Transcript new-allan-load-spreading-0111.ppt

LOAD SPREADING
APPROACHES
David Allan, Scott Mansfield, Eric Gray, Janos Farkas
Ericsson
INTRODUCTION
› We want load spreading to operate within the layer
– We do not have per hop “layer violations” in frame processing
– OAM can “fate share” with ECMP flows without having to
impersonate payload
› This requires a frame to carry entropy information such
that an equal cost multi-path decision can be made at
each hop
– Entropy information is populated by a DPI function at the network
ingress
– The expectation is that next hop interface selection between the
set of next hop paths will be some function of the entropy
information
Ericsson | 2010-09-10 | Page 2
FIRST, THE OBVIOUS; IF THE SPREADING FuNCTION IS THE SAME AT
EVERY HOP:
Common hash and common next hop count at each stage is pathological
Ericsson | 2010-09-10 | Page 3
The size of the entropy field will not change this
WITH A UNIQUE HASH AT EACH HOP/Stage
Need to re-randomize each subset of traffic at each
stage in the hierarchy to use all links
Ericsson | 2010-09-10 | Page 4
HoW Entropy label is expected to work
› Each node determines the next hop as a function of both
the entropy information in the packet, and locally generated
random information
– In order to try to perfectly re-randomize any arbitrary subset of flow
labels received
› First blush is BIG is better for quantity of entropy carried in
the packet, but is this really true?
Ericsson | 2010-09-10 | Page 5
Thought Experiment
› Question: How much entropy information do we need?
› Experiment: For a given size of entropy information field, lets run it
through a number of stages and see if and observe how it degenerates
when compared to perfection
– For 3 interfaces at each hop, perfection is 1/3, 1/9, 1/27 etc. of original
number of flows assuming we perfectly randomize at every hop
› Practical reality is we will not perfectly randomize progressive subsets of the
original complete set of flows
› The example shows the average of the CV from “perfection” for 10 tries
of 5 stages of randomization with the fan out being 3 interfaces per
stage
Ericsson | 2010-09-10 | Page 6
Sidebar – quality of randomization
› Using a function of a random number and packet entropy information for
randomization at each hop will tend to produce subsets correlated in some
fashion
› This leads to poorer randomization at each hop
– Interface = f(packet entropy, random value) % (# of interfaces)
› What we want is a non-lossy and non-correlated transform in order to
maximize randomization of any arbitrary subset of flow IDs
› Using a table lookup with the table populated with random values allows
us to produce non-lossy uncorrelated subsets
– Interface = table [packet entropy] % (# of interfaces)
› However, the larger the packet entropy field is, the less practical a table
look up is to implement
–
hence we illustrate results for both in the large entropy field case
Ericsson | 2010-09-10 | Page 7
HOW MUCH ENTROPY INFORMATION IS NEEDED?
CV of load Spreading vs. size of Entropy label
10
1
1
2
3
4
5
0.1
Table 6 bit
Table 12 bit
Function 20 bit
Table 20 bit
CV
0.01
0.001
0.0001
0.00001
0.000001
Number of stages of randomization
Ericsson | 2010-09-10 | Page 8
ADVANTAGES OF USING THE B-VID as a FLOW LABEL
› The CV is still < 20% of load after 5 stages of switching for 12 bits of
entropy
› Use of VID offers backwards compatibility with older implementations
and offering some of the benefits of ECMP load spreading while using
those implementations
– They cannot do the fine grained processing of the B-VID such that each VID
value gets unique treatment as it would blow out the FDB
› But each SVL VID range normally associated with an MSTI could have a
randomly selected outgoing interface as part of FDB design
- Hence the 6 bits of entropy example on the curves
› This still requires SPB “intelligence” at nodes utilizing existing
implementations
– A completely “brain dead” nodes that simply behaves as a shared segment
will cause packet duplication when flooding… as all SPB/ECMP peers are
receiving promiscuously
– This is true no matter what flow label approach we adopt
Ericsson | 2010-09-10 | Page 9
ADVANTAGES OF USING THE B-VID as a FLOW LABEL/2
› Much smaller entropy label allows randomization
transforms that produce an uncorrelated set
– Indirection table allows for superior randomization
› Much smaller entropy label simplifies troubleshooting
– A full “blind” test needs to exercise only 4094 path permutations
› 802.1ag CFM will require minimal or no changes to work
Ericsson | 2010-09-10 | Page 10
CONCLUSIONS
› We do not need a large entropy token to get acceptable
results
› There are advantages to using the B-VID as such a token
– We can offer some ECMP benefits with primarily control plane
changes now
– We can incrementally modify the technology base over time to
improve the quality of load spreading
› O(MSTI) evolving to O(B-VID) for the amount of per-hop entropy
available
– The edge function is common
› B-VID = f(entropy source)
– OAM procedures are simplified and can work with the existing
toolset
Ericsson | 2010-09-10 | Page 11