Deadlock-free local rerouting for handling multiple local link failures in hierarchical network topologies

Abstract

A method for communication includes partitioning local links in a subnetwork of a packet data network into at least first and second groups. For each local link that connects a first upper-tier switch to a first lower-tier switch in the subnetwork, a corresponding detour route is defined, passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch, and from the second lower-tier switch over a second local link to a second upper-tier switch, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch. Upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, data packets arriving from the network at the first upper-tier switch are rerouted to pass via the corresponding detour route to the first lower-tier switch.

Description

FIELD OF THE INVENTION

The present invention relates generally to communication networks, and particularly to methods and systems for rerouting network traffic in cases of link failure.

BACKGROUND

Communication networks used in applications such as High-Performance Computing (HPC) systems and datacenters typically comprise a large number of network elements interconnected by links. A high-reliability communication network should be able to recover quickly from link failure events.

Methods for recovery from link failures are known in the art. For example, a white paper by Mellanox® Technologies (2018) describes a solution called SHIELD™ (Self-Healing Interconnect Enhancement for Intelligent Datacenters), which takes advantage of the intelligence already built into the latest generation of InfiniBand switches. By making the fabric capable of self-healing autonomy, the speed with which communications can be corrected in the face of a link failure can be sped up by 5000×, fast enough to save communications from expensive retransmissions or absolute failure.

Aspects of the SHIELD solution are described, for example, in U.S. Pat. No. 9,729,473, which describes a method in a network element that includes multiple interfaces for connecting to a communication network. The method includes receiving via an ingress interface packets that are not allowed to undergo re-routing and that are addressed to a destination via a first egress interface. The packets are forwarded via the first egress interface when there is a valid path from the first egress interface to the destination. When there is no valid path from the first egress interface to the destination, a second egress interface is selected from a group of multiple egress interfaces that have respective paths to the destination and are assigned to packets for which re-routing is allowed, and the packets are forwarded via the second egress interface until recovering a path to the destination.

U.S. Pat. No. 9,137,143 describes a multiprocessor computer system comprising a dragonfly processor interconnect network that comprises a plurality of processor nodes and a plurality of routers. The routers are operable to adaptively route data by selecting from among a plurality of network paths from a target node to a destination node in the dragonfly network based on one or more of network congestion information from neighboring routers and failed network link information from neighboring routers.

SUMMARY

Embodiments of the present invention that are described hereinbelow provide improved methods and systems for rerouting network traffic in cases of link failure.

There is therefore provided, in accordance with an embodiment of the invention, a method for communication, which includes configuring primary routes for delivery of data packets in a network including multiple subnetworks interconnected by trunk links. Each subnetwork includes multiple switches, which are configured to communicate via at least first and second virtual lanes and are arranged in a bipartite topology including an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches. The local links in each subnetwork are partitioned into at least first and second groups.

For each local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, a corresponding detour route is defined, passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch within the given subnetwork, and from the second lower-tier switch over a second local link to a second upper-tier switch in the given subnetwork, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch. Upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, the data packets arriving at the first upper-tier switch for transmission to the first lower-tier switch are rerouted so that the data packets pass through the corresponding detour route to the first lower-tier switch, while transmitting the data packets over the second and third local links via the second virtual lane.

In some embodiments, the lower-tier switches include leaf switches, and the method includes delivering the data packets received in the given subnetwork via the detour route to a network end-node that is connected to the first lower-tier switch.

Additionally or alternatively, partitioning the links in each subnetwork includes defining a third group of the links, wherein defining the corresponding detour route includes configuring detour routes through the first and second groups of the links for the data packets that are destined for network end-nodes that are connected to the given subnetwork, and wherein the method includes defining at least one dive-through route over which a data packet enters the subnetwork via the first virtual lane through a first trunk link to one of the upper-tier switches in the subnetwork, passes through one of the links in the third group to one of the lower-tier switches in the subnetwork and through another one of the links via the second virtual lane to another one of the upper-tier switches in the subnetwork, from which the data packet exits the subnetwork through a second trunk link.

In a disclosed embodiment, defining the corresponding detour route includes defining multiple detour routes for use in case of failures of multiple ones of the local links in the given subnetwork.

In some embodiments, defining the corresponding detour route includes, for any given local link, applying the corresponding detour route in rerouting the data packets arriving in the given subnetwork over any of a plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link. In a disclosed embodiment, configuring the primary routes includes defining an adaptive routing group including the primary routes that are destined to the given subnetwork irrespective of the lower-tier switches to which the primary routes are directed, such that upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over detour routes within the given subnetwork.

There is also provided, in accordance with an embodiment of the invention, a communication system, which includes a network including multiple subnetworks interconnected by trunk links. Each subnetwork includes multiple switches, which are configured to communicate via at least first and second virtual lanes and are arranged in a bipartite topology including an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches. A management processor is configured to define primary routes for delivery of data packets in the network, to partition the local links in each subnetwork into at least first and second groups, and for each local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, to define a corresponding detour route passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch within the given subnetwork, and from the second lower-tier switch over a second local link to a second upper-tier switch in the given subnetwork, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch. The switches in the given subnetwork are configured, upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, to reroute the data packets arriving at the first upper-tier switch for transmission to the first lower-tier switch so that the data packets pass through the corresponding detour route to the first lower-tier switch, while transmitting the data packets over the second and third local links via the second virtual lane.

There is additionally provided, in accordance with an embodiment of the invention, a method for communication, which includes partitioning local links in a subnetwork of a packet data network into at least first and second groups. For each local link that connects a first upper-tier switch to a first lower-tier switch in the subnetwork, a corresponding detour route is defined, passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch, and from the second lower-tier switch over a second local link to a second upper-tier switch, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch. Upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, data packets arriving from the network at the first upper-tier switch for forwarding over the local link are rerouted so that the data packets pass via the corresponding detour route to the first lower-tier switch.

In a disclosed embodiment, rerouting the data packets includes receiving the data packets from the network, and forwarding the data packets via a specified virtual lane from the second lower-tier switch over the second local link to the second upper-tier switch.

There is further provided, in accordance with an embodiment of the invention, method for communication, which includes configuring primary routes for delivery of data packets in a network including multiple subnetworks interconnected by trunk links. Each subnetwork includes multiple switches in a bipartite topology including an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches. For any given local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, a corresponding detour route is defined, passing through the local links within the given subnetwork from the first upper-tier switch to a second lower-tier switch, and from the second lower-tier switch to a second upper-tier switch and from the second upper-tier switch to the first lower-tier switch. When the given local link is a part of a plurality of the primary routes, the corresponding detour route is applied to all of the plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link. Upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, the data packets arriving at the first spine switch in the given subnetwork are rerouted to pass via the corresponding detour route to the first lower-tier switch.

In a disclosed embodiment, configuring the primary routes includes defining an adaptive routing group including the primary routes that are destined to the given subnetwork irrespective of the lower-tier switches to which the primary routes are directed, such that upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over detour routes within the given subnetwork.

The present invention will be more fully understood from the following detailed description of the embodiments thereof, taken together with the drawings in which:

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1A is a block diagram that schematically illustrates a multi-level subnetwork, in accordance with an embodiment of the invention;

FIG. 1B is a block diagram of a hierarchical packet communication network comprising interconnected multiple subnetworks, in accordance with an embodiment of the invention;

FIG. 2 is a block diagram that schematically illustrates establishment of a detour route following a link failure in a multi-level subnetwork, in accordance with an embodiment of the invention;

FIG. 3 is a block diagram that schematically illustrates establishment of a detour route following a link failure in a multi-level subnetwork, in accordance with another embodiment of the invention; and

FIG. 4 is a block diagram that schematically illustrates establishment of a “dive-through” route in a multi-level subnetwork, in accordance with an embodiment of the invention.

DETAILED DESCRIPTION OF EMBODIMENTS

Overview

When a link in a packet data network fails, packets that would normally be forwarded via this link may be dropped until an alternative path to the destination is established. In principle, a central entity in the network may detect or be notified of a failed link and may then reconfigure one or more of the network elements to create an alternative path to the destination. Centralized recovery methods of this sort, however, react too slowly for the needs of high-performance, high-speed networks.

U.S. patent application Ser. No. 17/016,464, filed Sep. 10, 2020, whose disclosure is incorporated herein by reference, describes methods and systems for recovering from link failures in a network using predefined local detour paths. This approach provides a fast recovery mechanism, and the detour path can remain in use until the link is fixed or until a central entity reconfigures an alternative path. The techniques described in this patent application are particularly applicable, for example, to topologies comprising interconnected groups of switches (or other network elements) having bipartite topologies, in which the switches are arranged in upper and lower tiers, which are connected by a mesh of local links. In such a topology, for example, the switches may include a lower tier of leaf switches, which connect to end-nodes of the network, and an upper tier of spine switches, which are connected by trunk links to the other groups.

In lossless networks, rerouting of packets in response to link failures can give rise to problems of deadlock due to flow control loops. Lossless networks typically apply flow control techniques to prevent buffer overfill and packet drop events. Rerouting in such networks, however, may result in buffers of multiple switches interconnected cyclically becoming full, preventing these switches from further sending or receiving packets. This sort of condition is referred to as a “deadlock condition” or simply “deadlock.” One way to avoid deadlock is to transmit packets over multiple virtual lanes (VLs) on each link, but this approach requires that the switches have large memories in order to allocate sufficient buffer space for all the virtual lanes. To optimize the use of memory resources, it is generally desirable that only a small number of virtual lanes be used throughout the network, for example two virtual lanes.

In response to this problem, the above-mentioned U.S. patent application Ser. No. 17/016,464 describes a deadlock-free solution that can be applied by the switches in a bipartite group in case of a local link failure: The packets are rerouted to their destination switch via a detour path, which includes a detour link that delivers packets in a spine-to-leaf direction and another detour link that delivers packets in a leaf-to-spine direction within the group. This solution is capable of handling any single link failure within any given group of switches without requiring that more than two virtual lanes be used throughout the network.

Embodiments of the present invention that are described herein extend this solution to cover the possibility of multiple local link failures within a bipartite group of switches. These embodiments are applicable particularly to networks comprising multiple subnetworks interconnected by trunk links, each subnetwork comprising multiple switches arranged in a bipartite topology. A management processor (for example, a computer processor performing the function of a subnet manager) configures primary routes, over which packets are delivered through the network from a given source to a given destination, and detour routes for use in case of failures of the primary routes.

To define the detour routes, the management processor partitions the local links in each subnetwork into at least two groups. For each local link that is a part of one or more of the primary routes and connects a given upper-tier switch to a given lower-tier switch in one of the subnetworks, the corresponding detour route passes from the given upper-tier switch to a second lower-tier switch within the subnetwork through a first local link belonging, for example, to the first of the two groups. The route then passes from this second lower-tier switch over a second local link (which may belong to either of the groups) to a second upper-tier switch in the subnetwork, and from the second upper-tier switch over a third local link belonging to the second group to the original, given lower-tier switch.

Upon a failure of a local link, between a certain upper-tier switch and a certain lower-tier switch, for which a detour route of this sort has been defined, the switches in the subnetwork reroute data packets arriving at the given upper-tier switch to pass through the corresponding detour route to the lower-tier switch. Assuming that the packets arrive at the upper tier switch via a first virtual lane (for example from a trunk link on VL0), over which packets are transmitted from their source nodes through the network, the switches will transmit the data packets over the second and third local links of the detour route via a different, second virtual lane (VL1 in this example). It can be shown that this scheme will operate without deadlocks, using only two virtual lanes, even when multiple local links in a given subnetwork fail simultaneously. Specifically, in a bipartite subnetwork having n switches in each of the upper and lower tiers, the number of local link failures that can be handled by the present detour routes is n/2−1.

In some embodiments of the present invention, advance definition of detour routes for the local links in each subnetwork is used in defining adaptive routing (AR) groups over the entire network. For this purpose, as explained above, the management processor defines a corresponding detour route for each local link (between a certain upper-tier switch and a certain lower-tier switch in a given subnetwork) that is a part of one or more primary routes in the network. When a given local link is a part of multiple primary routes, the corresponding detour route is applied to all of these primary routes. Thus, these primary routes continue to pass through the same upper-tier switch in the subnetwork notwithstanding failures of the given local link.

Using this scheme, the management processor can define an AR group comprising the primary routes that are destined to any given subnetwork irrespective of the lower-tier switches through which the primary routes pass. Upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over the corresponding detour routes within the given subnetwork. Thus, the switches outside any given subnetwork need not even be informed of local link failures in the subnetwork and continue to transmit packets along the primary routes irrespective of such failures, since local link failures are handled internally within the subnetworks. This approach is useful in conserving memory in the switches and reducing the need for inter-switch failure notifications.

System Description

Reference is now made to FIGS. 1A and 1B, which are block diagrams that schematically illustrate a communication system 20 in which a network recovers from link failures using deadlock-free detour paths, in accordance with an embodiment of the invention. System 20 comprises multiple subsystems 22, each comprising a respective subnetwork 26, one of which is shown in detail in FIG. 1A. Subnetworks 26 are labeled SN1 . . . SN5 in FIG. 1B and are interconnected by trunk links 52. Network end-nodes 24, such as host computers, in system 20 communicate with one another over local links 30 within each subnetwork 26 and trunk links 52 between the subnetworks. For the sake of simplicity, system 20 is shown as including only five subnetworks, with eight switches arranged in two tiers in each subnetwork. In practice, however, the methods described herein can be applied in much larger networks, with greater numbers of subnetworks and more switches in each subnetwork.

As shown in FIG. 1A, each subnetwork 26 is arranged in a bipartite topology, including lower-tier switches 28 and upper-tier switches 29, which are connected to the lower-tier switches by a mesh of local links 30. In the pictured example, lower-tier switches 28 are identified as leaf switches, which are denoted L0 . . . L3 and are connected to respective end-nodes 24; while upper-tier switches 29 are identified as spine switches and denoted S0 . . . S3. End-nodes 24 are connected only to the leaf switches, while the spine switches are connected only to the leaf switches and to trunk links 52. In the description that follows, switches 28 and 29 may be referred to simply as “leaves” and “spines,” respectively, for the sake of brevity. The principles of the present invention, however, are by no means limited to this specific sort of bipartite topology and may alternatively be applied to any bipartite pair of tiers in a hierarchical subnetwork that includes two or more such tiers.

In the example of FIG. 1A, switches 28 and 29 are interconnected in a fully connected bipartite topology, i.e., each upper-tier switch 29 connects to all lower-tier switches 28, and each lower-tier switch connects to all upper-tier switches. A bipartite topology of this sort can also be viewed as a two-level Fat-Tree (FT) topology or two adjacent levels in a multi-level FT topology.

The topology is defined as “bipartite” because it is analogous to a bipartite graph, meaning a graph in which the vertices are partitioned into two sets, and every edge in the graph has an endpoint in each of the two sets. In the present context, a bipartite topology comprises multiple upper-tier switches and multiple lower-tier switches. Each upper-tier switch is connected to one or more lower-tier switches, and each lower-tier switch is connected to one or more upper-tier switches, but upper-to-upper and lower-to-lower links are not allowed. A bipartite topology is referred to as “fully connected” when each upper-tier switch connects to all lower-tier switches, and vice versa. In general, however, a bipartite topology is not necessarily fully connected.

In the present embodiment and in the description that follows, subnetwork 26 comprise switches 28 and 29. As shown in the inset in FIG. 1A, each switch 28 comprises multiple ports 32, a packet processor 34, and a memory 36, for example for storing packets in queues while awaiting transmission. Switches 29 (not shown in the inset) are of similar construction. The switch connects via ports 32 to other switches and/or to network end-nodes 24. In FIG. 1A, four ports of each switch that are used for inter-tier connections are within subnetwork 26 are numbered 0 . . . 3, while four ports used for external connections are numbered 4 . . . 7. The term “switch” is used in the present description and in the claims to refer broadly to network elements having multiple ports and capabilities of switching data packets among the parts. Other network elements with these features, such as routers and bridges, are also considered to be “switches” in the present context.

Each port 32 serves both as an ingress port for receiving packets from another switch or from a network end-node, and as an egress port for transmitting packets to another switch or to a network end-node. Ports 32 typically comprise hardware circuit components capable of performing physical layer (PHY) and media access control (MAC) interface functions, as are known in the art.

Packet processor 34 applies various processing operations to packets received via ports 32, such as verification, filtering, prioritization, and forwarding. These processing operations may also include modifying the Virtual Lane (VL) with which the packets are associated. As noted earlier, switches 28, 29 support at least two virtual lanes (VL0 and VL1) and may support a larger number of virtual lanes. Typically, packet processor 34 comprise hard-wired and/or programmable hardware logic circuits, which are configured to carry out the functions described herein at high speed. Additionally or alternatively, at least some of the functions of packet processor 34 may be implemented in software on a suitable programmable processing core.

Packet processor 34 performs packet forwarding using a forwarding table 38, based, for example, on source and destination addresses carried in the packet headers. Using the forwarding table, the switch determines the egress port for each received packet. Switch 28 stores received packets that have been processed and await transmission in queues in memory 36. In alternative embodiments, other packet processing and forwarding methods, such as Access Control Lists (ACLs), can also be used.

Switch 28 comprises a switch controller 40, typically comprising a programmable processor, which is programmed in software to carry out the functions that are described herein. Among other tasks, switch controller 40 updates forwarding table 38 to include detour paths that can be used to resolve network malfunctions due to link failures. Typically these detour paths are defined by a management processor, such as a programmable processor 56 in a subnet manager (SM) 54 (FIG. 1B) in system 20. Additionally or alternatively, the detour paths may be defined by another entity in system 20, for example by interaction among switches 28, 29 in each subnetwork 26, in communication with SM 54 or independently.

SM 54 typically comprises processor 56 and a memory and may be configured as a standalone device, with communication links to switches 28 and 29, or may be integrated into one of the switches or other components of system 20. Although SM 54 is shown in FIG. 1B as a single physical entity, the functions of SM 54 may alternatively be distributed among multiple processing units in system 20. SM 54 carries out the routing and fault-handling functions that are described herein, along with other management functions that are known in the art, under the control of suitable software. This software may be provided to SM 54 in electronic form, for example over a network. Alternatively or additionally, the software may be stored on tangible, non-transitory computer-readable media, such as electronic, optical, or magnetic memory media.

As noted above, switch controller 40 stores one or more predefined detour paths, which are defined by a management processor, such as processor 56 in SM 54, using the methods that are described herein. Upon a link failure, switch controller 40 directs traffic to an alternative port leading to the respective detour path. In other embodiments, switch controller 40 communicates with other switches in network 26 for selecting a detour path, and applies the selected detour path by updating forwarding table 38. Additionally or alternatively, upon detecting a failure of one of local links 30, switch controller 40 notifies subnet manager 54, so that the subnet manager can take appropriate action in rerouting in the network to circumvent the failed link. The use of the detour route in the meanwhile ensures that there will be little or no interruption of packet traffic in the network while the processes of failure notification and rerouting are going on. Once the new routes have been established, the detour route may no longer be required.

In FIG. 1B, a network end-node 24s, referred to as a source node and labeled “SRC node,” sends packets over its local subnetwork 26 (SN1) to a destination end-node 24d, denoted “DST node,” in another subnetwork 26 (SN4). SM 54 configures a primary route for this packet flow, for example a three-hop route beginning with local link 30 from L2 to S2 within SN1, then via trunk link 52 to S2 in SN4, and via local link 30 to L2 in SN4, which delivers the data packets to the DST node.

By convention, switch L2 in SN1 transmits the packets from their source node on VL0. This choice is arbitrary, and any virtual lane can be used for transmission of packets from their source. To ensure deadlock-free operation while conserving virtual lanes (and thus conserving memory resources), however, it is desirable that the same virtual lane, such as VL0, be used for transmission of all data packets from their respective source nodes in the source subnetwork. Switchover to a second virtual lane, such as VL1, within subnetworks 26 takes place when appropriate, specifically when the packet is forwarded by an upper-tier switch 29 from one trunk link 52 to another or is forwarded within a subnetwork from a lower-tier switch to an upper-tier switch, as described further hereinbelow.

Alternatively or additionally, SM 54 may define a four-hop route, for example via one of upper-tier switches in SN5, or a six-hop “dive-through” route, passing through upper-tier switches 29 and one of lower-tier switches 28 in SN5. (Such routes may similarly be defined through other subnetworks, such as SN2 and SN3. Each “hop” corresponds to a single link on the path of the packet to its destination.) In the dive-through route, for example, data packets transmitted from SRC node 24s through switch L2 are conveyed through a first trunk link 52 to one of upper-tier switches 29 in SN5, for example from S1 in SN1 to S0 in SN5. The packets thus enter SN5 and pass through local links 30 from S0 to one of lower-tier switches 28, such as switch L1 in SN5, and then through another one of the local links 30 to another one of the upper-tier switches, for example to switch S3 is SN5. The data packets are transmitted from SN1 and enter SN5 on VL0, but for purposes of deadlock avoidance, switch L1 in SN5 transmits the packets onward using VL1. The data packets exit SN5 on VL1 through a second trunk link 52, for example from switch S3 is SN5 to switch S1 in SN4. This switch then transmits the packets over local link 30 to switch L2, which passes the packets to DST end-node 24d. Details of the criteria applied by SM 54 is defining dive-through routes of this sort are described further hereinbelow with reference to FIG. 4.

Defining and Applying Detour Routes

FIGS. 2 and 3 are block diagrams that schematically illustrates establishment of detour routes following a link failure in one of subnetworks 26, in accordance with embodiments of the invention. SM 54 defines a corresponding detour route of this sort for each local link 30 in any given subnetwork 26 that is a part of one or more of the primary routes in a network, such as the routes described above between source node 24s and destination node 24d in system 20. FIGS. 2 and 3 show two variants on the same principles of definition of these detour routes, with different choices of the middle link in the routes, as explained below.

To ensure that the detour routes will be deadlock-free, SM 54 partitions local links 30 into two groups. In FIGS. 2-4, local links in one of the groups (referred to arbitrarily as the first group) are represented by solid arrows, while those in the other (second) group are represented by dashed arrows. It can be useful to define the two groups to contain roughly equal numbers of the local links, with the links in each group connecting to most or all of switches 28 and 29. Alternatively, other groupings may be used.

Referring now to FIG. 2, a primary route from a source node 60 in another subnetwork passes through an upper-tier switch S1 in subnetwork 26 to a lower-tier switch L2, via a local link 66. (The primary route then typically passes from switch L2 to a destination end-node, as shown in FIG. 1B, for example.) For each such local link, SM 54 (FIG. 1B) defines a corresponding detour route using links from both of the groups of local links. In the example shown in FIG. 2, the detour route passes through a first local link 68 belonging to the first group, connecting the first upper-tier switch S1 to a different lower-tier switch L1 within subnetwork 26, and then from this lower-tier switch over a second local link 70 to a second upper-tier switch S3. In the present example, local link 70 belongs to the first group, as well. The detour route then passes from this second upper-tier switch S3 over a third local link 72 belonging to the second group to the original lower-tier switch L2.

In the scenario shown in FIG. 2, link 66 has failed. Therefore, when packets on a primary route that included link 66 enter subnetwork 26, switch S1 reroutes the packets over the detour route defined above, via links 68, 70 and 72. The packets may arrive at switch S1 on either VL0 or VL1. In either case, however, to ensure deadlock-free operation, switches L1 and S3 will transmit the packets over the second and third local links of the detour route (links 70 and 72) via VL1.

Typically, SM 54 defines multiple detour routes for use in case of failures of different local links in subnetwork 26. Two or more of these detour routes can be used concurrently when multiple link failures occur simultaneously in the subnetwork.

As another example, FIG. 3 shows a detour route for use in case of failure of a local link 74 from upper-tier switch S1 to lower-tier switch L4. The detour route includes a first local link 76 from switch S1 to switch L1 within subnetwork 26, and then a second local link 78 to a second upper-tier switch S2 and a third local link 80 to the original lower-tier switch L4. Local link 76 belongs to the first group of links, while local links 78 and 80 belong to the second group. As in FIG. 2, the packets are transmitted over links 78 and 80 using VL1.

Any given local link in a given subnetwork 26, such as links 66 and 74 in FIGS. 2 and 3, can typically be a part of multiple different primary routes through system 20 that terminate at destination nodes that are connected to this subnet. For example, local link 66 may be a part of multiple primary routes that connect various source nodes 24 in SN1, as well as in SN5 (and other subnetworks), to destination nodes that are connected to switch L2 in SN4. In such cases, switch S1 will apply the detour route that is shown in FIG. 2 in rerouting the data packets arriving in subnetwork 26 over any of the primary routes of which local link 66 is a part. Thus, the primary routes will continue to pass through the same upper-tier switch S1 notwithstanding failures of local link 66.

Rerouting of the packets within subnetwork 26 in this manner is transparent to the other subnetworks in system 20. Switches 28, 29 in this subnetwork will notify SM 54 of the failure, and SM 54 will typically respond by updating forwarding tables 38 (FIG. 1A) of the switches within the subnetwork. Until this update takes place, however, traffic on all the applicable primary routes is transmitted without interruption over the detour route.

This application of preconfigured detour routes is also useful in simplifying the adaptive routing information that is provided to the switches in system 20 and thus in reducing the volume of memory that the switches must use to store this adaptive routing information. For this purpose, SM 54 defines adaptive routing groups in system 20. Each such group comprises the primary routes that are destined to a given subnetwork, irrespective of the lower-tier switches (i.e., the leaf switches in the present example) within the subnetwork to which the primary routes are directed. Upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group will be rerouted over detour routes within the given subnetwork.

Fault routing information is saved for each adaptive routing group to prevent sending data packets toward failed links. Since local failures are handled locally inside each subnetwork, the fault routing information is correct for the entire subnetwork, thus enabling the adaptive routing group information to be stored per subnetwork, rather than per destination leaf switch. When adaptive routing groups of this sort are defined with respect to the primary routes that terminate in a given subnetwork, switches in other subnetworks need not be aware at all of the lower-tier switches within the given subnetwork that take part in detour routing. There is thus no need for adaptive routing information to be stored by switches in other subnetworks with respect to the leaf switches through which the primary routes reach their destinations in the given subnetwork. This adaptive routing solution is a useful adjunct to the particular detour routing schemes that are described above; but it may alternatively be applied in conjunction with other methods for defining detour paths within subnetworks, such as the methods defined in the above-mentioned U.S. patent application Ser. No. 17/016,464.

FIG. 4 is a block diagram that schematically illustrates a “dive-through” route in subnetwork 26, in accordance with another embodiment of the invention. For purposes of dive-through routing, SM 54 partitions the local links in the subnetwork to define a third group, in addition to the first and second groups that are used in defining detour paths for data packets that are destined for network end-nodes in subnetwork 26. In the present example, the first group of links includes a local link 94, indicated by a solid arrow, and the second group includes a local link 96, indicated by a dashed arrow. A local link 98 in the third group is indicated by a dash-dot arrow. SM 54 can choose the number of links to include in the third group, for example based on the volume of dive-through traffic that is expected to traverse subnetwork 26.

The dive-through route shown in FIG. 4 is used for forwarding of data packets transmitted from source node 60 over one of trunk links 52 (FIG. 1B) to an upper-tier switch S1 in subnetwork 26. The packets are assumed to arrive on VL0. The dive-through route passes through link 98 in the third group to a lower-tier switch L5 in subnetwork 26, and from switch L5 through another link 100 to another one of the upper-tier switches S6. To prevent deadlocks, switch L5 transmits the packet over link 100 on VL1, and switch S6 forwards the packet on this same virtual lane out of subnetwork 26 through another trunk link 102 to the destination subnetwork. Link 100 may belong to any of the groups of local links.

Although the embodiments shown in FIGS. 2-4 rely on partitioning of the links in a subnetwork for the purpose of defining deadlock-free detour routes, other partitioning schemes may alternatively to be used for this purpose, mutatis mutandis. For example, switches 28 and 29 in a given subnetwork may be partitioned into two or three groups to serve as the basis for defining detour routes between the groups. As another alternative, the ports of each switch may be partitioned between the groups.

Documents incorporated by reference in the present patent application, and specifically the above-mentioned U.S. patent application Ser. No. 17/016,464, are to be considered an integral part of the application except that to the extent any terms are defined in these incorporated documents in a manner that conflicts with the definitions made explicitly or implicitly in the present specification, only the definitions in the present specification should be considered.

It will be appreciated that the embodiments described above are cited by way of example, and that the present invention is not limited to what has been particularly shown and described hereinabove. Rather, the scope of the present invention includes both combinations and subcombinations of the various features described hereinabove, as well as variations and modifications thereof which would occur to persons skilled in the art upon reading the foregoing description and which are not disclosed in the prior art.

Claims

1. A method for communication, comprising: configuring primary routes for delivery of data packets in a network comprising multiple subnetworks interconnected by trunk links, each subnetwork comprising multiple switches, which are configured to communicate via at least first and second virtual lanes and are arranged in a bipartite topology comprising an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches;partitioning the local links in each subnetwork into at least first and second groups;for each local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, defining a corresponding detour route passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch within the given subnetwork, and from the second lower-tier switch over a second local link to a second upper-tier switch in the given subnetwork, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch; andupon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, rerouting the data packets arriving at the first upper-tier switch for transmission to the first lower-tier switch so that the data packets pass through the corresponding detour route to the first lower-tier switch, while transmitting the data packets over the second and third local links via the second virtual lane.

2. The method according to claim 1, wherein the lower-tier switches comprise leaf switches, and the method comprises delivering the data packets received in the given subnetwork via the detour route to a network end-node that is connected to the first lower-tier switch.

3. The method according to claim 1, wherein partitioning the links in each subnetwork comprises defining a third group of the links, wherein defining the corresponding detour route comprises configuring detour routes through the first and second groups of the links for the data packets that are destined for network end-nodes that are connected to the given subnetwork, andwherein the method comprises defining at least one dive-through route over which a data packet enters the subnetwork via the first virtual lane through a first trunk link to one of the upper-tier switches in the subnetwork, passes through one of the links in the third group to one of the lower-tier switches in the subnetwork and through another one of the links via the second virtual lane to another one of the upper-tier switches in the subnetwork, from which the data packet exits the subnetwork through a second trunk link.

4. The method according to claim 1, wherein defining the corresponding detour route comprises defining multiple detour routes for use in case of failures of multiple ones of the local links in the given subnetwork.

5. The method according to claim 1, wherein defining the corresponding detour route comprises, for any given local link, applying the corresponding detour route in rerouting the data packets arriving in the given subnetwork over any of a plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link.

6. The method according to claim 5, wherein configuring the primary routes comprises defining an adaptive routing group comprising the primary routes that are destined to the given subnetwork irrespective of the lower-tier switches to which the primary routes are directed, such that upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over detour routes within the given subnetwork.

7. A communication system, comprising: a network comprising multiple subnetworks interconnected by trunk links, each subnetwork comprising multiple switches, which are configured to communicate via at least first and second virtual lanes and are arranged in a bipartite topology comprising an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches; anda management processor, which is configured to define primary routes for delivery of data packets in the network, to partition the local links in each subnetwork into at least first and second groups, and for each local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, to define a corresponding detour route passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch within the given subnetwork, and from the second lower-tier switch over a second local link to a second upper-tier switch in the given subnetwork, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch,wherein the switches in the given subnetwork are configured, upon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, to reroute the data packets arriving at the first upper-tier switch for transmission to the first lower-tier switch so that the data packets pass through the corresponding detour route to the first lower-tier switch, while transmitting the data packets over the second and third local links via the second virtual lane.

8. The system according to claim 7, wherein the lower-tier switches comprise leaf switches, which are configured to deliver the data packets received in the given subnetwork via the detour route to a network end-node that is connected to one of the leaf switches.

9. The system according to claim 7, wherein the management processor is configured to partition the links in each subnetwork to define a third group of the links, wherein the detour routes are configured through the first and second groups of the links for the data packets that are destined for network end-nodes that are connected to the given subnetwork, andwherein the management processor is configured to define at least one dive-through route over which a data packet enters the subnetwork via the first virtual lane through a first trunk link to one of the upper-tier switches in the subnetwork, passes through one of the links in the third group to one of the lower-tier switches in the subnetwork and through another one of the links via the second virtual lane to another one of the upper-tier switches in the subnetwork, from which the data packet exits the subnetwork through a second trunk link.

10. The system according to claim 7, wherein the management processor is configured to define multiple detour routes for use in case of failures of multiple ones of the local links in the given subnetwork.

11. The system according to claim 7, wherein for any given local link, the corresponding detour route is applied in rerouting the data packets arriving in the given subnetwork over any of a plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link.

12. The system according to claim 11, wherein the management processor is configured to define an adaptive routing group comprising the primary routes that are destined to the given subnetwork irrespective of the lower-tier switches to which the primary routes are directed, such that upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over detour routes within the given subnetwork.

13. A method for communication, comprising: partitioning local links in a subnetwork of a packet data network into at least first and second groups;for each local link that connects a first upper-tier switch to a first lower-tier switch in the subnetwork, defining a corresponding detour route passing through a first local link belonging to the first group from the first upper-tier switch to a second lower-tier switch, and from the second lower-tier switch over a second local link to a second upper-tier switch, and from the second upper-tier switch over a third local link belonging to the second group to the first lower-tier switch; andupon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, rerouting data packets arriving from the network at the first upper-tier switch for forwarding over the local link, so that the data packets pass via the corresponding detour route to the first lower-tier switch.

14. The method according to claim 13, wherein defining the corresponding detour route comprises defining multiple detour routes for use in case of failures of multiple ones of the local links in the given subnetwork.

15. The method according to claim 13, wherein defining the corresponding detour route comprises, for any given local link, applying the corresponding detour route in rerouting the data packets arriving in the given subnetwork over any of a plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link.

16. The method according to claim 13, wherein rerouting the data packets comprises receiving the data packets from the network, and forwarding the data packets via a specified virtual lane from the second lower-tier switch over the second local link to the second upper-tier switch.

17. A method for communication, comprising: configuring primary routes for delivery of data packets in a network comprising multiple subnetworks interconnected by trunk links, each subnetwork comprising multiple switches in a bipartite topology comprising an upper tier containing upper-tier switches, which are connected to one or more of the trunk links, and a lower tier containing lower-tier switches, which are connected by local links to the upper-tier switches;for any given local link that is a part of one or more of the primary routes and connects a first upper-tier switch to a first lower-tier switch in a given subnetwork, defining a corresponding detour route passing through the local links within the given subnetwork from the first upper-tier switch to a second lower-tier switch, and from the second lower-tier switch to a second upper-tier switch and from the second upper-tier switch to the first lower-tier switch;when the given local link is a part of a plurality of the primary routes, applying the corresponding detour route to all of the plurality of the primary routes of which the given local link is a part, such that the primary routes continue to pass through the first upper-tier switch notwithstanding failures of the given local link; andupon a failure of the local link connecting the first upper-tier switch to the first lower-tier switch, rerouting the data packets arriving at the first spine switch in the given subnetwork to pass via the corresponding detour route to the first lower-tier switch.

18. The method according to claim 17, wherein configuring the primary routes comprises defining an adaptive routing group comprising the primary routes that are destined to the given subnetwork irrespective of the lower-tier switches to which the primary routes are directed, such that upon failures of any of the local links within the given subnetwork, the primary routes in the adaptive routing group are rerouted over detour routes within the given subnetwork.