Internet Engineering Task Force Curtis Villamizar
INTERNET-DRAFT ANS
draft-ietf-ospf-omp-00 March 13, 1998
OSPF Optimized Multipath (OSPF-OMP)
Status of this Memo
This document is an Internet-Draft. 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-Drafts.
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.''
To view the entire list of current Internet-Drafts, please check the
``1id-abstracts.txt'' listing contained in the Internet-Drafts Shadow
Directories on ftp.is.co.za (Africa), ftp.nordu.net (Europe),
munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or
ftp.isi.edu (US West Coast).
Abstract
OSPF may form multiple equal cost paths between points. This is true
of any link state protocol. In the absense of any explicit support to
take advantage of this, a path may be chosen arbitrarily. Techniques
have been utilized to divide traffic somewhat evenly among the
available paths. These techniques have been referred to as Equal Cost
Multipath (ECMP). An unequal division of traffic among the available
paths is generally preferable. Routers generally have no knowledge of
traffic loading on distant links and therefore have no basis to
optimize the allocation of traffic.
Optimized Mulitpath is a compatible extension to OSPF, utilizing the
Opaque LSA to distribute loading information, proposing a means to
adjust forwarding, and providing an algorithm to make the adjustments
gradually enough to insure stability yet provide reasonably fast
adjustment when needed.
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
1 Overview
Networks running OSPF are often heavily loaded. Topologies often
evolve to include multiple paths. Multiple paths may be initially
designed to provide redundancy but also result from incremental
addition of circuits to accomodate traffic growth. The redundant
paths provide a potential to distribute traffic loading and reduce
congestion. Optimized Mulitpath (OMP) provides a means for OSPF to
make better use of this potential to distribute loading.
Early attempts to provide load sensitive routing involved changing
link costs according to loading. These attempts were doomed to
failure because the adjustment increment was grossly course and
oscillation was inevitable [?].
A widely utilized technique to improve loading is known as Equal Cost
Multipath (ECMP). If the topology is such that equal cost paths exist,
then an attempt is made to divide traffic equally among the paths.
Three methods of splitting traffic have been used.
1. Per packet round robin forwarding.
2. Dividing destination prefixes among available next hops in the
forwarding entries.
3. Dividing traffic according to a hash function applied to the source
and desination pair.
The ``per packet round robin forwarding'' technique is only applicable
if the delays on the paths are almost equal. The delay difference
must be small relative to packet serialization time. Delay
differences greater than three times the packet serialization time can
cause terrible TCP performance. for example, packet 2, 4, and 6 may
arrive before packet 1, triggering TCP fast retransmit. The result
will be limiting TCP to a very small window and very poor performance
over long delay paths.
The delay differences must be quite small. A 532 byte packet is
serialized onto a DS1 link in under 2.8 msec. At DS3 speed,
serialization is accomplished in under 100 usec. At OC12 it is under
7 usec. For this reason ``per packet round robin forwarding'' is not
applicable to a high speed WAN.
Dividing destination prefixes among available next hops provides a
very course and unpredictable load split. Long prefixes are
problematic. In reaching an end node, the majority of traffic is
often destined to a single prefix. This technique is applicable to a
Villamizar Expires September 13, 1998 [Page 2]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
high speed WAN but with the drawbacks just mentioned better techniques
are needed.
The ``source/destination hash'' based technique was used as far back
as the T3-NSFNET in the NSS routers. A hash function, such as CRC-16,
is applied over the source address and destination address. The hash
space is then split evenly among the available paths by either setting
threshholds or performing a modulo operation. Traffic between any
given source and destination remain on the same path. Because the
technique is based on host addresses, and uses both the source and
destination address, it does not suffer the course granularity problem
of the prefix based technique, even when forwarding to a single
prefix. Source/destination hash is the best technique available for a
high speed WAN.
The forwarding decision for the ``source/destination hash'' based
technique is quite simple. When a packet arrives, look up the for-
warding entry in the radix tree. The next hop entry can be an array
index into a set of structures, each containing one or more actual
next hops. If more than one next hop is present, compute a CRC16
value based on the source and destination addresses. The CRC16 can be
implemented in hardware and computed in parallel to the radix tree
lookup in high speed implementations, and discarded if not needed.
Each next hop entry in the structure must contain a boundary value and
the next hop itself. An integer ``less than'' comparison is made
against the boundary value determining whether to use this next hop or move
to the next a comparison. In hardware the full set of comparisons can be
made simultaneously for up to some number of next hops. This yields the
next hop to use.
For ECMP, the boundary values are set by first dividing one more than
the maximum value that the hash computation can return (65536 for
CRC16) by the number of available next hops and then setting the Nth
boundary to N times that number (with the Nth value fixed at one more
than the maximum value regardless of underflow caused by trucating
during division, 65536 for CRC16).
An equal load split is not always optimal. Consider the example in
Figure 1 with the offered traffic in Table 1. If all of the link
costs are set equally, then the link N1---N3 is significantly
overloaded (135.75%) while the path N1---N2---N3 is lightly loaded
(45.25% and 22.62%). If the cost on the N1---N3 link is equal to the
cost of the N1---N2---N3 path, then N1 will try to split the load
destined toward N3 across the two paths.
Given the offered traffic in Table 1 the loading on N1---N3 is reduced
to 67.87% but the link loading on the path N2---N3 becomes 113.12%.
Ideally node N1 should put 1/3 of the traffic toward N3 on the path
N1---N2---N3 and 2/3 on the path N1---N3. To know to do this N1 must
know the loading on N2--N3.
Villamizar Expires September 13, 1998 [Page 3]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
.----.
/ \
| N2 |
\ /
`----'
// \\
// \\
// \\
.----. .----.
/ \ / \
| N1 | ======== | N3 |
\ / \ /
`----' `----'
Figure 1: A very simple application of ECMP
Nodes Traffic Node Names
n3-n1 60.000 Node 3 -> Node 1
n1-n3 60.000 Node 1 -> Node 3
n3-n2 20.000 Node 3 -> Node 2
n2-n3 20.000 Node 2 -> Node 3
n2-n1 10.000 Node 2 -> Node 1
n1-n2 10.000 Node 1 -> Node 2
Table 1: Traffic loading for the example in Figure 1
Villamizar Expires September 13, 1998 [Page 4]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
This is where Optimized Multipath (OMP) provides additional benefit
over ECMP. Ignoring for the moment how node N1 knows to put 1/3 of the
traffic toward N3 on the path N1---N2---N3, the way this is
accomplished from a forwarding standpoint is to move the boundary in
the forwarding structure from the default value of 1/2 of 65536 to
about 1/3 of 65536. If there are a very large set of source and
destination host addresses pairs, then the traffic will be split among
the 65536 possible hash values. This provides the means for a very
fine granularity of adjustment.
Having explained how a fine granularity of forwarding adjustment can
be accomplished, what remains is to define how nodes in a large topol-
ogy can know what the loading levels are elsewhere in the topology and
defining an algorithm which can allow autonomous unsyncronized
decisions on the parts of many routers in a topology to quickly
converge on a near optimal loading without the risk of oscillation.
2 Flooding Loading Information
Loading information is flooded within an OSPF area using Opaque LSAs
[1]. Area local scope (link-state type 10) link state attributes are
flooded containing an ``Opaque Type'' of LSA_OMP_LINK_LOAD or
LSA_OMP_PATH_LOAD. The type LSA_OMP_LINK_LOAD Opaque LSA is used to
flood link loading information within an area. The type
LSA_OMP_PATH_LOAD Opaque LSA is used to flood loading information for
use with inter-area routes. Loading information obtained from an
exterior routing protocol may also be considered if available. The
means of passing loading information in an exterior routing protocol
is beyond the scope of this document.
2.1 Link Loading Information
Within an area link loading is flooded using the type LSA_OMP_LINK_LOAD
Opaque LSA. The ``Opaque ID'' must contain a 24 bit integer that is
unique to the advertising router and link or virtual link. The method
of assignment of these 24 bit integers is a local matter. A router
must be capable of being configured to put a fixed value in N of the
bits and use the remainin 24-N bits to uniquely identify an interface.
The ``Opaque Information'' in the type LSA_OMP_LINK_LOAD Opaque LSA
contains the following.
1. a measure of link loading in each direction as a fraction of link
capacity,
Villamizar Expires September 13, 1998 [Page 5]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
2. a measure of packets dropped due to queue overflow in each
direction (if known) expressed as a fraction,
3. the link capacity in killobits per second (or unity if less than
1000 bits per second).
Generally the number of ouput packets dropped will be known. In
designs where drops occur on the input (generally a bad design
practice), the rate of input queue drops should be recorded. These
measures of loading and drop are computed using the interface counters
generally maintained for SNMP purposes, plus a running count of output
queue drops if available. The counters are sampled every
OMP_SAMPLE_INTERVAL seconds.
The previous value of each of the counters is substracted from the
current value. The counters to be sampled are the following.
1. bytes out,
2. bytes in,
3. packets out,
4. packets in,
5. output queue drops,
6. input queue drops.
A value of instantaneous load in each direction is based on byte count
and link capacity. An instantaneous output queue drop rate is based
on queue drops and packet count. Each of these is combined with a
running filtered value according to the following method. The running
total is shifted down by OMP_SHIFT_BITS bits and subtracted from the
running total. The instantaneous value is shifted down by
OMP_SHIFT_BITS bits and added to the running total.
The last time that a type LSA_OMP_LINK_LOAD Opaque LSA with the same
Opaque ID was sent is recorded and the values sent are recorded. For
the purpose of determining when to reflood, an equivalent loading
figure is used. The computation of equivalent loading is described in
Section 2.3.
The higher of the current equivalent loading computation and the
previous is used when determining whether to send the type
LSA_OMP_LINK_LOAD Opaque LSA. The type LSA_OMP_LINK_LOAD Opaque LSA is
sent if any of the following is true.
Villamizar Expires September 13, 1998 [Page 6]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
1. The equivalent load is over 100% and the change in equivalent load
since last resent is over 5% and 30 seconds has elapsed since last
sent.
2. The equivalent load is over 100% and the change in equivalent load
since last resent is over 2% and 90 seconds has elapsed since last
sent.
3. The equivalent load is over 100% and 3 minutes has elapsed since
last sent.
4. The equivalent load is over 90% and and the change in equivalent
load since last resent is over 5% and 1 minute has elapsed since
last sent.
5. The equivalent load is over 90% and the change in equivalent load
since last resent is over 2% and 4 minutes has elapsed since last
sent.
6. The equivalent load is over 90% and 10 minutes has elapsed since
last sent.
7. The equivalent load is over 70% and the change in equivalent load
since last resent is over 10% and 1 minute has elapsed since last
sent.
8. The equivalent load is over 70% and the change in equivalent load
since last resent is over 5% and 2 minutes has elapsed since last
sent.
9. The equivalent load is over 70% and the change in equivalent load
since last resent is over 2% and 8 minutes has elapsed since last
sent.
10. The equivalent load is over 70% and 15 minutes has elapsed since
last sent.
11. The equivalent load is over 50% and the change in equivalent load
since last resent is over 10% and 1 minute has elapsed since last
sent.
12. The equivalent load is over 50% and the change in equivalent load
since last resent is over 5% and 5 minutes has elapsed since last
sent.
13. The equivalent load is over 25% and the change in equivalent load
since last resent is over 25% and 2 minutes has elapsed since last
sent.
14. The equivalent load is over 25% and 20 minutes has elapsed since
last sent.
15. The type LSA_OMP_LINK_LOAD Opaque LSA has never been sent.
Villamizar Expires September 13, 1998 [Page 7]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
The point of this graduated scale is to reduce the amount of flooding
that is occurring unless links are in trouble or undergoing a
significant traffic shift. Change may occur in a quiescent network
due to failure external to the network that causes traffic to take
alternate paths. In this case, the more frequent flooding will
trigger a faster convergence. Traffic shift may also occur due to
shedding of loading by the OMP algortihm itself as the algorithm
converges in response to an external change.
2.2 Path Loading Information
Path loading information regarding an adjacent area is flooded by an
Area Border Router (ABR) using the type LSA_OMP_PATH_LOAD Opaque LSA.
The ``Opaque ID'' must contain a 24 bit integer that is unique to the
router and an advertised summary LSA. The method of assignment of
these 24 bit integers is a local matter. A router must be capable of
being configured to put a fixed value in N of the bits and use the
remainin 24-N bits to uniquely identify the summary LSA.
The ``Opaque Information'' in the type LSA_OMP_PATH_LOAD Opaque LSA
contains the following.
1. the highest loading in the direction toward the destination as a
fraction of link capacity,
2. a measure of total packet drop due to queue overflow in the
direction toward the destination expressed as a fraction,
3. the smallest link capacity on the path to the destination.
These values are taken from the link on the path from the ABR to the
destination of the summary LSA. The link with the highest loading may
not be the link with the lowest capacity. The queue drop value is one
minus the product of fraction of packets that are not dropped at each
measurement point on the path (input and output in the direction of
the path). The following computation is used.
path-loss = 1 - product((1 - link-loss-in) * (1 -
link-loss-out))
The path loading and path loss rate are filtered according to the same
algorithm defined in the prior section. Rather than polling a set of
counters the current value of the path loading and path loss rate is
Villamizar Expires September 13, 1998 [Page 8]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
used. An equivalent load is calculated for each path to a summary LSA
destination as described in Section 2.3. A type LSA_OMP_PATH_LOAD
Opaque LSA is flooded according to the same rate schedule as described
in the prior section.
An ABR may be configured to not send type LSA_OMP_PATH_LOAD Opaque LSA
into any given area.
2.3 Computing equivalent loading
The equivalent load is the actual percent loading multiplied by a
factor that provides an estimate of the extent to which TCP would be
slowing down to avoid congestion. This estimate is based on the link
bandwidth and loss rate, knowledge of TCP dynamics, and some
assumption about the characteristics of the TCP flows being passed
through the link. Some of the assumptions must be configured.
If loss is low, the equivalent load will be equal to the actual
percent loading. If loss is high and loading is at or near 100%, then
the equivalent load calculation provides a means of deciding which
links are more heavily overloaded. The equivalent load figure is not
intended to be an accurate pridiction of offerred load, simply a
metric for use in deciding which link to offload.
Mathis and Mahdavi provide the following estimate of loss given TCP
window size and round trip time [2].
p < (MSS / (BW * RTT))**2
The basis for the estimate is that TCP slows down roughly in
proportion to the inverse of the square root of loss. There is no way
to know how fast TCP would be going if no loss were present if there
are other bottlenecks. A somewhat arbitrary assumption is made that
TCP would go no faster than if loss were at 0.5%. If loss is greater
than 0.5% then TCP performance would be reduced. The equivalent
loading is estimated using the following computation.
equiv-load = load * K * sqrt(1 - ((1 - loss-in) * (1 -
loss-out)))
The inverse of the square root of 0.1% is 10 so 10 may be used for the
value of ``K''. A square root is somewhat time consuming to compute,
so a table lookup can be done to avoid this computation. Increments
Villamizar Expires September 13, 1998 [Page 9]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
of 0.5% would yield a 200 entry table. The computation could then be
done with a table lookup, a shift, and an integer multiply. At most
this needs to be done on links with loss once every
OMP_SAMPLE_INTERVAL seconds.
The conversion of loss to estimated loading is not at all accurate.
The non-linearity does affect the time to converge though convergence
still occurs as long as loss is positively correlated to loading.
This is discussed further in Section D.1.
3 Adjusting Equal Cost Path Loadings
Adjustments are made to a next hop structure to reflect differences in
loading on the paths as reported by the type LSA_OMP_LINK_LOAD Opaque
LSA and type LSA_OMP_PATH_LOAD Opaque LSA. Section 3.2 describes the
selection of a ``critically loaded segment'' which is used to
determine when to make adjustments and the size of the adjustments.
An adjustment to loading of a given set of equal cost paths is made
when one of the following conditions are true.
1. The LSA for the ``critically loaded segment'' has been reflooded.
2. The difference between the equivalent load of the ``critically
loaded segment'' and the lightest loaded path is greater than 5%
and the equivalent load of the ``critically loaded segment'' is
greater than 100% and 90 seconds has elapsed since the last
adjustment.
3. The difference between the equivalent load of the ``critically
loaded segment'' and the lightest loaded path is greater than 3%
and the equivalent load of the ``critically loaded segment'' is
greater than 100% and 9 180 seconds has elapsed since the last
adjustment.
4. The difference between the equivalent load of the ``critically
loaded segment'' and the lightest loaded path is greater than 5%
and the equivalent load of the ``critically loaded segment'' is
greater than 90% and 120 seconds has elapsed since the last
adjustment.
5. The difference between the equivalent load of the ``critically
loaded segment'' and the lightest loaded path is greater than 3%
and the equivalent load of the ``critically loaded segment'' is
greater than 90% and 240 seconds has elapsed since the last
adjustment.
6. 300 seconds has elapsed since the last adjustment.
Villamizar Expires September 13, 1998 [Page 10]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
The reflooding algorithm is designed to be slightly less aggressive
than the adjustment algorithm. This reduces the need to continuously
flood small changes except in conditions of overload or substantial
change in loading. Some overshoot may occur due to adjustments made
in the absence of accurate knowledge of loading.
3.1 Next hop structures
For intra-AS routes, a separate next hop structure must exist for each
destination router or network. For inter-AS routes, a separate struc-
ture must exist for each intra-AS route if a type LSA_OMP_PATH_LOAD
Opaque LSA exists for the summary LSA and more than one ABR is
advertising the summary route and the equivalent load for the summary
LSA is greater than 50% and the equivalent load is not sufficiently
smaller than the intra-AS loading. If a separate structure is not
used for the intra-AS route, then the next hop structure associated
with the reaching the ABR or set of ABRs is used. For external
routes, if an equivalent loading exists, and more than one ASBR is ad-
vertising the route, and the equivalent load is greater than 50% and the
equivalent load is not sufficiently smaller than the internal route load-
ing associated with the external next hop, then a separate structure is
used. If a separate structure is not used for an external route, then the
next hop structure associated with the reaching external next hop is used.
Hysterysis must be used in the algorithm for determining if an
equivalent load on a summary LSA or external route is considered
sufficiently smaller than the intra-AS equivalent load or if an
external route is considered sufficiently smaller than the inter-AS
equivalent load. For for the purpose of describing this algorithm one
euivalent load is referred to as the more external, and the other as
the more internal equivalent load.
If the more external equivalent load exceeds the more internal equiva-
lent load by 5% and the more internal equivalent load is under 90%,
then a separate next hop structure is created. If the more external
equivalent load falls below 20% of the more internal equivalent load
or the more internal equivalent load exceeds 95%, then an existing
separate next hop structure is marked for removal and combined with
the more internal next hop structure (see Section 3.4). The more
external equivalent load should not fall significantly below the more
internal unless either the traffic toward the more external destina-
tion increases or the loading on the more internal increases, since
the more internal equivalent load will become the critical segment on the
separate next hop structure if the load is sufficiently shifted but is un-
likely to overshoot by 20%. These threshholds should be configurable at
least per type of routes (inter-AS or external).
Villamizar Expires September 13, 1998 [Page 11]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
3.2 Critcally loaded segment
For every set of intra-AS paths, one link or part of the path is
identified as the ``critcally loaded'' segment. This is the part of
the path with the highest equivalent load as defined in Section 2.3.
For an inter-AS route with a separate next hop structure, the
critcally loaded segment is the critcally loaded segment for the
intra-AS set of paths, or the summary LSA if the equivalent load on
the summary LSA is greater. For an external route with a separate
next hop structure, the critcally loaded segment is the critcally
loaded segment for the internal route or the external route if the
equivalent load on the external route is greater.
Each next hop structure has exactly one ``critcally loaded'' segment.
An Opaque LSA may be the critcally loaded segment for no next hop
structures if it is lightly loaded. An Opaque LSA may be the
critcally loaded segment for many next hop structures if it is heavily
loaded.
3.3 Load Adjustment Rate
In order to assure stability the rate of adjustment must be
sufficiently limited. An adaptive adjustment rate method is used.
When the SPF is recalculated paths may disappear and new paths may
appear. If a new equal cost path is added to a set of existing set of
paths or a single path, the new path would have previously not carried
any traffic of the traffic. The next hop structure is initialized or
modified if there had previously been equal cost paths such that the
new path is unused. The procedures for handling links coming up or
going down is covered in Section 4.
A ``critcally loaded'' segment for a next hop structure is determined
as described in Section 3.2. When the type LSA_OMP_LINK_LOAD Opaque
LSA or type LSA_OMP_PATH_LOAD Opaque LSA for this segment is updated,
load is shed toward all equal cost paths that do not involve that
segment. A separate set of variables controlling rate of adjustment
is kept for each alternate path.
The number of paths may exceed the number of next hops. The
distinction between paths which share a next hop is important if one
of the paths sharing a next hop goes down (see Section 4). This
distinction is only needed in making the computations. When moving
the next hop structure into the data structures used for forwarding,
paths which share a common next hop may be combined. The following
variables are kept for each path in a next hop structure.
Villamizar Expires September 13, 1998 [Page 12]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
1. The current ``traffic share'' (an integer, the range is 0 to 65355
for a CRC16 hash),
2. The current ``move increment'' (an integer, the range is 0 to 65355
for a CRC16 hash),
3. The number of moves in the same direction, referred to as the
``move count''.
If there is no prior history for a path, then the move increment is
initialized to about 1% or 650. The number of moves in the same
direction is initialized to 3. No loading adjustment is made on the
first iteration.
If critcally loaded segment has not changed, or if the current path
did not contain the previous critcally loaded segment, then the
adjustment is continuing in the same direction. If the critcally
loaded segment has just changed and the path being shed load toward
contains the prior critcally loaded segment, then the adjustment
direction has reversed for this path.
If the adjustment direction has reversed, the number of moves in the
same direction is set to zero and the move increment is reduced by
half. This move increment is then used.
If the adjustment is continuing in the same direction, the number of
moves in the same direction is considered before increasing the move
increment. This value is here referred to as the move count. The
move increment is updated according to the following conditions.
1. If the move count is greater than 4, the move increment is
increased by its current value divided by the number of equal cost
paths in the next hop structure.
2. If the move count is less than or equal to 4, the move increment is
increased by 1/2 its current value divided by the number of equal
cost paths in the next hop structure.
The move increment is never less than one and the increase in move
increment is never less than one. The move increment is never allowed
to exceed the size of the hash space divided by the number of equal
cost paths in the next hop structure.
The dramatic decrease in move increment when move direction is
reversed and the slow increase in move increment when it remains in
the same direction keeps the algorithm stable. The exponential nature
of the increase allows the algorithm to track externally caused
changes in traffic loading.
Villamizar Expires September 13, 1998 [Page 13]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
The amount of hash space allocated to a path is incremented by the
move amount and the amount of hash space allocated to the critcally
loaded path or paths are decremented by this amount.
This process is repeated for each alternate path. The new hash space
boundaries are then moved to the forwarding engine.
3.4 Creating and destroying next hop structures
Section 3.1 describes the conditions under which a next hop structure
would be needed for an inter-AS route or AS external route. Briefly,
a separate next hop structure is needed if the loading indicated by
the type LSA_OMP_PATH_LOAD Opaque LSA or exterior routing protocol is
sufficiently high to require separate balancing for traffic to the
summary-LSA or exterior route and the intra-AS loading is sufficiently
low.
When a separate next hop structure is created, the same available
paths appear in the structure, leading to the same set of ABR or ASBR.
The balance on these available paths should be copied from the
existing more internal next hop structure. By initializing the new
next hop structure this way, a sudden change in loading is avoided if
the summary route or external route sinks a great deal of traffic.
When a separate next hop structure can be destroyed, the traffic
should be transitioned gradually. The next hop structure must be
marked for deletion. The traffic share in this separate next hop
structure should be gradually changed so that it exactly matches the
traffic share in the more internal next hop structure. The gradual
change should follow the adjustment rate schedule described in
Section 3.3 where the move increment is increased gradually as moves
continue in the same direction. Once the separate next hop structure
marked for deletion matches the more internal next hop structure, the
summary route or external route can be changed to point to the more
internal next hop structure and the deletion can be made.
4 Dealing with Link Failures
Link failures do occur for various reasons. OSPF routing will
converge to a new set of paths. Whatever load balance had previously
existed will be upset and the load balancing will have to converge to
a new load balanced state.
Links which are intermitent may be the most harmful. The OSPF
``Hello'' protocol is inadequate for handling intermitent links. When
such a link is up it may draw traffic during periods of high loss,
Villamizar Expires September 13, 1998 [Page 14]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
even brief periods of complete loss.
The inadequacies of the OSPF ``Hello'' protocol is well known and many
implementations provide lower level protocol state information to OSPF
to indicate a link in the ``down'' state. For example, indications
may include carrier loss, excessive framing errors, unavailable
seconds, or loss indications from PPP LQM.
Even where the use of a link is avoided by providing indication of
lower level link availability, intermitent links are still
problematic. During a brief period immediately after a link state
attribute is initially flooded OSPF state can be inconsistent among
routers within the AS. This inconsistency can cause intermittent
routing loops and have a severe short term impact on link loading. An
oscillating link can cause high levels of loss and is generally better
off held in the neighbor adjacency ``down'' state. The algorithm
described in the [4] can be used when advertising OSPF type 1 or type
2 LSA (router and network LSAs).
Regardless as to whether router and network LSAs are damped, neighbor
adjacency state changes will occur and router and network LSAs will
have to be handled. The LSA may indicate an up transition or a down
transition. In either an up or down transition, when the SPF
algorithm is applied, existing paths from the router doing the SPF to
specific destinations may no longer be usable and new paths may become
usable. In the case of an up transition, some paths may no longer be
usable because their cost is no longer among those tied for the best.
In the case of down transitions, new paths may become usable because
they are now the best path still available.
When a path becomes unusable, paths which previously had the same cost
may remain. This can only occur on an LSA down transition. A new
next hop entry should be created in which the proportion of
source/destination hash space allocated to the now infeasible path is
distributed to the remaining paths proportionally to their prior
allocation. Very high loading percentages should result, triggering
an increase in LSA_OMP_LINK_LOAD Opaque LSA flooding rate until
convergence is approached.
When a new path becomes usable it may be tied for best with paths car-
rying existing traffic. This can only occur on an LSA up transition.
A new next hop entry should be created in which the loading on the new
path is zero. If such a path were to oscillate, little or no load
would be affected. If the path remains usable, the shift of load to
this path will accellerate until a balance is reached.
If a completely new set of best paths becomes available, the load
should be split across the available paths, proportional to 10% of
link capacity plus the remaining link bandwidth as determined by prior
LSA_OMP_LINK_LOAD Opaque LSA values.
Villamizar Expires September 13, 1998 [Page 15]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
A Data Formats
@@ This is obviously a very important piece and is missing.
B Concise Statement of the Algorithms
@@ MIB counters -> pseudo code -- pull code from the simulation and
past it here.
C Changes to Existing OSPF Implementations
@@ BSD and gated would make a nice reference implementation
C.1 Forwarding
@@ ip_output.c or equiv
C.2 Routing process interface to forwarding
@@ route socket or equiv
C.3 The routing process
@@ gated or equiv
D Algorithm Performance
@@ see http://engr.ans.net/ospf-omp
D.1 Conversion from Loss to Equivalent Load
@@ black art - consult local TCP guru - very few exist
Villamizar Expires September 13, 1998 [Page 16]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
D.2 Performance when tracking traffic load change
@@ see http://engr.ans.net/ospf-omp/ramp.html
D.3 Performance when traffic loading is constant
@@ it converges and oscillates a tiny bit - about 0.5% in simulations.
D.4 Convergence after a major perturbation
@@ see simulations at http://engr.ans.net/ospf-omp - we're going kill
a link and watch the thing converge at some later date.
References
[1] Rob Coltun. The ospf opaque lsa option. Internet Draft (Work in
Progress) draft-ietf-ospf-opaque-04, Internet Engineering Task
Force, 2 1998. ftp://ds.internic.net/internet-drafts/draft-ietf-
ospf-opaque-04.txt.
[2] M. Mathis, J. Semke, J. Mahdavi, and T. Ott. The macroscopic
behavior of the TCP congestion avoidance algorithm. ACM Computer
Communication Review, 27(3), July 1997.
[3] J. Moy. Ospf version 2. Technical Report RFC 2178, Internet Engi-
neering Task Force, 1997. ftp://ds.internic.net/rfc/rfc2178.txt.
[4] C. Villamizar, R. Govindan, and R. Chandra. Bgp route flap damp-
ing. Internet Draft (Work in Progress) draft-ietf-idr-route-damp-
02, Internet Engineering Task Force, 2
1998. ftp://ds.internic.net/internet-drafts/draft-ietf-idr-route-
damp-02.txt.
Security Considerations
@@ You can't be sure if loading the *.doc version of this draft into
your word processor may infect you with a dangerous virus so a *.doc
version has not been made available. [Something vaguely serious to go
in here at a later date, subject to IESG whim de jour.] @@
Villamizar Expires September 13, 1998 [Page 17]
INTERNET-DRAFT OSPF Optimized Multipath (OSPF-OMP) March 13, 1998
Author's Addresses
Curtis Villamizar
ANS Communications
<curtis@ans.net>
Villamizar Expires September 13, 1998 [Page 18]
