OPSAWG                                                      R. Krishnan
Internet Draft                                   Brocade Communications
Intended status: Informational                                  L. Yong
Expires: December 13, 2014                                   Huawei USA
                                                            A. Ghanwani
                                                                Ning So
                                                    Tata Communications
                                                          B. Khasnabish
                                                        ZTE Corporation
                                                          June 13, 2014

     Mechanisms for Optimizing LAG/ECMP Component Link Utilization in



 Status of this Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79. This document may not be modified,
   and derivative works of it may not be created, except to publish it
   as an RFC and to translate it into languages other than English.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as Internet-

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   The list of current Internet-Drafts can be accessed at

   The list of Internet-Draft Shadow Directories can be accessed at

   This Internet-Draft will expire on December 13, 2014.

Copyright Notice

   Copyright (c) 2014 IETF Trust and the persons identified as the
   document authors. All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document. Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document. Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.


   Demands on networking infrastructure are growing exponentially due to
   bandwidth hungry applications such as rich media applications and
   inter-data center communications. In this context, it is important to
   optimally use the bandwidth in wired networks that extensively use
   link aggregation groups and equal cost multi-paths as techniques for
   bandwidth scaling. This draft explores some of the mechanisms useful
   for achieving this.

Table of Contents

   1. Introduction...................................................3
      1.1. Acronyms..................................................4
      1.2. Terminology...............................................4
   2. Flow Categorization............................................5
   3. Hash-based Load Distribution in LAG/ECMP.......................6
   4. Mechanisms for Optimizing LAG/ECMP Component Link Utilization..7
      4.1. Differences in LAG vs ECMP................................8
      4.2. Operational Overview......................................9
      4.3. Large Flow Recognition...................................10
         4.3.1. Flow Identification.................................10
         4.3.2. Criteria and Techniques for Large Flow Recognition..11
         4.3.3. Sampling Techniques.................................11
         4.3.4. Inline Data Path Measurement........................13
         4.3.5. Use of More Than One Method for Large Flow
      4.4. Load Rebalancing Options.................................14
         4.4.1. Alternative Placement of Large Flows................14
         4.4.2. Redistributing Small Flows..........................15
         4.4.3. Component Link Protection Considerations............15
         4.4.4. Load Rebalancing Algorithms.........................15
         4.4.5. Load Rebalancing Example............................16
   5. Information Model for Flow Rebalancing........................17
      5.1. Configuration Parameters for Flow Rebalancing............17
      5.2. System Configuration and Identification Parameters.......18
      5.3. Information for Alternative Placement of Large Flows.....19
      5.4. Information for Redistribution of Small Flows............19
      5.5. Export of Flow Information...............................20
      5.6. Monitoring information...................................20
         5.6.1. Interface (link) utilization........................20
         5.6.2. Other monitoring information........................20
   6. Operational Considerations....................................21
      6.1. Rebalancing Frequency....................................21
      6.2. Handling Route Changes...................................21
      6.3. Forwarding Resources.....................................21
   7. IANA Considerations...........................................22
   8. Security Considerations.......................................22
   9. Contributing Authors..........................................22
   10. Acknowledgements.............................................22
   11. References...................................................23
      11.1. Normative References....................................23
      11.2. Informative References..................................23

1. Introduction

   Networks extensively use link aggregation groups (LAG) [802.1AX] and
   equal cost multi-paths (ECMP) [RFC 2991] as techniques for capacity
   scaling. For the problems addressed by this document, network traffic
   can be predominantly categorized into two traffic types: long-lived
   large flows and other flows.  These other flows, which include long-
   lived small flows, short-lived small flows, and short-lived large
   flows, are referred to as "small flows" in this document.  Long-lived
   large flows are simply referred to as "large flows."

   Stateless hash-based techniques [ITCOM, RFC 2991, RFC 2992, RFC 6790]
   are often used to distribute both large flows and small flows over
   the component links in a LAG/ECMP. However the traffic may not be
   evenly distributed over the component links due to the traffic

   This draft describes mechanisms for optimizing LAG/ECMP component
   link utilization while using hash-based techniques. The mechanisms
   comprise the following steps -- recognizing large flows in a router;
   and assigning the large flows to specific LAG/ECMP component links or
   redistributing the small flows when a component link on the router is

   It is useful to keep in mind that in typical use cases for this
   mechanism the large flows are those that consume a significant amount
   of bandwidth on a link, e.g. greater than 5% of link bandwidth.  The
   number of such flows would necessarily be fairly small, e.g. on the
   order of 10's or 100's per LAG/ECMP.  In other words, the number of
   large flows is NOT expected to be on the order of millions of flows.
   Examples of such large flows would be IPsec tunnels in service
   provider backbone networks or storage backup traffic in data center

1.1. Acronyms

   DOS: Denial of Service

   ECMP: Equal Cost Multi-path

   GRE: Generic Routing Encapsulation

   LAG: Link Aggregation Group

   MPLS: Multiprotocol Label Switching

   NVGRE: Network Virtualization using Generic Routing Encapsulation

   PBR: Policy Based Routing

   QoS: Quality of Service

   STT: Stateless Transport Tunneling

   TCAM: Ternary Content Addressable Memory

   VXLAN: Virtual Extensible LAN

1.2. Terminology

   Central management entity: Refers to an entity that is capable of
   monitoring information about link utilization and flows in routers
   across the network and may be capable of making traffic engineering
   decisions for placement of large flows.  It may include the functions
   of a collector if the routers employ a sampling technique [RFC 7011].

   ECMP component link: An individual nexthop within an ECMP group.  An
   ECMP component link may itself comprise a LAG.

   ECMP table: A table that is used as the nexthop of an ECMP route that
   comprises the set of component links and the weights associated with
   each of those component links.  The weights are used to determine
   which values of the hash function map to a given component link.

   LAG component link: An individual link within a LAG.  A LAG component
   link is typically a physical link.

   LAG table: A table that is used as the output port which is a LAG
   that comprises the set of component links and the weights associated
   with each of those component links.  The weights are used to
   determine which values of the hash function map to a given component

   Large flow(s): Refers to long-lived large flow(s).

   Small flow(s): Refers to any of, or a combination of, long-lived
   small flow(s), short-lived small flows, and short-lived large

2. Flow Categorization

   In general, based on the size and duration, a flow can be categorized
   into any one of the following four types, as shown in Figure 1:

      (a) Short-lived Large Flow (SLLF),
      (b) Short-lived Small Flow (SLSF),
      (c) Long-lived Large Flow (LLLF), and
      (d) Long-lived Small Flow (LLSF).

        Flow Size Bandwidth
            |                    |                    |
      Large |      SLLF          |       LLLF         |
      Flow  |                    |                    |
            |                    |                    |
      Small |      SLSF          |       LLSF         |
      Flow  |                    |                    |
            +--------------------+--------------------+-->Flow Duration
                 Short-lived            Long-lived
                 Flow                   Flow

          Figure 1: Flow Categorization

   In this document, as mentioned earlier, we categorize long-lived
   large flows as "large flows", and all of the others -- long-lived
   small flows, short-lived small flows, and short-lived large flows
   as "small flows".

3. Hash-based Load Distribution in LAG/ECMP

   Hash-based techniques are often used for traffic load balancing to
   select among multiple available paths within a LAG/ECMP group. The
   advantages of hash-based techniques for load distribution are the
   preservation of the packet sequence in a flow and the real-time
   distribution without maintaining per-flow state in the router. Hash-
   based techniques use a combination of fields in the packet's headers
   to identify a flow, and the hash function computed using these fields
   is used to generate a unique number that identifies a link/path in a
   LAG/ECMP group. The result of the hashing procedure is a many-to-one
   mapping of flows to component links.

   If the traffic mix constitutes flows such that the result of the hash
   function across these flows is fairly uniform so that a similar
   number of flows is mapped to each component link, if the individual
   flow rates are much smaller as compared to the link capacity, and if
   the rate differences are not dramatic, hash-based techniques produce
   good results with respect to utilization of the individual component
   links. However, if one or more of these conditions are not met, hash-
   based techniques may result in imbalance in the loads on individual
   component links.

   One example is illustrated in Figure 2.  In Figure 2, there are two
   routers, R1 and R2, and there is a LAG between them which has 3
   component links (1), (2), (3).  There are a total of 10 flows that
   need to be distributed across the links in this LAG.  The result of
   applying the hash-based technique is as follows:

     .  Component link (1) has 3 flows -- 2 small flows and 1 large
        flow -- and the link utilization is normal.

     .  Component link (2) has 3 flows -- 3 small flows and no large
        flow -- and the link utilization is light.

          o The absence of any large flow causes the component link

     .  Component link (3) has 4 flows -- 2 small flows and 2 large
        flows -- and the link capacity is exceeded resulting in

          o The presence of 2 large flows causes congestion on this
             component link.

                  +-----------+ ->     +-----------+
                  |           | ->     |           |
                  |           | ===>   |           |
                  |        (1)|--------|(1)        |
                  |           | ->     |           |
                  |           | ->     |           |
                  |   (R1)    | ->     |    (R2)   |
                  |        (2)|--------|(2)        |
                  |           | ->     |           |
                  |           | ->     |           |
                  |           | ===>   |           |
                  |           | ===>   |           |
                  |        (3)|--------|(3)        |
                  |           |        |           |
                  +-----------+        +-----------+

            Where: ->   small flow
                   ===> large flow

                Figure 2: Unevenly Utilized Component Links

   This document presents mechanisms for addressing the imbalance in
   load distribution resulting from commonly used hash-based techniques
   for LAG/ECMP that were shown in the above example. The mechanisms use
   large flow awareness to compensate for the imbalance in load

4. Mechanisms for Optimizing LAG/ECMP Component Link Utilization

   The suggested mechanisms in this draft are about a local optimization
   solution; they are local in the sense that both the identification of
   large flows and re-balancing of the load can be accomplished
   completely within individual nodes in the network without the need
   for interaction with other nodes.

   This approach may not yield a global optimization of the placement of
   large flows across multiple nodes in a network, which may be
   desirable in some networks. On the other hand, a local approach may
   be adequate for some environments for the following reasons:

      1) Different links within a network experience different levels of
   utilization and, thus, a "targeted" solution is needed for those hot-
   spots in the network.  An example is the utilization of a LAG between
   two routers that needs to be optimized.

      2) Some networks may lack end-to-end visibility, e.g. when a
   certain network, under the control of a given operator, is a transit
   network for traffic from other networks that are not under the
   control of the same operator.

4.1. Differences in LAG vs ECMP

   While the mechanisms explained herein are applicable to both LAGs and
   ECMP groups, it is useful to note that there are some key differences
   between the two that may impact how effective the mechanism is.  This
   relates, in part, to the localized information with which the scheme
   is intended to operate.

   A LAG is usually established across links that are between 2 adjacent
   routers.  As a result, the scope of problem of optimizing the
   bandwidth utilization on the component links is fairly narrow.  It
   simply involves re-balancing the load across the component links
   between these two routers, and there is no impact whatsoever to other
   parts of the network.  The scheme works equally well for unicast and
   multicast flows.

   On the other hand, with ECMP, redistributing the load across
   component links that are part of the ECMP group may impact traffic
   patterns at all of the nodes that are downstream of the given router
   between itself and the destination.  The local optimization may
   result in congestion at a downstream node.  (In its simplest form, an
   ECMP group may be used to distribute traffic on component links that
   are between two adjacent routers, and in that case, the ECMP group is
   no different than a LAG for the purpose of this discussion.  It
   should be noted that an ECMP component link may itself comprise a
   LAG, in which case the scheme may be further applied to the component
   links within the LAG.)

                      +-----+     +-----+
                      | S1  |     | S2  |
                      +-----+     +-----+
                       / \ \       / /\
                      / +---------+ /  \
                     / /  \  \     /    \
                    / /    \  +------+   \
                   / /      \    /    \   \
                +-----+    +-----+   +-----+
                | L1  |    | L2  |   | L3  |
                +-----+    +-----+   +-----+

                Figure 3: Two-level Clos Network

   To demonstrate the limitations of local optimization, consider a two-
   level Clos network topology as shown in Figure 3 with three leaf
   nodes (L1, L2, L3) and two spine nodes (S1, S2). Assume all of the
   links are 10 Gbps.

   Let L1 have two flows of 4 Gbps each towards L3, and let L2 have one
   flow of 7 Gbps also towards L3.  If L1 balances the load optimally
   between S1 and S2, and L2 sends the flow via S1, then the downlink
   from S1 to L3 would get congested resulting in packet discards.  On
   the other hand, if L1 had sent both its flows towards S1 and L2 had
   sent its flow towards S2, there would have been no congestion at
   either S1 or S2.

   The other issue with applying this scheme to ECMP groups is that it
   may not apply equally to unicast and multicast traffic because of the
   way multicast trees are constructed.

   Finally, it is possible for a single physical link to participate as
   a component link in multiple ECMP groups, whereas with LAGs, a link
   can participate as a component link of only one LAG.

4.2. Operational Overview

   The various steps in optimizing LAG/ECMP component link utilization
   in networks are detailed below:

   Step 1) This involves large flow recognition in routers and
   maintaining the mapping of the large flow to the component link that
   it uses. The recognition of large flows is explained in Section 4.3.

   Step 2) The egress component links are periodically scanned for link
   utilization and the imbalance for the LAG/ECMP group is monitored. If
   the imbalance exceeds a certain imbalance threshold, then re-
   balancing is triggered. Measurement of the imbalance is discussed
   further in 5.1. Additional criteria may also be used to determine
   whether or not to trigger rebalancing, such as the maximum
   utilization of any of the component links, in addition to the

   Step 3) As a part of rebalancing, the operator can choose to
   rebalance the large flows on to lightly loaded component links of the
   LAG/ECMP group, redistribute the small flows on the congested link to
   other component links of the group, or a combination of both.

   All of the steps identified above can be done locally within the
   router itself or could involve the use of a central management

   Providing large flow information to a central management entity
   provides the capability to globally optimize flow distribution as
   described in Section 4.1.  Consider the following example.  A router
   may have 3 ECMP nexthops that lead down paths P1, P2, and P3.  A
   couple of hops downstream on path P1 there may be a congested link,
   while paths P2 and P3 may be under-utilized. This is something that
   the local router does not have visibility into.  With the help of a
   central management entity, the operator could redistribute some of
   the flows from P1 to P2 and/or P3 resulting in a more optimized flow
   of traffic.

   The mechanisms described above are especially useful when bundling
   links of different bandwidths for e.g. 10 Gbps and 100 Gbps as
   described in [ID.ietf-rtgwg-cl-requirement].

4.3. Large Flow Recognition

4.3.1. Flow Identification

   A flow (large flow or small flow) can be defined as a sequence of
   packets for which ordered delivery should be maintained.  Flows are
   typically identified using one or more fields from the packet header,
   for example:

     .  Layer 2: source MAC address, destination MAC address, VLAN ID.

     .  IP header: IP Protocol, IP source address, IP destination
        address, flow label (IPv6 only), TCP/UDP source port, TCP/UDP
        destination port.

     .  MPLS Labels.

   For tunneling protocols like Generic Routing Encapsulation (GRE)
   [RFC 2784], Virtual eXtensible Local Area Network (VXLAN) [VXLAN],
   Network Virtualization using Generic Routing Encapsulation (NVGRE)
   [NVGRE], Stateless Transport Tunneling (STT) [STT], Layer 2 Tunneling
   Protocol (L2TP) [RFC 3931], etc., flow identification is possible
   based on inner and/or outer headers as well as fields introduced by
   the tunnel header, as any or all such fields may be used for load
   balancing decisions [RFC 5640].  The above list is not exhaustive.
   The mechanisms described in this document are agnostic to the fields
   that are used for flow identification.

   This method of flow identification is consistent with that of IPFIX
   [RFC 7011].

4.3.2. Criteria and Techniques for Large Flow Recognition

   From a bandwidth and time duration perspective, in order to recognize
   large flows we define an observation interval and observe the
   bandwidth of the flow over that interval.  A flow that exceeds a
   certain minimum bandwidth threshold over that observation interval
   would be considered a large flow.

   The two parameters -- the observation interval, and the minimum
   bandwidth threshold over that observation interval -- should be
   programmable to facilitate handling of different use cases and
   traffic characteristics. For example, a flow which is at or above 10%
   of link bandwidth for a time period of at least 1 second could be
   declared a large flow [DevoFlow].

   In order to avoid excessive churn in the rebalancing, once a flow has
   been recognized as a large flow, it should continue to be recognized
   as a large flow for as long as the traffic received during an
   observation interval exceeds some fraction of the bandwidth
   threshold, for example 80% of the bandwidth threshold.

   Various techniques to recognize a large flow are described below.

4.3.3. Sampling Techniques

   A number of routers support sampling techniques such as sFlow [sFlow-
   v5, sFlow-LAG], PSAMP [RFC 5475] and NetFlow Sampling [RFC 3954].
   For the purpose of large flow recognition, sampling needs to be
   enabled on all of the egress ports in the router where such
   measurements are desired.

   Using sFlow as an example, processing in a sFlow collector will
   provide an approximate indication of the large flows mapping to each
   of the component links in each LAG/ECMP group.  It is possible to
   implement this part of the collector function in the control plane of
   the router reducing dependence on an external management station,
   assuming sufficient control plane resources are available.

   If egress sampling is not available, ingress sampling can suffice
   since the central management entity used by the sampling technique
   typically has multi-node visibility and can use the samples from an
   immediately downstream node to make measurements for egress traffic
   at the local node.

   The option of using ingress sampling for this purpose may not be
   available if the downstream device is under the control of a
   different operator, or if the downstream device does not support

   Alternatively, since sampling techniques require that the sample be
   annotated with the packet's egress port information, ingress sampling
   may suffice.  However, this means that sampling would have to be
   enabled on all ports, rather than only on those ports where such
   monitoring is desired. There is one situation in which this approach
   may not work.  If there are tunnels that originate from the given
   router, and if the resulting tunnel comprises the large flow, then
   this cannot be deduced from ingress sampling at the given router.
   Instead, if egress sampling is unavailable, then ingress sampling
   from the downstream router must be used.

   To illustrate the use of ingress versus egress sampling, we refer to
   Figure 2.  Since we are looking at rebalancing flows at R1, we would
   need to enable egress sampling on ports (1), (2), and (3) on R1.  If
   egress sampling is not available, and if R2 is also under the control
   of the same administrator, enabling ingress sampling on R2's ports
   (1), (2), and (3) would also work, but it would necessitate the
   involvement of a central management entity in order for R1 to obtain
   large flow information for each of its links.  Finally, R1 can enable
   ingress sampling only on all of its ports (not just the ports that
   are part of the LAG/ECMP group being monitored) and that would
   suffice if the sampling technique annotates the samples with the
   egress port information.

   The advantages and disadvantages of sampling techniques are as


     .  Supported in most existing routers.

     .  Requires minimal router resources.


     .  In order to minimize the error inherent in sampling, there is a
        minimum delay for the recognition time of large flows, and in
        the time that it takes to react to this information.

   With sampling, the detection of large flows can be done on the order
   of one second [DevoFlow].  A discussion on determining the
   appropriate sampling frequency is available in the following
   reference [SAMP-BASIC].

4.3.4. Inline Data Path Measurement

   Implementations may perform recognition of large flows by performing
   measurements on traffic in the data path of a router. Such an
   approach would be expected to operate at the interface speed on every
   interface, accounting for all packets processed by the data path of
   the router.  An example of such an approach is described in IPFIX
   [RFC 5470].

   Using inline data path measurement, a faster and more accurate
   indication of large flows mapped to each of the component links in a
   LAG/ECMP group may be possible (as compared to the sampling-based

   The advantages and disadvantages of inline data path measurement are:


     .  As link speeds get higher, sampling rates are typically reduced
        to keep the number of samples manageable which places a lower
        bound on the detection time.  With inline data path measurement,
        large flows can be recognized in shorter windows on higher link
        speeds since every packet is accounted for [NDTM].

     .  Eliminates the potential dependence on an external management
        station for large flow recognition.


     .  It is more resource intensive in terms of the tables sizes
        required for monitoring all flows in order to perform the

   As mentioned earlier, the observation interval for determining a
   large flow and the bandwidth threshold for classifying a flow as a
   large flow should be programmable parameters in a router.

   The implementation details of inline data path measurement of large
   flows is vendor dependent and beyond the scope of this document.

4.3.5. Use of More Than One Method for Large Flow Recognition

   It is possible that a router may have line cards that support a
   sampling technique while other line cards support inline data path
   measurement of large flows.  As long as there is a way for the router
   to reliably determine the mapping of large flows to component links
   of a LAG/ECMP group, it is acceptable for the router to use more than
   one method for large flow recognition.

   If both methods are supported, inline data path measurement may be
   preferable because of its speed of detection [FLOW-ACC].

4.4. Load Rebalancing Options

   Below are suggested techniques for load rebalancing.  Equipment
   vendors may implement more than one technique, including those not
   described in this document, allowing the operator to choose between

   Note that regardless of the method used, perfect rebalancing of large
   flows may not be possible since flows arrive and depart at different
   times.  Also, any flows that are moved from one component link to
   another may experience momentary packet reordering.

4.4.1. Alternative Placement of Large Flows

   Within a LAG/ECMP group, the member component links with least
   average port utilization are identified.  Some large flow(s) from the
   heavily loaded component links are then moved to those lightly-loaded
   member component links using a policy-based routing (PBR) rule in the
   ingress processing element(s) in the routers.

   With this approach, only certain large flows are subjected to
   momentary flow re-ordering.

   When a large flow is moved, this will increase the utilization of the
   link that it moved to potentially creating imbalance in the
   utilization once again across the component links.  Therefore, when
   moving large flows, care must be taken to account for the existing
   load, and what the future load will be after large flow has been
   moved.  Further, the appearance of new large flows may require a
   rearrangement of the placement of existing flows.

   Consider a case where there is a LAG compromising four 10 Gbps
   component links and there are four large flows, each of 1 Gbps.
   These flows are each placed on one of the component links.
   Subsequent, a fifth large flow of 2 Gbps is recognized and to
   maintain equitable load distribution, it may require placement of one
   of the existing 1 Gbps flow to a different component link.  And this
   would still result in some imbalance in the utilization across the
   component links.

4.4.2. Redistributing Small Flows

   Some large flows may consume the entire bandwidth of the component
   link(s). In this case, it would be desirable for the small flows to
   not use the congested component link(s). This can be accomplished in
   one of the following ways.

   This method works on some existing router hardware. The idea is to
   prevent, or reduce the probability, that the small flow hashes into
   the congested component link(s).

     .  The LAG/ECMP table is modified to include only non-congested
        component link(s). Small flows hash into this table to be mapped
        to a destination component link. Alternatively, if certain
        component links are heavily loaded, but not congested, the
        output of the hash function can be adjusted to account for large
        flow loading on each of the component links.

     .  The PBR rules for large flows (refer to Section 4.4.1) must
        have strict precedence over the LAG/ECMP table lookup result.

   With this approach the small flows that are moved would be subject to

4.4.3. Component Link Protection Considerations

   If desired, certain component links may be reserved for link
   protection. These reserved component links are not used for any flows
   in the absence of any failures. In the case when the component
   link(s) fail, all the flows on the failed component link(s) are moved
   to the reserved component link(s). The mapping table of large flows
   to component link simply replaces the failed component link with the
   reserved link. Likewise, the LAG/ECMP table replaces the failed
   component link with the reserved link.

4.4.4. Load Rebalancing Algorithms

   Specific algorithms for placement of large flows are out of scope of
   this document.  One possibility is to formulate the problem for large
   flow placement as the well-known bin-packing problem and make use of
   the various heuristics that are available for that problem [bin-

4.4.5. Load Rebalancing Example

   Optimizing LAG/ECMP component utilization for the use case in Figure
   2 is depicted below in Figure 4. The large flow rebalancing explained
   in Section 4.4 is used. The improved link utilization is as follows:

     .  Component link (1) has 3 flows -- 2 small flows and 1 large
        flow -- and the link utilization is normal.

     .  Component link (2) has 4 flows -- 3 small flows and 1 large
        flow -- and the link utilization is normal now.

     .  Component link (3) has 3 flows -- 2 small flows and 1 large
        flow -- and the link utilization is normal now.

                  +-----------+ ->     +-----------+
                  |           | ->     |           |
                  |           | ===>   |           |
                  |        (1)|--------|(1)        |
                  |           |        |           |
                  |           | ===>   |           |
                  |           | ->     |           |
                  |           | ->     |           |
                  |  (R1)     | ->     |   (R2)    |
                  |        (2)|--------|(2)        |
                  |           |        |           |
                  |           | ->     |           |
                  |           | ->     |           |
                  |           | ===>   |           |
                  |        (3)|--------|(3)        |
                  |           |        |           |
                  +-----------+        +-----------+

            Where: ->   small flow
                   ===> large flow

                 Figure 4: Evenly Utilized Composite Links

   Basically, the use of the mechanisms described in Section 4.4.1
   resulted in a rebalancing of flows where one of the large flows on
   component link (3) which was previously congested was moved to
   component link (2) which was previously under-utilized.

5. Information Model for Flow Rebalancing

   In order to support flow rebalancing in a router from an external
   system, the exchange of some information is necessary between the
   router and the external system. This section provides an exemplary
   information model covering the various components needed for the
   purpose.  The model is intended to be informational and may be used
   as input for development of a data model.

5.1. Configuration Parameters for Flow Rebalancing

   The following parameters are required the configuration of this

     .  Large flow recognition parameters:

          o Observation interval: The observation interval is the time
             period in seconds over which the packet arrivals are
             observed for the purpose of large flow recognition.

          o Minimum bandwidth threshold: The minimum bandwidth threshold
             would be configured as a percentage of link speed and
             translated into a number of bytes over the observation
             interval.  A flow for which the number of bytes received,
             for a given observation interval, exceeds this number would
             be recognized as a large flow.

          o Minimum bandwidth threshold for large flow maintenance: The
             minimum bandwidth threshold for large flow maintenance is
             used to provide hysteresis for large flow recognition.
             Once a flow is recognized as a large flow, it continues to
             be recognized as a large flow until it falls below this
             threshold.  This is also configured as a percentage of link
             speed and is typically lower than the minimum bandwidth
             threshold defined above.

     .  Imbalance threshold: A measure of the deviation of the
        component link utilizations from the utilization of the overall
        LAG/ECMP group.  Since component links can be of a different
        speed, the imbalance can be computed as follows.  Let the
        utilization of each component link in a LAG/ECMP group with n
        links of speed b_1, b_2 ... b_n, be u_1, u_2 ... u_n. The mean
        utilization is computed is u_ave = [ (u_1 x b_1) + (u_2 x b_2) +
        ... + (u_n x b_n) ] / [b_1 + b_2 + ... + b_n].  The imbalance is
        then computed as max_{i=1 ... n} max_{i=1...n} | u_i - u_ave |.

     .  Rebalancing interval: The minimum amount of time between
        rebalancing events.  This parameter ensures that rebalancing is
        not invoked too frequently as it impacts packet ordering.

   These parameters may be configured on a system-wide basis or it may
   apply to an individual LAG.  It may be applied to an ECMP group
   provided the component links are not shared with any other ECMP

5.2. System Configuration and Identification Parameters

   The following parameters are useful for router configuration and
   operation when using the mechanisms in this document.

     .  IP address: The IP address of a specific router that the
        feature is being configured on, or that the large flow placement
        is being applied to.

     .  LAG ID: Identifies the LAG on a given router. The LAG ID may be
        required when configuring this feature (to apply a specific set
        of large flow identification parameters to the LAG) and will be
        required when specifying flow placement to achieve the desired

     .  Component Link ID: Identifies the component link within a LAG
        or ECMP group.  This is required when specifying flow placement
        to achieve the desired rebalancing.

     .  Component Link Weight: The relative weight to be applied to
        traffic for a given component link when using hash-based
        techniques for load distribution.

     .  ECMP group: Identifies a particular ECMP group.  The ECMP group
        may be required when configuring this feature (to apply a
        specific set of large flow identification parameters to the ECMP
        group) and will be required when specifying flow placement to
        achieve the desired rebalancing.  We note that multiple ECMP
        groups can share an overlapping set (or non-overlapping subset)
        of component links.  This document does not deal with the
        complexity of addressing such configurations.

   The feature may be configured globally for all LAGs and/or for all
   ECMP groups, or it may be configured specifically for a given LAG or
   ECMP group.

5.3. Information for Alternative Placement of Large Flows

   In cases where large flow recognition is handled by an external
   management station (see Section 4.3.3), an information model for
   flows is required to allow the import of large flow information to
   the router.

   Typical fields use for identifying large flows were discussed in
   Section 4.3.1.  The IPFIX information model [RFC 7012] can be
   leveraged for large flow identification.

   Large Flow placement is achieved by specifying the relevant flow
   information along with the following:

     .  For LAG: Router's IP address, LAG ID, LAG component link ID.

     .  For ECMP: Router's IP address, ECMP group, ECMP component link

   In the case where the ECMP component link itself comprises a LAG, we
   would have to specify the parameters for both the ECMP group as well
   as the LAG to which the large flow is being directed.

5.4. Information for Redistribution of Small Flows

   Redistribution of small flows is done using the following:

     .  For LAG: The LAG ID and the component link IDs along with the
        relative weight of traffic to be assigned to each component link
        ID are required.

     .  For ECMP: The ECMP group and the ECMP Nexthop along with the
        relative weight of traffic to be assigned to each ECMP Nexthop
        are required.

   It is possible to have an ECMP nexthop that itself comprises a LAG.
   In that case, we would have to specify the new weights for both the
   ECMP nexthops within the ECMP group as well as the component links
   within the LAG.

   In the case where an ECMP component link itself comprises a LAG, we
   would have to specify new weights for both the component links within
   the ECMP group as well as the component links within the LAG.

5.5. Export of Flow Information

   Exporting large flow information is required when large flow
   recognition is being done on a router, but the decision to rebalance
   is being made in an external management station. Large flow
   information includes flow identification and the component link ID
   that the flow currently is assigned to. Other information such as
   flow QoS and bandwidth may be exported too.

   The IPFIX information model [RFC 7012] can be leveraged for large
   flow identification.

5.6. Monitoring information

5.6.1. Interface (link) utilization

   The incoming bytes (ifInOctets), outgoing bytes (ifOutOctets) and
   interface speed (ifSpeed) can be measured from the Interface table
   (iftable) MIB [RFC 1213].

   The link utilization can then be computed as follows:

   Incoming link utilization = (ifInOctets/8) / ifSpeed

   Outgoing link utilization = (ifOutOctets/8) / ifSpeed

   For high speed Ethernet links, the etherStatsHighCapacityTable MIB
   [RFC 3273] can be used.

   For scalability, it is recommended to use the counter push mechanism
   in [sflow-v5] for the interface counters.  Doing so would help avoid
   counter polling through the MIB interface.

   The outgoing link utilization of the component links within a
   LAG/ECMP group can be used to compute the imbalance (See Section 5.1)
   for the LAG/ECMP group.

5.6.2. Other monitoring information

   Additional monitoring information that is useful includes:

     .  Number of times rebalancing was done.

     .  Time since the last rebalancing event.

     .  The number of large flows currently rebalanced by the scheme.

     .  A list of the large flows that have been rebalanced including

          o the rate of each large flow at the time of the last
             rebalancing for that flow,

          o the time that rebalancing was last performed for the given
             large flow, and

          o the interfaces that the large flows was (re)directed to.

     .  The settings for the weights of the interfaces within a
        LAG/ECMP used by the small flows which depend on hashing.

6. Operational Considerations

6.1. Rebalancing Frequency

   Flows should be rebalanced only when the imbalance in the utilization
   across component links exceeds a certain threshold.  Frequent
   rebalancing to achieve precise equitable utilization across component
   links could be counter-productive as it may result in moving flows
   back and forth between the component links impacting packet ordering
   and system stability.  This applies regardless of whether large flows
   or small flows are redistributed.  It should be noted that reordering
   is a concern for TCP flows with even a few packets because three out-
   of-order packets would trigger sufficient duplicate ACKs to the
   sender resulting in a retransmission [RFC 5681].

   The operator would have to experiment with various values of the
   large flow recognition parameters (minimum bandwidth threshold,
   observation interval) and the imbalance threshold across component
   links to tune the solution for their environment.

6.2. Handling Route Changes

   Large flow rebalancing must be aware of any changes to the FIB.  In
   cases where the nexthop of a route no longer to points to the LAG, or
   to an ECMP group, any PBR entries added as described in Section 4.4.1
   and 4.4.2 must be withdrawn in order to avoid the creation of
   forwarding loops.

6.3. Forwarding Resources

   Hash-based techniques used for load balancing with LAG/ECMP are
   usually stateless.  The mechanisms described in this document require
   additional resources in the forwarding plane of routers for creating
   PBR rules that are capable of overriding the forwarding decision from
   the hash-based approach.  These resources may limit the number of
   flows that can be rebalanced and may also impact the latency
   experienced by packets due to the additional lookups that are

7. IANA Considerations

   This memo includes no request to IANA.

8. Security Considerations

   This document does not directly impact the security of the Internet
   infrastructure or its applications. In fact, it could help if there
   is a DOS attack pattern which causes a hash imbalance resulting in
   heavy overloading of large flows to certain LAG/ECMP component

   An attacker with knowledge of the large flow recognition algorithm
   and any stateless distribution method can generate flows that are
   distributed in a way that overloads a specific path.  This could be
   used to cause the creation of PBR rules that exhaust the available
   rule capacity on nodes.  If PBR rules are consequently discarded,
   this could result in congestion on the attacker-selected path.
   Alternatively, tracking large numbers of PBR rules could result in
   performance degradation.

9. Contributing Authors

   Sanjay Khanna
   Cisco Systems
   Email: sanjakha@gmail.com

10. Acknowledgements

   The authors would like to thank the following individuals for their
   review and valuable feedback on earlier versions of this document:
   Shane Amante, Fred Baker, Michael Bugenhagen, Zhen Cao, Brian
   Carpenter, Benoit Claise, Michael Fargano, Wes George, Sriganesh
   Kini, Roman Krzanowski, Andrew Malis, Dave McDysan, Pete Moyer,
   Peter Phaal, Dan Romascanu, Curtis Villamizar, Jianrong Wong, George
   Yum, and Weifeng Zhang.  As a part of the IETF Last Call process,
   valuable comments were received from Martin Thomson, Thomson and Carlos

11. References

11.1. Normative References

   [802.1AX] IEEE Standards Association, "IEEE Std 802.1AX-2008 IEEE
   Standard for Local and Metropolitan Area Networks - Link
   Aggregation", 2008.

   [RFC 2991] Thaler, D. and C. Hopps, "Multipath Issues in Unicast and
   Multicast," November 2000.

   [RFC 7011] Claise, B. et al., "Specification of the IP Flow
   Information Export (IPFIX) Protocol for the Exchange of IP Traffic
   Flow Information," September 2013.

   [RFC 7012] Claise, B. and B. Trammell, "Information Model for IP Flow
   Information Export (IPFIX)," September 2013.

   [sFlow-v5] Phaal, P. and M. Lavine, "sFlow version 5,"
   http://www.sflow.org/sflow_version_5.txt, July 2004.

11.2. Informative References

   [bin-pack] Coffman, Jr., E., M. Garey, and D. Johnson. Approximation
   Algorithms for Bin-Packing -- An Updated Survey. In Algorithm Design
   for Computer System Design, ed. by Ausiello, Lucertini, and Serafini.
   Springer-Verlag, 1984.

   [CAIDA] "Caida Internet Traffic Analysis," http://www.caida.org/home.

   [DevoFlow] Mogul, J., et al., "DevoFlow: Cost-Effective Flow
   Management for High Performance Enterprise Networks," Proceedings of
   the ACM SIGCOMM, August 2011.

   [FLOW-ACC] Zseby, T., et al., "Packet sampling for flow accounting:
   challenges and limitations," Proceedings of the 9th international
   conference on Passive and active network measurement, 2008.

   [ID.ietf-rtgwg-cl-requirement] Villamizar, C. et al., "Requirements
   for MPLS over a Composite Link," September 2013.

   [ITCOM] Jo, J., et al., "Internet traffic load balancing using
   dynamic hashing with flow volume," SPIE ITCOM, 2002.

   [NDTM] Estan, C. and G. Varghese, "New directions in traffic
   measurement and accounting," Proceedings of ACM SIGCOMM, August 2002.

   [NVGRE] Sridharan, M. et al., "NVGRE: Network Virtualization using
   Generic Routing Encapsulation," draft-sridharan-virtualization-
   nvgre-04, February 2014.

   [RFC 2784] Farinacci, D. et al., "Generic Routing Encapsulation
   (GRE)," March 2000.

   [RFC 6790] Kompella, K. et al., "The Use of Entropy Labels in MPLS
   Forwarding," November 2012.

   [RFC 1213] McCloghrie, K., "Management Information Base for Network
   Management of TCP/IP-based internets: MIB-II," March 1991.

   [RFC 2992] Hopps, C., "Analysis of an Equal-Cost Multi-Path
   Algorithm," November 2000.

   [RFC 3273] Waldbusser, S., "Remote Network Monitoring Management
   Information Base for High Capacity Networks," July 2002.

   [RFC 3931] Lau, J. (Ed.), M. Townsley (Ed.), and I. Goyret (Ed.),
   "Layer 2 Tunneling Protocol - Version 3," March 2005.

   [RFC 3954] Claise, B., "Cisco Systems NetFlow Services Export Version
   9," October 2004.

   [RFC 5470] G. Sadasivan et al., "Architecture for IP Flow Information
   Export," March 2009.

   [RFC 5475] Zseby, T. et al., "Sampling and Filtering Techniques for
   IP Packet Selection," March 2009.

   [RFC 5640] Filsfils, C., P. Mohapatra, and C. Pignataro, "Load
   Balancing for Mesh Softwires," August 2009.

   [RFC 5681] Allman, M. et al., "TCP Congestion Control," September

   [SAMP-BASIC] Phaal, P. and S. Panchen, "Packet Sampling Basics,"

   [sFlow-LAG] Phaal, P. and A. Ghanwani, "sFlow LAG counters
   structure," http://www.sflow.org/sflow_lag.txt, September 2012.

   [STT] Davie, B. (Ed.) and J. Gross, "A Stateless Transport Tunneling
   Protocol for Network Virtualization (STT)," draft-davie-stt-06, March

   [VXLAN] Mahalingam, M. et al., "VXLAN: A Framework for Overlaying
   Virtualized Layer 2 Networks over Layer 3 Networks," draft-
   mahalingam-dutt-dcops-vxlan-09, April 2014.

   [YONG] Yong, L., "Enhanced ECMP and Large Flow Aware Transport,"
   draft-yong-pwe3-enhance-ecmp-lfat-01, September 2010.

   Appendix A. Internet Traffic Analysis and Load Balancing Simulation

   Internet traffic [CAIDA] has been analyzed to obtain flow statistics
   such as the number of packets in a flow and the flow duration. The
   five tuples in the packet header (IP addresses, TCP/UDP Ports, and IP
   protocol) are used for flow identification. The analysis indicates
   that < ~2% of the flows take ~30% of total traffic volume while the
   rest of the flows (> ~98%) contributes ~70% [YONG].

   The simulation has shown that given Internet traffic pattern, the
   hash-based technique does not evenly distribute the flows over ECMP
   paths. Some paths may be > 90% loaded while others are < 40% loaded.
   The more ECMP paths exist, the more severe the misbalancing. This
   implies that hash-based distribution can cause some paths to become
   congested while other paths are underutilized [YONG].

   The simulation also shows substantial improvement by using the large
   flow-aware hash-based distribution technique described in this
   document. In using the same simulated traffic, the improved
   rebalancing can achieve < 10% load differences among the paths. It
   proves how large flow-aware hash-based distribution can effectively
   compensate the uneven load balancing caused by hashing and the
   traffic characteristics [YONG].

Authors' Addresses

   Ram Krishnan
   Brocade Communications
   San Jose, 95134, USA
   Phone: +1-408-406-7890
   Email: ramkri123@gmail.com

   Lucy Yong
   Huawei USA
   5340 Legacy Drive
   Plano, TX 75025, USA
   Phone: +1-469-277-5837
   Email: lucy.yong@huawei.com

   Anoop Ghanwani
   San Jose, CA 95134
   Phone: +1-408-571-3228
   Email: anoop@alumni.duke.edu

   Ning So
   Tata Communications
   Plano, TX 75082, USA
   Phone: +1-972-955-0914
   Email: ning.so@tatacommunications.com

   Bhumip Khasnabish
   ZTE Corporation
   New Jersey, 07960, USA
   Phone: +1-781-752-8003
   Email: vumip1@gmail.com