| Internet-Draft | EVPN Weighted Multi-Pathing | May 2021 |
| Malhotra, et al. | Expires 11 November 2021 | [Page] |
EVPN enables all-active multi-homing for a CE device connected to two or more PEs via a LAG, such that bridged and routed traffic from remote PEs to hosts attached to the Ethernet Segment can be equally load balanced (it uses Equal Cost Multi Path) across the multi-homing PEs. EVPN also enables multi-homing for IP subnets advertised in IP Prefix routes, so that routed traffic from remote PEs to those IP subnets can be load balanced. This document defines extensions to EVPN procedures to optimally handle unequal access bandwidth distribution across a set of multi-homing PEs in order to:¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
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."¶
This Internet-Draft will expire on 11 November 2021.¶
Copyright (c) 2021 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 (https://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.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].¶
"Local PE" in the context of an ESI refers to a provider edge switch OR router that physically hosts the ESI.¶
"Remote PE" in the context of an ESI refers to a provider edge switch OR router in an EVPN overlay, whose overlay reachability to the ESI is via the Local PE.¶
In an EVPN-IRB based network overlay, with a CE multi-homed via a EVPN all-active multi-homing, bridged and routed traffic from remote PEs can be equally load balanced (ECMPed) across the multi-homing PEs:¶
All of the above load balancing and DF election procedures implicitly assume equal bandwidth distribution between the CE and the set of multi-homing PEs. Essentially, with this assumption of equal "access" bandwidth distribution across all PEs, ALL remote traffic is equally load balanced across the multi-homing PEs. This assumption of equal access bandwidth distribution can be restrictive with respect to adding / removing links in a multi-homed LAG interface and may also be easily broken on individual link failures. A solution to handle unequal access bandwidth distribution across a set of multi-homing EVPN PEs is proposed in this document. Primary motivation behind this proposal is to enable greater flexibility with respect to adding / removing member PE-CE links, as needed and to optimally handle PE-CE link failures.¶
+------------------------+
| Underlay Network Fabric|
+------------------------+
+-----+ +-----+
| PE1 | | PE2 |
+-----+ +-----+
\ /
\ ESI-1 /
\ /
+\---/+
| \ / |
+--+--+
|
CE1
Consider CE1 that is dual-homed to PE1 and PE2 via EVPN all-active multi-homing with single member links of equal bandwidth to each PE (aka, equal access bandwidth distribution across PE1 and PE2). If the provider wants to increase link bandwidth to CE1, it must add a link to both PE1 and PE2 in order to maintain equal access bandwidth distribution and inter-work with EVPN ECMP load balancing. In other words, for a dual-homed CE, total number of CE links must be provisioned in multiples of 2 (2, 4, 6, and so on). For a triple-homed CE, number of CE links must be provisioned in multiples of three (3, 6, 9, and so on). To generalize, for a CE that is multi-homed to "n" PEs, number of PE-CE physical links provisioned must be an integral multiple of "n". This is restrictive in case of dual-homing and very quickly becomes prohibitive in case of multi-homing.¶
Instead, a provider may wish to increase PE-CE bandwidth OR number of links in any link increments. As an example, for CE1 dual-homed to PE1 and PE2 in all-active mode, provider may wish to add a third link to only PE1 to increase total bandwidth for this CE by 50%, rather than being required to increase access bandwidth by 100% by adding a link to each of the two PEs. While existing EVPN based all-active load balancing procedures do not necessarily preclude such asymmetric access bandwidth distribution among the PEs providing redundancy, it may result in unexpected traffic loss due to congestion in the access interface towards CE. This traffic loss is due to the fact that PE1 and PE2 will continue to be treated as equal cost paths at remote PEs, and as a result may attract approximately equal amount of CE1 destined traffic, even when PE2 only has half the bandwidth to CE1 as PE1. This may lead to congestion and traffic loss on the PE2-CE1 link. If bandwidth distribution to CE1 across PE1 and PE2 is 2:1, traffic from remote hosts must also be load balanced across PE1 and PE2 in 2:1 manner.¶
More importantly, unequal PE-CE bandwidth distribution described above may occur during regular operation following a link failure, even when PE-CE links were provisioned to provide equal bandwidth distribution across multi-homing PEs.¶
+------------------------+
| Underlay Network Fabric|
+------------------------+
+-----+ +-----+
| PE1 | | PE2 |
+-----+ +-----+
\\ //
\\ ESI-1 //
\\ /X
+\\---//+
| \\ // |
+---+---+
|
CE1
Consider a CE1 that is multi-homed to PE1 and PE2 via a LAG with two member links to each PE. On a PE2-CE1 physical link failure, LAG represented by an Ethernet Segment ESI-1 on PE2 stays up, however, its bandwidth is cut in half. With existing ECMP procedures, both PE1 and PE2 may continue to attract equal amount of traffic from remote PEs, even when PE1 has double the bandwidth to CE1. If bandwidth distribution to CE1 across PE1 and PE2 is 2:1, traffic from remote hosts must also be load balanced across PE1 and PE2 in 2:1 manner to avoid unexpected congestion and traffic loss on PE2-CE1 links within the LAG. As an alternative, min-link on LAGs is sometimes used to bring down the LAG interface on member link failures. This however results in loss of available bandwidth in the network, and is not ideal.¶
+-----------------------+
|Underlay Network Fabric|
+-----------------------+
+-----+ +-----+ +-----+ +-----+
| PE1 | | PE2 | ..... | PEx | | PEn |
+-----+ +-----+ +-----+ +-----+
\ \ // //
\ L1 \ L2 // Lx // Ln
\ \ // //
+-\-------\-----------//--------//-+
| \ \ ESI-1 // // |
+----------------------------------+
|
CE
To generalize, if total link bandwidth to a CE is distributed across "n" multi-homing PEs, with Lx being the total bandwidth to PEx across all links, traffic from remote PEs to this CE must be load balanced unequally across [PE1, PE2, ....., PEn] such that, fraction of total unicast and BUM flows destined for CE that are serviced by PEx is:¶
Lx / [L1+L2+.....+Ln]¶
Figure 3 illustrates a scenario where PE1..PEn are attached to a multi-homed Ethernet Segment, however this document generalizes this requirement so that the unequal load balancing can be applied to PEs attached to a vES or to a multi-homed subnet advertised by EVPN IP Prefix routes.¶
The solution proposed below includes extensions to EVPN procedures to achieve the above.¶
In order to achieve weighted load balancing to an ES or vES for overlay unicast traffic, Ethernet A-D per ES route (EVPN Route Type 1) is leveraged to signal the Ethernet Segment weight to remote PEs. Using Ethernet A-D per-ES route to signal the Ethernet Segment weight provides a mechanism that reacts to changes in access bandwidth or number of access links in a service and host independent manner. Remote PEs computing the MAC path-lists based on global and aliasing Ethernet A-D routes now have the ability to setup weighted load balancing path-lists based on the ESI access bandwidth or number of links received from each PE that the ES is multi-homed to.¶
In order to achieve weighted load balancing of overlay BUM traffic, EVPN ES route (Route Type 4) is leveraged to signal the ESI weight to PEs within an ESI's redundancy group to influence per-service DF election. PEs in an ESI redundancy group now have the ability to do service carving in proportion to each PE's relative ESI weight.¶
Unequal load balancing to multi-homed subnets is achieved by signaling the weight along with the IP Prefix routes advertised for the subnet.¶
Procedures to accomplish this are described in greater detail next.¶
A new EVPN Link Bandwidth extended community is defined for the solution specified in this document:¶
EVPN Link Bandwidth Extended Community value field is used to carry total bandwidth of PE's all physical links in an ethernet segment, expressed in Mbits/sec (MegabitsPerSecond) represented as an unsigned integer. Note however that the load balancing algorithm defined in this document uses ratio of Link Bandwidths. Hence, the operator may choose a different unit or use the community as a generalized weight that may be set to link count, locally configured weight, or a value computed based on an attribute other than link bandwidth. In such case, the operator MUST ensure consistent usage of the unit across all PEs in an ethernet segment. This may involve multiple routing domains/Autonomous Systems.¶
In order to facilitate this, as well as avoid interop issues because of provisioning error, one octet in the extended community's six octet 'value' field is used to explicitly signal if the weight encoded in the remaining five octets is link bandwidth expressed in Mbits/sec or a generalized weight value. This results in the following encoding for EVPN link bandwidth extended community:¶
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Type | Sub-Type | Value-Units | | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + | Value-Weight | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Value-Units is encoded as:¶
Generalized weight units are intentionally left arbritrary to allow for flexibility in its usage for different applications without having to define new encoding for each non-default application. Implementations SHOULD support the default units of Mbits/sec, while support of non-default generalized weight is considered optional.¶
Link bandwidth extended community described in [BGP-LINK-BW] for layer 3 VPNs was considered for re-use here. This Link bandwidth extended community is however defined in [BGP-LINK-BW] as optional non-transitive. Since it is not possible to change deployed behavior of extended-community defined in [BGP-LINK-BW], it was decided to define a new one. In inter-AS scenarios, link-bandwidth needs to be signaled to eBGP neighbors. When signaled across AS boundary, this attribute can be used to achieve optimal load-balancing towards source PEs from a different AS. This is applicable both when next-hop is changed or unchanged across AS boundaries.¶
A PE that is part of an Ethernet Segment's redundancy group would advertise an additional "EVPN link bandwidth" extended community attribute with Ethernet A-D per-ES route (EVPN Route Type 1), that carries total bandwidth of PE's physical links in an Ethernet Segment. New EVPN link bandwidth extended community defined in this document is used for this purpose.¶
A receiving PE MUST check for consistent 'Value-Units' received in the EVPN link bandwidth exteneded community from each remote PE in an Ethernet Segment. In case of any inconsistency in 'Value-Units' across PEs in an Ethernet Segment, this attribute is to be ignored and remote PEs are to follow regular ECMP forwarding to that Ethernet Segment. Once consistency of 'Value-Units' is validated, receiving PE SHOULD use the 'Value-Weight' received from each PE to compute a relative (normalized) weight for each remote PE, per-ES, and then use this relative weight to compute a weighted path-list to be used for load balancing, as opposed to using an ECMP path-list for load balancing across the PE paths. PE Weight and resulting weighted path-list computation at remote PEs is a local matter. An example computation algorithm is shown below to illustrate the idea:¶
if,¶
L(x,y) : link bandwidth advertised by PE-x for ESI-y¶
W(x,y) : normalized weight assigned to PE-x for ESI-y¶
H(y) : Highest Common Factor (HCF) of [L(1,y), L(2,y), ....., L(n,y)]¶
then, the normalized weight assigned to PE-x for ESI-y may be computed as follows:¶
W(x,y) = L(x,y) / H(y)¶
For a MAC+IP route (EVPN Route Type 2) received with ESI-y, receiving PE may compute MAC and IP forwarding path-list weighted by the above normalized weights.¶
As an example, for a CE dual-homed to PE-1, PE-2, PE-3 via 2, 1, and 1 GE physical links respectively, as part of a LAG represented by ESI-10:¶
L(1, 10) = 2000 Mbits/sec¶
L(2, 10) = 1000 Mbits/sec¶
L(3, 10) = 1000 Mbits/sec¶
H(10) = 1000¶
Normalized weights assigned to each PE for ESI-10 are as follows:¶
W(1, 10) = 2000 / 1000 = 2.¶
W(2, 10) = 1000 / 1000 = 1.¶
W(3, 10) = 1000 / 1000 = 1.¶
For a remote MAC+IP host route received with ESI-10, forwarding load balancing path-list may now be computed as: [PE-1, PE-1, PE-2, PE-3] instead of [PE-1, PE-2, PE-3]. This now results in load balancing of all traffic destined for ESI-10 across the three multi-homing PEs in proportion to ESI-10 bandwidth at each PE.¶
Weighted path-list computation must only be done for an ESI if a link bandwidth attribute is received from all of the PE's advertising reachability to that ESI via Ethernet A-D per-ES Route Type 1. In an unlikely event that link bandwidth attribute is not received from one or more subset of PEs, forwarding path-list should be computed using regular ECMP semantics. Note that a default weight cannot be assumed for a PE that does not advertise its link bandwidth as the weight attribute t be used in path-list computation is relative.¶
Optionally, load sharing of per-service DF role, weighted by individual PE's link-bandwidth share within a multi-homed ES may also be achieved.¶
In order to do that, a new DF Election Capability [RFC8584] called "BW" (Bandwidth Weighted DF Election) is defined. BW MAY be used along with some DF Election Types, as described in the following sections.¶
[RFC8584] defines a new extended community for PEs within a redundancy group to signal and agree on uniform DF Election Type and Capabilities for each ES. This document requests IANA for a bit in the DF Election extended community Bitmap:¶
Bit 28: BW (Bandwidth Weighted DF Election)¶
ES routes advertised with the BW bit set will indicate the desire of the advertising PE to consider the link-bandwidth in the DF Election algorithm defined by the value in the "DF Type".¶
As per [RFC8584], all the PEs in the ES MUST advertise the same Capabilities and DF Type, otherwise the PEs will fall back to Default [RFC7432] DF Election procedure.¶
The BW Capability MAY be advertised with the following DF Types:¶
The following sections describe how the DF Election procedures are modified for the above DF Types when the BW Capability is used.¶
When all the PEs in the Ethernet Segment (ES) agree to use the BW Capability with DF Type 0, the Default DF Election procedure as defined in [RFC7432] is modified as follows:¶
Considering the same example as in Section 3, the candidate PE list for DF election is:¶
[PE-1, PE-1, PE-2, PE-3].¶
The DF for a given VLAN-a on ES-10 is now computed as (VLAN-a % 4). This would result in the DF role being distributed across PE1, PE2, and PE3 in portion to each PE's normalized weight for ES-10.¶
[RFC8584] introduces Highest Random Weight (HRW) algorithm (DF Type 1) for DF election in order to solve potential DF election skew depending on Ethernet tag space distribution. [EVPN-PER-MCAST-FLOW-DF] further extends HRW algorithm for per-multicast flow based hash computations (DF Type 4). This section describes extensions to HRW Algorithm for EVPN DF Election specified in [RFC8584] and in [EVPN-PER-MCAST-FLOW-DF] in order to achieve DF election distribution that is weighted by link bandwidth.¶
A new variable called "bandwidth increment" is computed for each [PE, ES] advertising the ES link bandwidth attribute as follows:¶
In the context of an ES,¶
L(i) = Link bandwidth advertised by PE(i) for this ES¶
L(min) = lowest link bandwidth advertised across all PEs for this ES¶
Bandwidth increment, "b(i)" for a given PE(i) advertising a link bandwidth of L(i) is defined as an integer value computed as:¶
b(i) = L(i) / L(min)¶
As an example,¶
with PE(1) = 10, PE(2) = 10, PE(3) = 20¶
bandwidth increment for each PE would be computed as:¶
b(1) = 1, b(2) = 1, b(3) = 2¶
with PE(1) = 10, PE(2) = 10, PE(3) = 10¶
bandwidth increment for each PE would be computed as:¶
b(1) = 1, b(2) = 1, b(3) = 1¶
Note that the bandwidth increment must always be an integer, including, in an unlikely scenario of a PE's link bandwidth not being an exact multiple of L(min). If it computes to a non-integer value (including as a result of link failure), it MUST be rounded down to an integer.¶
HRW algorithm as described in [RFC8584] and in [EVPN-PER-MCAST-FLOW-DF] computes a random hash value for each PE(i), where, (0 < i <= N), PE(i) is the PE at ordinal i, and Address(i) is the IP address of PE(i).¶
For 'N' PEs sharing an Ethernet segment, this results in 'N' candidate hash computations. The PE that has the highest hash value is selected as the DF.¶
We refer to this hash value as "affinity" in this document. Hash or affinity computation for each PE(i) is extended to be computed one per bandwidth increment associated with PE(i) instead of a single affinity computation per PE(i).¶
PE(i) with b(i) = j, results in j affinity computations:¶
affinity(i, x), where 1 < x <= j¶
This essentially results in number of candidate HRW hash computations for each PE that is directly proportional to that PE's relative bandwidth within an ES and hence gives PE(i) a probability of being DF in proportion to it's relative bandwidth within an ES.¶
As an example, consider an ES that is multi-homed to two PEs, PE1 and PE2, with equal bandwidth distribution across PE1 and PE2. This would result in a total of two candidate hash computations:¶
affinity(PE1, 1)¶
affinity(PE2, 1)¶
Now, consider a scenario with PE1's link bandwidth as 2x that of PE2. This would result in a total of three candidate hash computations to be used for DF election:¶
affinity(PE1, 1)¶
affinity(PE1, 2)¶
affinity(PE2, 1)¶
which would give PE1 2/3 probability of getting elected as a DF, in proportion to its relative bandwidth in the ES.¶
Depending on the chosen HRW hash function, affinity function MUST be extended to include bandwidth increment in the computation.¶
For e.g.,¶
affinity function specified in [EVPN-PER-MCAST-FLOW-DF] MAY be extended as follows to incorporate bandwidth increment j:¶
affinity(S,G,V, ESI, Address(i,j)) = (1103515245.((1103515245.Address(i).j + 12345) XOR D(S,G,V,ESI))+12345) (mod 2^31)¶
affinity or random function specified in [RFC8584] MAY be extended as follows to incorporate bandwidth increment j:¶
affinity(v, Es, Address(i,j)) = (1103515245((1103515245.Address(i).j + 12345) XOR D(v,Es))+12345)(mod 2^31)¶
This section applies to ES'es where all the PEs in the ES agree use the BW Capability with DF Type 2. The BW Capability modifies the Preference DF Election procedure [EVPN-DF-PREF], by adding the LBW value as a tie-breaker as follows:¶
Section 4.1, bullet (f) in [EVPN-DF-PREF] now considers the LBW value:¶
f) In case of equal Preference in two or more PEs in the ES, the tie-breakers will be the DP bit, the LBW value and the lowest IP PE in that order. For instance:¶
While incorporating link bandwidth into the DF election process provides optimal BUM traffic distribution across the ES links, it also implies that DF elections are re-adjusted on link failures or bandwidth changes. If the operator does not wish to have this level of churn in their DF election, then they should not advertise the BW capability. Not advertising BW capability may result in less than optimal BUM traffic distribution while still retaining the ability to allow a remote ingress PE to do weighted ECMP for its unicast traffic to a set of multi-homed PEs.¶
PE-CE link bandwidth availability may sometimes vary in real-time disproportionately across PE-CE links within a multi-homed ESI due to various factors such as flow based hashing combined with fat flows and unbalanced hashing. Reacting to real-time available bandwidth is at this time outside the scope of this document.¶
EVPN Link bandwidth extended community may also be used to achieve unequal load-balancing of prefix routed traffic by including this extended community in EVPN Route Type 5. When included in EVPN RT-5, its value is to be interpreted as advertising PE's relative weight for the prefix included in this RT-5. Receiving PE will then compute the forwarding path-list for the prefix route using weighted paths received from each remote PE.¶
EVPN-LAG based multi-homing on an IRB gateway may also be deployed together with non-EVPN routing, such as global routing or an L3VPN routing control plane. Key property that differentiates this set of use cases from EVPN IRB use cases discussed earlier is that EVPN control plane is used only to enable LAG interface based multi-homing and NOT as an overlay VPN control plane. EVPN control plane in this case enables:¶
Applicability of weighted ECMP procedures proposed in this document to these set of use cases is an area of further consideration.¶
This document raises no new security issues for EVPN.¶
[RFC8584] defines a new extended community for PEs within a redundancy group to signal and agree on uniform DF Election Type and Capabilities for each ES. This document requests IANA for a bit in the DF Election extended community Bitmap:¶
Bit 28: BW (Bandwidth Weighted DF Election)¶
A new EVPN Link Bandwidth extended community is defined to signal local ES link bandwidth to remote PEs. This extended-community is defined of type 0x06 (EVPN). IANA is requested to assign a sub-type value of 0x10 for the EVPN Link bandwidth extended community, of type 0x06 (EVPN). EVPN Link Bandwidth extended community is defined as transitive.¶
IANA is requested to set up a registry called "Value-Units" for the 1-octet field in the EVPN Link Bandwidth Extended Community. New registrations will be made through the "RFC Required" procedure defined in [RFC8126]. The following initial values in that registry exist:¶
Value Name Reference ---- ---------------- ------------- 0 Weight in units of Mbps This document 1 Generalized Weight This document 2-255 Unassigned¶
Authors would like to thank Satya Mohanty for valuable review and inputs with respect to HRW and weighted HRW algorithm refinements proposed in this document. Authors would also like to thank Bruno Decraene and Sergey Fomin for valuable review and comments.¶
Satya Ranjan Mohanty Cisco Systems US Email: satyamoh@cisco.com¶