Active Queue Management for Web Traffic

1
Active Queue Management for Web Traffic

• Mark Claypool, Bob Kinicki and Matt Hartling
• Worcester Polytechnic Institute
• Computer Science Department
• Worcester, MA 01609
• claypool,rek_at_cs.wpi.edu

2
Outline

• Motivation
• RED
• SHRED Algorithm
• Performance Metrics
• Web Traffic Model
• Topology and Experimental Procedures
• RED, SHRED and Drop Tail Results
• Conclusions

3
Motivation for Active Queue Management

• Congestion is still an Internet problem.
• Short TCP Web flows dominate the Internet.
• Mice Web objects yield short-lived flows
  smaller than 2KB.
• TCP slow-start provides biased performance for
  short Web flows.

4
Motivation for SHRED

• TCP uses cwnd to limit a flows sending rate.
• Fast Retransmit is ineffective when cwnd is less
  than four and for last three packets of a flow.
• Retransmission Time Out RTO penalty is high in
  the first few packets of a flow.

5
RED Routers

• Random Early Detection (RED) detects congestion
  early by maintaining an exponentially-weighted
  average queue size.
• RED probabilistically drops packets before the
  queue overflows to signal congestion to TCP
  sources.
• RED attempts to avoid global synchronization and
  bursty packet drops.

6
RED
packet
minth
maxth
minth average queue length threshold for
triggering probabilistic drops/marks. maxth
average queue length threshold for triggering
forced drops.
7
RED Parameters

• qavg average queue size
• qavg (1-wq) qavg wq instantaneous queue
  size
• wq weighting factor 0.001 lt
  wq lt 0.004
• maxp maximum dropping/marking probability
• pb maxp (qavg minth) / (maxth
  minth)
• pa pb / (1 count pb)
• buffer_size the size of the router queue in
  packets.

8
RED Router Mechanism
1
Dropping/Marking Probability
maxp
0
Min-threshold
Queue Size
Max-threshold
Average Queue Length (avgq)
9
SHort-lived flow friendly REDSHRED

• Basic SHRED Idea
• To lower the drop probability for flows with
  small cwnds and to increase the drop probability
  for flows with relatively large cwnds.

10
SHRED

• SHRED uses an edge hint and inserts the
  current value of TCP cwnd into IP packet header.
• Upon packet arrival at SHRED router
• cwndavg (1 wc) cwndavg (wc) cwndsample
• where
• wc set to 0.002

11
SHRED

• SHRED modifies minth and maxp
• minth-mod minth
• (maxth minth) x (1
  cwndsample / cwndavg )
• maxp-mod maxp x (maxth minth-mod) / (maxth
  minth)
• and re-computes pb
• pb maxp-mod x (qavg minth-mod) /(maxth
  minth-mod)

12
SHRED Mechanismusing gentle RED
13
Web Traffic Characterization

• General Web flow modeling through congestion
  yields increased response times that in turn
  decrease the load generated by a Web client.
• The model constructed has multiple objects per
  Web page downloaded in parallel and followed by a
  waiting period determined by the page generation
  rate.

14
Web Traffic Characterization

• For ns-2 simulations
• Pareto II used to generate Web objects min
  12 bytes, max 2MB, average object size 10KB,
  1.2 shape parameter.
• Canonical experiment 1 object per page (unless
  otherwise specified).

15
Performance Metrics

• Object transmission time - the time to transfer a
  single Web object from a server to the client.
• Web response time the time to download all
  objects in a Web page.
• goodput (Mbps) - the rate at which packets arrive
  at the receiver. Goodput differs from throughput
  in that retransmissions are excluded from goodput.

16
Performance Metrics

• Jains fairness
• For any given set of user throughputs (x1, x2, ,
  xn), the fairness index to the set is defined
• f (x1, x2, , xn)
• Percentage of packets dropped per flow.

17
Simulation Topology
Web Source
Web Sink
100 Mbps 1 ms.
100 Mbps 1 ms.
10 Mbps 60 ms.
Router
Congested Router
FTP Source
FTP Sink
RED Parameters Minth 30 pkts maxth 90
pkts maxp 0.1 wq 0.0008 avg pkt 974
bytes maxq 225 pkts


FTP Source
FTP Sink
18
Experimental Procedures

• Simulated RED, SHRED and Drop Tail.
• A few early longer duration experiments were
  conducted to determine point when simulations
  were stable.
• All experiments were 160 simulated seconds.
• Measurements were taken after 20 seconds of
  warm-up period.
• Simulated both FTP traffic and Web traffic using
  TCP Reno.

19
Traffic Mixes

• Web-only experiments
• similar to RED-Tuning paper procedures.
• Web-mixed experiments
• FTP flows fixed at 10.
• Web flows varied from 40 to 80 of bottlenecked
  bandwidth (10Mbps).

20
Traffic Mixes (cont.)

• FTP-mixed experiments
• Web traffic load fixed at 50.
• FTP flows varied from 0 to 40.
• FTP-only experiments
• No Web flows.
• FTP flows varied from 10 to 100.

21
Fig 3a Web-only Transmission Time CDF100 load
140 active flows
22
Fig 3b Web-only Transmission TimeCDF Tail100
load 140 active flows
23
Fig4a Web-mixed Transmission Time CDF 70 Web
load 10 FTP flows
24
Fig 4b Web-mixed Transmission TimeCDF Tail70
Web load 10 FTP flows
25
Fig. 5 Normalized Web Response TimeWeb-mixed
Experiments
26
Fig 6 Percent DropsWeb-mixed Experiments
27
SHRED Performance

• Web-only SHRED has best performance.
• Web-mixed SHRED closer to uncongested
  performance. SHRED is better than RED and Drop
  Tail in heavy-tail of CDF due to RTO issues.
• With normalized transmission time, SHRED 4
  better than RED which is 8 better than Drop
  Tail.
• As number of flows increase, the SHRED benefit in
  packet drops widens.

28
Fig. 7 Normalized Web Response TimeFTP-mixed
Experiments
29
Web Response Time

• More simulations run where there are multiple
  objects per page.
• Uniform random distribution of Web objects/page
  for (1 to 8), (1 to 16) and (1 to 32) for
  Web-only and Web-mixed experiments.
• Response time the time to download the whole
  page of objects.

30
Fig. 8 Normalized Web Response TimeWeb-mixed
Experiments
31
Table 1FTP-only Goodput (Mbps)
32
Table 2FTP-only Jains Fairness
33
Conclusions

• SHRED produces lower object transmission times
  than either RED or Drop Tail in We-only and mixed
  traffic simulations.
• SHRED yields significant response time
  improvement when there are multiple objects per
  page.
• SHRED improvements do not negatively impact FTP
  traffic.
• Basic SH scheme can be applied to other AQMs
  (e.g., research on PISA).