SRED: Stabilized Red

1
SRED Stabilized Red

• T.J. Ott, T.V.Lakshman and Larry H. Wang
• Presented by Prashant Ratanchandani

2
Main Idea

• Compare each arriving packet with a randomly
  chosen packet that preceded it into the buffer
• If both packets are of the same flow, declare a
  hit
• Estimate the number of active flows
• Find the candidates for a misbehaving flow
• J out of K hits information

3
Zombie List

• List of M recently seen flows
• Longer memory than the buffer alone
• Information for each zombie
• Countnumber of packets this zombie received
• Timestamp arrival time of most recently received
  packet

4
Zombie List Operations

• While the zombie list is not full, add the packet
  flow identifier to the list, increment count and
  set timestamp
• Zombie list is full
• Each arriving packet p compared with randomly
  chosen zombie
• If(hit) countcount1
  t timestamptimestampp
• If (!hit)
• with probability P, overwrite the flow
  identifier of zombie with arrived packet
• Packet may be dropped irrespective of a hit or not

5
Intuitive Argument

• Zombie list likely to lose memory every M/p
  packets, M being the length of the zombie list
• Fewer active flows imply there will be more hits
• Misbehaving flows shall cause more hits than
  well-behaved ones

6
Hits and Active Flows

• P(t)-estimate for the hit frequency at the
  arrival of the t-th packet at the buffer
• Hit(t)0, if no hit 1, if hit
• P(t)(1- ?)P(t-1)?Hit(t)
• For example ?p/M
• P(t) is an estimate of the frequency of hits for
  approximately the most recent M/p packets before
  packet t

7
Proposition

• P(t)-1 is a good estimate for the effective
  number of active flows in the time shortly before
  the arrival of packet t

8
Supporting argument

• P(arrived packet belongs to flow i)?i
• Zombie represents flow I with probability ?i
• Thus probability that the packet shall cause a
  hit P(Hit1)?(?i2)
• Assuming N flows of identical traffic intensity
  (?i1/N) P(Hit1)1/N

9
Proposition-Supporting Arguments

• P(t)-1 is an estimate of the effective number of
  active flows even in the asymmetrical case
• If ?I2-I then EP(t)3/16
• Flows taking more than 3/16 of the average
  bandwidth are taking their fair share and flows
  less than 3/16 are taking their fair share
• If N flows with probability ?I then
  1/Nlt?(?i2)lt1
• Update P(t) after every L packets
• P(new)(1-L?)P(old) ?H

10
Simple SRED

• Target buffer occupation Q0
• Set a drop probability p
• Square root law congestion window of each flow,
  cwnd ?p-1/2 MSSs
• Sum of N congestion windows N.p-1/2
• Q0 N p-1/2? p (N/ Q0)2
• p is proportional to N2

11
Candidate Probability Drop Function

• Buffer capacity B bytes
• Psred(q)pmax if B/3ltqltB pmax/4 if
  B/6ltqltB/3 0 if 0ltqltB/6
• Pmax chosen to be 0.15
• On packet arrival P(t) is updated
• Pzappsred(q)X min(1,1/(256 X P(t))2)

12
Probability function(contd.)

• psred has only three possible values
  (0,pmax,pmax/4)
• psred depends only on the instantaneous buffer
  occupation q
• When 1/256ltP(t)lt1 pzap(psred/65,536)X(number
  of flows)2
• When 0ltP(t)lt1/256 pzappsred
• If p becomes very large TCP sources spend most of
  the time in time-out
• When P(t) becomes small estimating P(t) is
  unreliable

13
Full SRED

• Pzappsred(q)X min(1,1/(256 X P(t))2)X(1Hit(t)/P(
  t))
• The old pzap is multiplied by
• 1 ?I/ ?(?j2)
• Increases the drop probability for overactive
  flows and reduces TCPs bias for flows with short
  RTTs

14
Simulation ResultsSymmetrical network

• Pmax0.15
• Transfer of infinitely large file
• Buffer occupancy below or slightly over B/3
• Loss always due to RED or SRED, and not buffer
  overflow
• Buffer occupancy for Nlt256 is independent of the
  number of flows

15
Simulation Results-Symmetrical network

• Buffer occupancy for Ngt256 increases with the
  number of flows.
• Buffer occupation almost always never decreases
  below B/6, where pzap0

16
Comparison-RED and SRED
17
Comparison-RED and SRED

• For RED, the buffer occupancy increases with the
  number of connections
• SRED buffer occupancy is at least B/3

18
Asymmetrical Network Simulations

• Connections with shorter round trip times get
  higher throughput
• Full SRED reduces this advantage by a small
  amount
• Full SRED is better when the number of active
  connections decreases
• Simple SRED and Full SRED both equally effective
  in stabilizing buffer occupation

19
Realistic Source Model

• 2000 sources
• 200-400 active at a time
• P(t)-1 underestimates the number of active
  flows
• Many flows supposed to be active are not active
  yet, or not active anymore
• Active connections have varying congestion
  windows

20
Buffer Occupancy-Real Sources
21
Misbehaving Sources

• Hit mechanism is used to identify candidates for
  misbehaving flows
• A flow taking more than its fair share is a hit
  with a high Count for the zombie
• Total Occurrence is the sum over all zombies that
  have the same flow of (Count1)
• Mechanisms are used to filter flows and find the
  misbehaving subset

22
Differences with RED

• SRED estimates number of active flows
• Misbehaving flows can be identified without
  keeping per-flow state
• Drop probabilities are adjusted according to
  number of active flows
• No computation of average queue length
• Only TCP flows are assumed

23
Main Contributions

• SRED provides a mechanism to estimate number of
  active flows and identify misbehaving flows
• SRED controls buffer occupancy by adjusting drop
  probabilities using estimated number of active
  flows

24
Future Work

• Parameter Tuning(pmax and psred)
• Research on how to declare a flow as misbehaving