirich, Garcia, Knowles, and Horst, Compaq Computer Corporation, Tandem Division, Sep. 21, 1998. As there described, the Server Net System Area Network (SAN) is a scalable interconnect technology designed as the primary interconnect for high availability information processing systems. These systems are characterized by round the clock availability in high profile locations where they support online transaction processing, telecommunications, internet service providers, and other applications. The ServerNet-II (SAN) achieves its high level availability by incorporating fault tolerant mechanisms at every architectural level. Failures in routing nodes that could impact the interconnect fabric are detected and isolated through self-checking logic. Failures in network links or routing elements are retried or re-routed along an alternate path through the fabric.

Heretofore, the multiple fabrics (X, Y) interconnecting multiple processors have generally had the same topologies with an equal number of switches interconnected together and with users in identical networks. The performance of each fabric or interconnect network can be defined in terms of inter-node distances, number of switch components, and data transfer capacity or bisection bandwidth. The bisection width is equal to the number of links in the weakest fabric bisection. While number of router hops does not significantly alter message latency in wormhole-routed networks, it does lower link occupancy and average link contention. The present invention is directed to enhancing network performance by employing asymmetric fabric topologies in a multi-fabric environment.

The invention is directed to a method of structuring a switch network having at least two groups of multi-port switches and the resulting network for interconnecting multi-port user nodes to enhance network performance such as increased bisection bandwidth, reduced inter-node distances, or reduced number of switches.

A first switch group is provided having a first plurality of multi-port switches interconnected with a first plurality of user ports whereby the first plurality of user ports are interconnected through one or more of the first plurality of switches. A second switch group is then provided which has a second plurality of multi-port switches interconnected with a second plurality of user ports whereby the second plurality of user ports are interconnected through one or more of the second plurality of switches. The first switch group and the second switch group are asymmetrical with at least one path between two nodes being of different length in the two groups of switches. By careful introduction of asymmetry between the switch groups, network performance is enhanced.

In the environment with two port (X,Y) user nodes, two switch groups are provided with the X ports of all user nodes being connected to the X switch group and the Y ports of all user nodes being connected to the Y switch group, each switch group comprising a plurality of multi-port crossbar switches. All of the plurality of X ports are interconnected through the X switch group and all the plurality of Y ports are interconnected through the Y switch groups. In structuring the asymmetry between switch groups, at least one distance -n (i.e., n-hop), a path between nodes) in the X switch group is not a distance -n in the Y switch group. Alternatively, or in addition thereto, the X switch group can have a different number of switches than has the Y switch group. In the embodiments where the switch groups are not interconnected, each switch group constitutes an independent fabric.

The invention and objects and features thereof will be more readily apparent from the following detailed description and appended claims when taken with the drawing.

FIG. 1A illustrates a conventional switch network having 16 user nodes (X, Y) interconnected by two identical fabrics (X, Y) each including two 12-port switches (routers), and FIG. 1B illustrates a similar switch network but with asymmetric topologies of the node connections in accordance with one embodiment of the invention.

FIG. 2 A and FIG. 2B illustrate a 24-node topology using 12-port switches in two (X,Y) asymmetric fabrics.

FIGS. 3A-3C illustrate a switch network for a 72-node 48-switch topology using 6-port switches in two (X,Y) asymmetric fabrics in accordance with another embodiment of the invention.

FIG. 4A illustrates another 72-node asymmetric topology using 18 12-port switches in asymmetric X and Y fabrics in accordance with the invention, and FIGS. 4B, 4 C are equivalent schematics of the X fabric and Y fabric, respectively.

As used herein a fabric is an independent switch group or router group, and a link is a connection between any two elements (end node or switch) and can be a bus or a pair of unidirectional connections.

Referring now to the drawing, FIG. 1A is a functional block diagram of a conventional switch network for 16 user nodes ( 1 - 16 ) with each node having two ports (X, Y) with the X ports connected through two 12-port switches 20 X, 22 X and the Y ports interconnected through two 12-port switches 20 Y, 22 Y. As illustrated, the two (X, Y) fabrics are identical in the connection of the two 12-port switches with each other and with the 16 nodes connected to the switches. Since each of the 12-port switches has 8 users connected thereto, 4 ports are provided for interconnection between switches. In this embodiment 56 pairs of nodes can be interconnected at a distance 1 (i.e., through only one switch) while 64 node pairs are interconnected at a distance two (i.e., through two switches). For example, node 1 (X or Y) is interconnectable with nodes 2 , 3 , 4 , 5 , 6 , 7 , or 8 through either switch 20 X or switch 20 Y. Similarly, node 9 (X, Y) can be connected to any one of nodes 10 , 11 , 12 , 13 , 14 , 15 , and 16 through a single switch 22 X, 22 Y. Node 1 (X, Y) is connectable to any one of nodes 9 through 16 through two switches ( 20 X, 22 X or 20 Y, 22 Y). A bisection of the network along a weakest dimension as illustrated by line AA shows a bisection width of 8 links, 4 links connecting the two Y switches and 4 links connecting the two X switches.

Consider now a reconnection of the 4 switches in an asymmetric topology as illustrated in FIG. 1 B. In this embodiment the X nodes 1 - 8 are again connected to switch 20 X and the X nodes 9 - 16 are again connected to switch 22 X. However, Y nodes 1 - 4 along with Y nodes 13 - 16 are connected to switch 20 Y, and Y nodes 5 - 12 are connected to switch 22 Y. A bisection of this network along line BB shows a bisection width of 16 links, thus doubling the transmission capacity of the network vis-a-vis FIG. 1 A. Further, 88 pairs of nodes are now interconnected through a single switch or distance 1 while 32 pairs of nodes are connected through two switches at a distance 2 . Accordingly, the asymmetric topologies of the X fabric and the Y fabric produce a doubling of bisection width and more than fifty percent increase in the number of pairs of nodes which are connected at a distance 1 .

FIGS. 2A and 2B illustrate X and Y fabrics each comprising three 12-port switches in a 24-node topology, in accordance with the prior art and with the invention respectively. The X nodes N 1 -N 8 are connected to switch R 1 X, X nodes N-N 16 are connected to switch R 2 X, and X nodes N 7 -N 24 are connected to switch R 3 X, each pair of X switches is interconnected by four links. Similarly, each pair of Y switches R 1 Y, R 2 Y and R 3 Y are interconnected by four links. However, Y nodes N 1 , N 4 , N 7 —N 22 in FIG. 2B are connected to switch R 1 Y, Y nodes N 2 -N 5 , N 8 . . . N 23 are connected to Y switch R 2 Y, and Y nodes N 3 , N 6 , N 9 . . . N 24 are connected to Y switch R 3 Y.

If the X fabric were identical to the Y fabric as in FIG. 2A prior art, then the shortest path between certain pairs of nodes (e.g. N 1 and N 9 ) would be two hops (distance 2 ). Only 84 pairs of nodes would be distance 1 and 192 pairs of nodes would be distance 2 in connections. Graph bisection is 16 links.

However, the asymmetric fabrics of FIG. 2B has 147 pairs of nodes with distance 1 connections and only 129 pairs of nodes are distance 2 connected. Further, a graph bisection width is 24 links. Therefore, the use of asymmetric fabrics reduces the average inter-node distance and increases the number of links in the bisection.

FIGS. 3A-3C illustrate a 72-node 48-switch topology in accordance with another embodiment of the invention. 72 nodes 30 as shown in FIGS. 3A, 3 B, and 3 C are interconnected by an X fabric including 24 switches 32 and by a Y fabric having 24 crossbar switches 34 . As shown in FIG. 3C, each node has an X fabric port 30 X and a Y fabric port 30 Y which are connected through an internal PCI bus 40 to two 40 MHz CPUs 42 . In this embodiment each switch or router is a 6-port ServerNet-I crossbar switch available from Tandem Computers Division of Compaq Computer Corporation. Each unit 36 includes two groups of nodes 30 as shown in FIG. 3 B.

Each group of nodes connects to the X fabric, and to the Y fabric, through two different routers. For example, each user group in unit 36 is connected to router 32 and to router 34 . Thus failure of one router in a fabric does not disrupt connection between any two nodes.

The weakest bisection C—C of the asymmetric fabric topology has 24 links (3 for each bisected switch “ 32 or 34 ”, plus each bisected line). The X and Y fabrics, when considered independently, have only 10 links in each of their bisections. Thus, if Y fabric were constructed identical to the X fabric, the net width of the weakest bisection would have been just 20 links. While each X, Y fabric has the same number of switches or routers, the interconnection of the various nodes to the switches differs in the X fabric as compared to the Y fabric, thus altering the shortest distance between nodes. Numbering the units in FIG. 3A from 1 to 9 (left to right), it can be seen that there is exactly one-hop path (either in the X fabric or in the Y fabric but not in both) between a group of nodes in one unit and an group of nodes in a different unit. for groups of nodes within the same unit, there are two one-hop paths in the X fabric and two in the Y fabric. The following matrix shows the routers (and the fabrics) providing one-hop paths between groups of nodes in various units.

FIGS. 4A, 4 B, and 4 C illustrate a 72-node asymmetric topology (FIG. 4A) using 12-port routers or switches and equivalent circuits for the X fabric (FIG. 4B) and the Y fabric (FIG. 4 C), respectively. Every group of 8 nodes is connected, via a maximum of 2 hops, to every other group of 8 nodes in either the X fabric or the Y fabric. There are 9 groups of 8 nodes each, or a total of 72 nodes.

The topology provides a way to connect 72 nodes using nine 12-port switches so that the maximum distance between any pair of nodes is only two router (switch) hops when using the preferred fabric. The matrix below shows the preferred fabric for every pair. With asymmetric X and Y fabrics as shown the weakest bisection has 28 links using 18 routers. If identical fabrics were used, the weakest bisection would have 24 links and the maximum distance between any pair of nodes would be 3 hops. Accordingly, the asymmetric topology again reduces distance and increases bisection width.

In the described embodiment, enhanced network performance has been realized through the addition of a second fabric differing in connectivity by at least one link from the first fabric, such that either the maximum distance between at least one pair of nodes is reduced or the net bisection width of the interconnection network is more than doubled after the second fabric is added. Similar improvements in network performance can be realized effectively by reducing the number of switches between switch groups or fabrics.

While the invention has been described with reference to specific embodiments, the description is illustrative of the invention and is not to be construed as limiting the invention. For example, the invention encompasses switch networks having more than 2 fabrics or switch groups. Also, the invention is described in the context of System Area Networks, but the invention applies equally well to other networks such as Fibre channel, ATM, and switched Ethernet. Thus, various modifications and applications may occur to those skilled in the art without departing from the true spirit and scope of the invention as defined by the appended claims.