How MPLS handles ECMP:
source: http://tools.ietf.org/html/draft-ietf-mpls-ecmp-bcp-01
Current EMCP Practices [as of 2005]
The MPLS label stack and Forwarding Equivalence Classes are defined
in [RFC3031]. The MPLS label stack does not carry a Protocol Identi-
fier. Instead the payload of an MPLS packet is identified by the
Forwarding Equivalence Class (FEC) of the bottom most label. Thus it
is not possible to know the payload type if one does not know the
label binding for the bottom most label. Since an LSR which is pro-
cessing a label stack need only know the binding for the label(s) it
must process, it is very often the case that LSRs along an LSP are
unable to determine the payload type of the carried contents.
As a means of potentially reducing delay and congestion, IP networks
have taken advantage of multiple paths through a network by splitting
traffic flows across those paths. The general name for this practice
is Equal Cost Multipath or ECMP. In general this is done by hashing
on various fields on the IP or contained headers. In practice,
within a network core, the hashing in based mainly or exclusively on
the IP source and destination addresses. The reason for splitting
aggregated flows in this manner is to minimize the re-ordering of
packets belonging to individual flows contained within the aggregated
flow. Within this document we use the term IP ECMP for this type of
forwarding algorithm.
In the early days of MPLS, the payload was almost exclusively IP.
Even today the overwhelming majority of carried traffic remains IP.
Providers of MPLS equipment sought to continue this IP ECMP behavior.
As shown above, it is not possible to know whether the payload of an
MPLS packet is IP at every place where IP ECMP needs to be performed.
Thus vendors have taken the liberty of guessing what the payload is.
By inspecting the first nibble beyond the label stack, it can be
inferred that a packet is not IPv4 or IPv6 if the value of the nibble
(where the IP version number would be found) is not 0x4 or 0x6
respectively. Most deployed LSRs will treat a packet whose first
nibble is equal to 0x4 as if the payload were IPv4 for purposes of IP
ECMP.
A consequence of this is that any application which defines a FEC
which does not take measures to prevent the values 0x4 and 0x6 from
occurring in the first nibble of the payload may be subject to IP
ECMP and thus having their flows take multiple paths and arriving
with considerable jitter and possibly out of order. While none of
this is in violation of the basic service offering of IP, it is
detrimental to the performance of various classes of applications.
It also complicates the measurement, monitoring and tracing of those
flows.
New MPLS payload types are emerging such as those specified by the
IETF PWE3 and AVT working groups. These payloads are not IP and, if
specified without constraint might be mistaken for IP.
It must also be noted that LSRs which correctly identify a payload as
not being IP, may still need to load-share this traffic across multi-
ple equal-cost paths. In this case a LABEL ECMP algorithm is
employed, where a hash is computed on all or part(s) of the label
stack. Any reserved label, no matter where it is located in the
stack, may be included in the computation for load balancing. Modi-
fication of the label stack between packets of a single flow could
result in re-ordering that flow. That is, were an explicit null or a
router-alert label to be added to a packet, that packet could take a
different path through the network.
Note that for some applications, being mistaken for IPv4 may not be
detrimental. The trivial case where the payload behind the top label
is a packet belonging to an MPLS IPv4 VPN. Here the real payload is
IP and most (if not all) deployed equipment will locate the end of
the label stack and correctly perform IP ECMP.
A less obvious case is when the packets of a given flow happen to
have constant values in the fields upon which IP ECMP would be per-
formed. For example if an ethernet frame immediately follows the
label and the LSR does not do ECMP on IPv6, then either the first
nibble will be 0x4 or it will be something else. If the nibble is
not 0x4 then no IP ECMP is performed, but Label ECMP may be per-
formed. If it is 0x4, then the constant values of the MAC addresses
overlay the fields that would be occupied by the source and destina-
tion addresses of an IP header.
