Patent Application
20120106347
The present invention relates generally to data communications networks, and in particular to techniques for facilitating load balancing in a data packet network backbone that utilizes shortest path bridging.
Load distribution or load spreading is a method by which data communications bandwidth is more effectively utilized and overall performance is improved in a network. More particularly, equal-cost multi-path (ECMP) is one common strategy for load spreading of unicast traffic in routed networks. ECMP can be utilized where the decision as to how to forward a packet to a given destination can resolve to any of multiple “equal-cost” paths, each of which tied for being the shortest (or lowest cost) path when running network traversal calculations. ECMP can be used in conjunction with most unicast routing protocols and nodes equipped with the required supporting data plane hardware, since it relies on a per-hop decision that is local to a single router and assumes promiscuous receipt and a complete forwarding table at every intermediate node. When ECMP is used at a given node in a network, the traffic from that node is divided pseudo-evenly across the set of equal-cost next hops. This process is implemented independently at each hop of the network where more than one path to a given destination exists.
In many implementations, at each point where the presence of multiple equal-cost next hops is encountered, each packet is inspected for a source of entropy, such as an Internet Protocol (IP) header, and a hash of header information is used to select the next hop for the particular packet. For highly aggregated traffic, this method will on average distribute the load evenly in regular topologies (i.e., symmetric topologies) and does offer some improvement in less regular topologies.
The Institute of Electrical and Electronics Engineers (IEEE) 802.1aq standard for shortest path bridging (SPB) has been developed to allow the construction of full-mesh, shortest-path connectivity in an Ethernet network architecture. SPB consolidates a number of control protocols into a single link state routing system supported by the intermediate system to intermediate system (IS-IS) protocol. This system is used for the computation of integrated and congruent unicast and multi-cast forwarding to construct basic Ethernet LAN connectivity.
Ethernet network architectures, including those supporting 802.1aq, do not support the per-hop multi-path forwarding approach discussed above. This lack of support is a consequence of the need for congruence between unicast and multicast traffic and because multicast is generally incompatible with ECMP. Accordingly, improved techniques for load balancing in Ethernet networks supporting SPB are needed.
In several embodiments of the invention, a flow classification process is used at the edge of the shortest path bridging network to determine a flow label for attachment to a client frame entering the network. Because any of several flow labels can be assigned to a client frame traversing the network to a particular egress node, the flow labels can be used by forwarding nodes to select among multiple equal-cost paths. In several embodiments, the flow label is calculated as a function of the client frame contents, which provide an entropy source for randomizing the selection of the flow label. This entropy source comprises the Internet Protocol (IP) header in the client frame, in some embodiments, but may comprise other client frame content in other cases, such as the frame's destination-MAC address and source-MAC address. In some embodiments, an IP header is used as an entropy source when it is available, and a destination-MAC address/source-MAC address pair used otherwise.
In some embodiments of the techniques disclosed herein, the forwarding process is similar to conventional Ethernet routing, involving a lookup of a next-hop interface in a forwarding database indexed by the destination address and the flow label. In a variant of this approach, the forwarding database comprises destination addresses paired with corresponding sets of next-hop interfaces, and the forwarding process includes the selection of one of the next-hop interfaces using a calculation based on the flow label and, in some embodiments, based further on a node-specific seed value.
One example method for facilitating load balancing in a data packet network backbone that utilizes shortest path bridging begins with the receipt of one or more client frames at an ingress node of the network backbone. A flow offset value is calculated for each client frame as a function of the contents of the client frame; this flow offset value is one of N predetermined values, where N defines a maximum number of possible paths from the ingress node to a destination-MAC address for the client frame. The method continues with the attaching of a backbone virtual LAN identifier (B-VID) to each client frame, wherein the B-VID is a function of a predetermined B-VID base value and the flow offset value for the client frame. The client frame is then forwarded according to the B-VID and the backbone destination-MAC address for the client frame.
In some cases, the flow offset value is generated as a function of an Internet Protocol header in the client frame. For example, the flow offset value may be generated by calculating a hash of the Internet Protocol header and mapping the calculated hash to one of the N predetermined values. In some embodiments, this mapping may comprise computing the flow offset value as the modulo remainder of the hash divided by N, and wherein the B-VID is the sum of the predetermined B-VID base value and the flow offset value.
In other cases, the flow offset value may be generated as a function of a source-MAC address for the client frame and either a client destination MAC-address or the backbone destination-MAC address for the client frame. In still other cases, the method may further include first determining whether an Internet Protocol header is present in the client frame, and generating the flow offset value as a function of the Internet Protocol header, if present, and otherwise generating the flow offset value as a function of the destination-MAC address and a source-MAC address for the client frame, if no IP header is present.
Another example method is directed to techniques for populating a forwarding database in a network node. According to this method, the number of equal-cost next hops corresponding to a backbone destination-MAC address is determined. Each backbone virtual LAN identifier (B-VID) from a predetermined set of B-VIDs is mapped to a next-hop interface, as a function of the B-VID and the number of equal-cost next hops. Finally, a forwarding database is populated with entries corresponding to each combination of backbone destination-MAC address and B-VID, the entries comprising identifiers for the next-hop interfaces mapped to the B-VIDs. This process may be repeated for each of the backbone destination addresses known to the network node.
In some embodiments, the predetermined set of B-VIDs is a predetermined range of B-VIDs. In these embodiments, each B-VID may be mapped to a next-hop interface by subtracting a predetermined B-VID base value from the B-VID, to obtain a flow offset value, calculating an interface offset value as the flow offset value modulo the number of equal-cost next hops, and mapping the B-VID to the next-hop interface corresponding to the interface offset value. In other embodiments, the predetermined set of B-VIDs includes a plurality of B-VID ranges, and each B-VID range is mapped to one of the next-hop interfaces.
In another variant of the methods summarized above for populating a forwarding database in a network node, a set of equal-cost next hops is determined for each destination-MAC address, and the forwarding database is populated with entries mapping the backbone destination-MAC address to identifiers for each of the equal-cost next hops in the corresponding set. When this technique for populating a forwarding database is used, a forwarding node uses a modified technique for forwarding frames. More particularly, the forwarding node receives a data frame labeled with a backbone virtual LAN identifier (B-VID) and a destination-MAC address, and selects next-hop identifier from the set of equal-cost next-hop identifiers stored in the forwarding database in association with the destination-MAC address, as a function of the B-VID. The node then forwards the data frame via the next-hop interface corresponding to the selected next-hop identifier. In some of these embodiments, the node selects the next-hop identifier by computing a pseudorandom number as a function of the B-VID and a node-specific seed value, and selects the next-hop identifier from the set of equal-cost next-hop identifiers as a function of this pseudorandom number.
In addition to the methods summarized below, several variants are disclosed and described in further detail in the following detailed description. Network nodes configured to carry out one or more of these methods are also described. Thus, the present invention is not limited to the above features and advantages. Indeed, those skilled in the art will recognize additional features and advantages upon reading the following detailed description, and upon viewing the accompanying drawings.
It should be understood that the following description, while indicating several embodiments of the invention, is given by way of illustration only. Various changes and modifications within the scope of the invention as claimed will become apparent to those skilled in the art upon reading the following description and viewing the attached drawings.
The IEEE has developed the 802.1aq standards for Shortest Path Bridging to provide logical Ethernet networks on native Ethernet infrastructure. 802.1aq-enabled nodes (SPB-capable bridges) use the Intermediate System to Intermediate System (IS-IS) link state protocol to advertise network topology and membership in logical networks. This well-understood routing protocol allows nodes in the network to determine the shortest path between nodes automatically. Effectively, each bridge advertises the nodes it knows about to all the other nodes, so that all the bridges in the network eventually have a similar view of the network. Each node can then calculate the shortest (lowest cost) route to a target node and forward frames to the next hop in that shortest path.
The 802.1aq standards describe two variants of SPB. SPBV (Shortest Path Bridging—VID) is backwards compatible with Spanning Tree Protocols (STP), using the “region” concept of the Multiple Spanning Tree Protocol (MSTP), and is generally suitable for small Virtual Local-Area Networks (VLANs) having fewer than one hundred bridges. The other variant, known as SPBM (Shortest Path Bridging—MAC) is designed for carrier-grade applications, including large Provider Backbone Bridge (PBB) networks having as many as one thousand bridges.
In an SPBM network, unicast frames received at the edge of the network from an attached Ethernet device are encapsulated in mac-in-mac IEEE 802.1ah header, which is used by core bridges for forwarding the frame across the network. When the frame egresses the SPBM network at the far side, the IEEE 802.1ah header is stripped off. Source learning of Client MAC (C-MAC) addresses in a logical network is performed at the edge of the network, so that frames are forwarded to the backbone MAC (B-MAC) address for the egress node closest to the destination client. The B-MAC addresses for the SPB bridges are distributed and/or algorithmically constructed by the control plane, eliminating the need for B-MAC learning. SPBM supports unicast and multicast, with all routing on symmetric shortest paths, and support is provided for many equal-cost shortest paths.
By contrast,
One advantage of shortest path bridging is that mesh networks can be formed to distribute load more evenly across the network topology, since choke points can be eliminated. Redundant links are actually used, rather than sitting idle until a failure. The improved efficiency provides the potential to flatten an existing network, by removing one or more tiers, while maintaining or even improving the network's performance.
On the other hand, an SPB network is more dynamic than conventional Ethernet networks, which complicates traffic management and network troubleshooting. Furthermore, load balancing in a real network can be quite complex. The network 200 illustrated in
One possible approach is to adapt ECMP techniques to SPBM. This could be done in a number of different ways, using any of several sources of entropy for selecting between the available equal-cost paths. One possibility is to hash a combination of the B-MAC addresses (the source-address, B-SA, and destination-address, B-DA) included in the 802.11ah header to provide a source of entropy for selecting among the equal-cost paths. However, this is a relatively poor source of entropy. A somewhat better source of entropy is the I-SID; a still better choice is to hash the C-MAC SA/DA information, while perhaps the best choice is to hash the Internet Protocol (IP) 5-tuple, if present. Hop selection based on the hashing of an entropy source can be performed on a per-hop basis, i.e., at each forwarding node, or performed at the edge of the bridging network, in which case some resulting token of information is carried in the data packet to summarize the hash.
A higher entropy source for randomizing data flow across multiple equal-cost forwarding interfaces provides better performance, thus hashing of the IP payload provides the best entropy, especially if the routers are Ethernet connected, since a single source/destination router MAC address pair will originate packets with a large variety of IP 5-tuples. However, per-hop hashing of the IP payload is problematic, for several reasons. First, this approach is a layer violation. Second, this approach is incompatible with the existing Ethernet technology base. Finally, this makes Ethernet Operations, Administration, and Management (OAM) very difficult, if not impossible, as the OAM protocols would need to impersonate IP packets to collect network performance data.
Because of these problems with per-hop hashing, edge-based assignment of flows to an Ethernet layer-specific flow label is a better approach. With this approach, existing OAM protocols can be made to work, layer violations are confined to the edge adaptation function, and minimal changes to the overall technology base are required.
The use of multiple, B-VID identified, ECT sets as currently documented in the 802.1aq standards can be considered one variation of a flow label. However, this technique has limitations with respect to the efficiency of link utilization. Proposed enhancements to ECT set generation apply computational techniques to improve the efficiency of path selection, but the requirement for symmetric congruence of unicast and multicast results in high computational complexity of path selection. The Internet Engineering Task Force's (IETF's) TRILL specifications include a separation of unicast and multicast techniques to permit unicast ECMP as part of its design—TRILL uses spanning trees for multicast in order to reduce computational complexity. IETF's PWE specifications include the concept of an entropy label which can be considered to be a flow label, as do the IPv6 standards.
The inventive techniques disclosed herein are based on an exploration of what can be done with the Ethernet technology base for unicast load spreading if the symmetric congruence is relaxed. Under this relaxation of the conventional congruence requirements, the “go” and “return” paths do no need to be the same. Further, unicast and flooded frames do not need to follow the same routing. As will be seen below, this approach allows for the possibility of simplified computation and simplified tie-breaking rules for routing computations.
To minimize overall changes to existing Ethernet devices, the techniques outlined herein will use the backbone virtual LAN identifier (B-VID) as a unicast flow label, as well as to delineate VIDs used for multicast. More particularly, the B-VID flow label, which is applied to packets at the edge of the SPB network, is similar to an entropy label but is not used for per-hop hashing. Rather, the edge nodes hash a suitable entropy source to produce the B-VID flow label for unicast packets. Core nodes of the SPB network can simply forward packets, using conventional forwarding database (FDB) techniques, based on the B-VID flow label and the destination B-MAC.
Using this approach, a single logical network will use several B-VID flow label values. Accordingly, to implement this approach, a means of agreeing on the number of and range of values of B-VID based flow labels for a logical network is required. Similarly, a mechanism for deciding which B-VIDs are assigned to multicast operation is needed. The existing algorithm negotiation framework for IS-IS hellos defined for 802.1aq can serve as the basis for this mechanism. It should be noted that this mechanism can also use ranges of B-VIDs, as is done in MSTP/SVL (Multiple Spanning Tree Protocol/Shared VLAN Learning) operation.
The nominal dataplane configuration in networks employing this technique is as specified in 802.1aq or 802.1Qay, where all B-VIDs used as flow labels are unblocked on all interfaces. Nodes exchange algorithm and topology information as per existing 802.1aq and IS-IS procedures.
When a node is computing shortest paths, it populates the forwarding database (FDB) so that there is a unique entry for each DA-MAC/B-VID tuple. An example algorithm begins with the assumption of a contiguous range of B-VID values starting from a pre-defined base value. To avoid unnecessarily correlating what is supposed to be a random process, a node may choose to combine a per-node random value with the flow label prior to interface selection. For each unicast flow label B-VID, the node performs modulo arithmetic of the B-VID value (and possibly the per-node random value) by the number of equal-cost next hops. The result identifies the interface (among the equal-cost interfaces) that the destination MAC/B-VID tuple will point to in the FDB. The FDB is populated accordingly. Multicast packets are handled separately, under this approach. Thus, a network may also compute one or more multicast distribution trees and assign each of these to a multicast B-VID, which are distinct from the B-VIDs used for unicast.
Behavior at edge nodes is somewhat more complex. Upon receipt of a client frame, an edge node examines the frame contents to perform a form of “flow classification,” which results in a value that, modulo the number of flow labels and with the base value added, selects the flow label to be applied to the packet as part of SPBM adaptation. Core nodes receiving the labeled frame will simply look up the B-DA in the FDB and select the next-hop interface as a function of the B-VID flow label and the number of next hops associated with the B-DA forwarding entry.
Those skilled in the art will appreciate that a technique is required to ensure that broadcast segments do not cause duplicate unicast packets. Unless the behavior of forwarding nodes is modified, then every node but the sender attached to a broadcast segment will accept frames broadcast on the segment. There are a few ways this can be handled. First, a key requirement is that for a B-VID/DA tuple there is only one egress from a broadcast segment. One solution is that any node adjacent to a broadcast segment/pseudo-node needs to also compute the Shortest Path First (SPF) tree for the pseudo-node that represents the broadcast segment in the routing system. With this approach, this node can determine whether it is on the shortest path for a DNB-VID tuple from the upstream node, and set up ingress filtering accordingly. Note that if nodes combine a local random value with the flow selection, all nodes attached to the shared segment will need a consistent view of the random value used for the pseudo node. There are numerous well understood techniques by which this can be achieved that are not central to this disclosure.
As suggested above, the flow identifier calculation is preferably based on a good source of entropy to best randomize the distribution of packets across multiple equal-cost paths in the SPB network. The flow identifier effectively captures this entropy, and is thus a form of entropy label. However, because the flow identifier already reflects the randomness of the entropy source, the core nodes need not perform complex calculations to forward the packets—rather, the packets can be forwarded using a FDB populated according to the backbone destination address and the flow identifiers.
OFFSET=H mod N
With this approach, the hash values are evenly divided among the N possible values (zero to N−1) for OFFSET . Finally, as shown at block 430, a flow identifier (B−VID) is calculated by adding the offset value to a pre-determined base flow identifier value (BASE):
B−VID=BASE+OFFSET
The approach illustrated in
In some cases, the source of entropy used for generating the flow label may vary, depending on what type of packet is being processed.
If the client-MAC to B-MAC binding already exists, however, then the edge node checks to see whether the packet has an IP header, as shown at block 530. If so, then the IP header is hashed, as shown at block 540, and the result used to calculate the B-VID, as indicated at block 550. On the other hand, if no IP header is present, then the edge node instead hashes the B-MAC source and destination addresses, as shown at block 535, and that result is used to calculate the B-VID at block 550. The B-VID may be calculated in the same manner as shown in blocks 420 and 430 of
Starting with the first B-VID of a set of B-VIDs assigned to the logical network (block 630), the B-VID is then mapped to one of the next-hop interfaces for the equal-cost paths, based on the B-VID and the number of next hops, as shown at block 640. If the B-VIDs in the set are consecutive values, then one simple way to map each B-VID value to a next-hop interface is to subtract the base B-VID value from the B-VID value, to obtain a flow offset value, calculate an interface offset value as the flow offset value modulo the number of equal-cost next hops, and then use the interface offset value as an index to the next-hop interface, among the next-hop interfaces corresponding to equal-cost paths for the destination address. In mathematical terms:
flow_offset=BVID−BVIDBASE
next_hop_index=(flow_offset mod n,
where n is the number of next-hop interfaces at the node that correspond to an equal-cost path to the destination address, and where the corresponding next-hop interfaces are indexed as next_hop(0) to next hop(n).
When local randomness is added to the flow offset to reduce the degree of correlation between handling of flows, one example would be:
flow_offset=BVID−BVID BASE
next_hop_index=(flow_offset xor seed_value)mod n.
The forwarding database is populated with the B-MAC destination address (DA-MAC), the B-VID, and an identifier for the next-hop interface, as shown at block 650. If there are more unprocessed B-VIDs in the set of B-VIDs for the current destination address, the process is repeated, as shown at blocks 660 and 665. Then, the process is repeated for each remaining destination address known to the network node, as shown at blocks 670 and 675. When the process is complete, the FDB for a given logical network includes multiple entries for each destination address, with each entry corresponding to one of the B-VIDs allocated for unicast forwarding in the logical network.
Table 1 illustrates an example corresponding to a small portion of an FDB, where there are four interfaces on the router, numbered I—0 to I—3, and where there are eight B-VIDs allocated to unicast forwarding for a given logical network, ranging from B-VID—0 to B-VID—8.
In this (simplistic) example, there happen to be two next-hop interfaces for each destination address, but the pair of next-hop interfaces differs for each of the three destination addresses shown. Thus, I—0 and I—3 are identifiers for the next-hops for the equal-cost paths to destination address B-MAC-DA-a, while the next-hop identifiers for the equal-cost paths to destination B-MAC-DA-b are 1—1 and I—2. For each destination address, the eight available B-VID values are split between the two equal-cost next-hops.
In a variant of the above techniques, one or more of the individual B-VIDs can be replaced with a range of B-VIDs. This approach may be suitable when there are a large number of flow labels available for a given logical network, compared to the maximum number of equal-cost paths that are supported. For example, if there are 4096 available flow labels (B-VIDs), but only 64 possible equal-cost paths, then the B-VIDs could be mapped to next-hop identifiers in “chunks” of 64. Thus, BVIDO in Table 1 might be replaced with a range, e.g., BVID0-BVID63, while BVID1 is replaced with a second range, e.g., BVID64 - BVID127.
Given a FDB populated according to the techniques described above and illustrated in
The above techniques for populating the FDB in backbone network nodes is generally compatible with existing Ethernet technology, in that the packet processing is based on a simple look-up and forward approach. Another approach to FDB and packet forwarding, which may require more extensive modifications to network node hardware, is illustrated in
As shown at block 730, an FDB is populated with the destination address and the corresponding set of next-hop identifiers. An example of this different type of FDB, corresponding to the same situation illustrated in Table 1, is illustrated in Table 2.
It will be immediately appreciated that an FDB according to this approach consumes less memory (i.e., represents less state information) than the FDB populated according to the earlier described approach.
A forwarding process 800 corresponding to this type of FDB is illustrated in
In some embodiments, the selection of the next-hop identifier based on the BVID uses a mathematical approach similar to that described above for populating the FDB. Thus, for example, an index can be computed using modulo division of the BVID value by the number of next-hop interfaces, e.g.,
next_hop_index=(BVID−BVIDBASE
where n is the number of next-hop interfaces in the set corresponding to the destination address. However, one advantage to this dynamic processing of the flow label is that randomization of the packet flows through the network can be improved by adding a seed value, such as a node-specific seed value, to the next-hop identification process. Thus, for example, the next-hop interface could be selected using a formula like:
next_hop_index=(BVID−BVIDBASE
where the value of seed varies from one node to the next, or from time to time, or both.
While the above discussion has described certain processes performed at edge nodes, i.e., edge nodes attached to client networks and/or client computers, and at forwarding nodes, which include nodes in the core of an SPB network. Of course, edge nodes may in some cases also be forwarding nodes, so one or several of the techniques may be implemented at a given node.
More generally speaking, these computer program instructions may be provided to a processor of a general purpose computer, a special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions described herein and illustrated in the accompanying block diagrams and flow diagrams. These computer program instructions may also be stored in a computer usable or computer-readable memory that may direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer usable or computer-readable memory produce an article of manufacture including instructions that implement the functions described herein. Several of the functional entities described herein may be implemented together on a single processor or computer, or each may be implemented on separate processors or computers. Those skilled in the art will recognize the advantages and disadvantages of grouping or distributing the functions described herein.
Of course, those skilled in the art will appreciate that the load balancing techniques described above are merely exemplary, and that various modifications and extensions of these procedures are possible. Those skilled in the art will also recognize that the functional aspects of the network node 900 described herein may be deployed in a number of different configurations, depending on the network structures, operator business models, and other factors.
With these and other variations and extensions in mind, those skilled in the art will appreciate that the foregoing description and the accompanying drawings represent non-limiting examples of the systems and apparatus taught herein for facilitating load balancing in a data packet network. As such, the present invention is not limited by the foregoing description and accompanying drawings. Instead, the present invention is limited only by the following claims and their legal equivalents.
This application claims priority under 35 U.S.C. §119(e) to Provisional Patent Application Ser. No. 61/408,037, filed 29 Oct. 2010, and also claims priority under 35 U.S.C. §119(e) to Provisional Patent Application Ser. No. 61/430,281, filed 6 Jan. 2011. The entire contents of each of the aforementioned provisional applications are incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 61408037 | Oct 2010 | US | |
| 61430281 | Jan 2011 | US |
