Presentation for the paper Core-Stateless Fair Queueing on SIGCOMM '98

1. Core-Stateless Fair Queueing
Achieving Approximately Fair Bandwidth
Allocations in High Speed Networks
Ion Stoica @ CMU
Scott Shenker @ Xerox PARC
Hui Zhang @ CMU
Presented by Wang Yuanxuan
SIGCOMM ’98

2. Is end-to-end congestion
control enough?

3. Limitations of End-to-end Protocols
• All end-host cooperate
– What about ill-behaved users?
• End-hosts implement homogeneous
control algorithms
– Assume all clients adopt TCP

4. Enforcing Fair Bandwidth Allocation
• Router controls bandwidth
• Each flow is ensured to receive its fair
share
– no matter what control algorithms

5. Fair Queueing

6. Fair Queueing (cont.)
• Maintain per flow state

7. Fair Queueing (cont.)
• A router needs to perform
– Per packet classification
– Per flow buffer management
– Per flow scheduling
• Any disadvantages?
– Complex
– Per flow management

8. Core-Stateless Fair Queueing
• Significantly reduces implementation
complexity
• Still achieves approximately fair
allocations

9. How?
decomposing edge and core routers
estimating incoming rate and fair
rate
drop packets probabilistically

10. Outline
• Previous Work
Design
• Algorithms
• Evaluation

11. Architecture
• Two types of nodes
– edge nodes
– core nodes
• Only edge routers maintain per flow
state

12. Probabilistic Dropping
• Given
– rate estimate carried in the label ri(t)
– fair share rate of the router α(t)
• The arrival rate of flow i at time t
min⁡[ 𝑟𝑖 𝑡 , α 𝑡 ]
• The packet is dropped with probability
max⁡(0, 1 −
α(𝑡)
𝑟𝑖(𝑡)
)

13. Edge Nodes & Core Nodes
• Edge nodes
– estimate the incoming rate of each flow, and
use it to label flow's packets
• All nodes
– periodically estimate the fair rate f
– upon packet arrival, compute the forwarding
probability P and forward it
– when a packet is forwarded reset its label

14. How to estimating
ri(t) and α(t)?

15. Example Algorithms
• Computation of Flow Arrival Rate
𝑟𝑖
𝑛𝑒𝑤
= 1 − 𝑒−
𝑇𝑖
𝑘
𝐾
𝑙𝑖
𝑘
𝑇𝑖
𝑘
+ 𝑒−𝑇𝑖 𝑘
/𝐾
𝑟𝑖
𝑜𝑙𝑑
• Link Fair Rate Estimation
ɑ 𝑛𝑒𝑤 = ɑ 𝑜𝑙𝑑
𝐶
𝐹 ri Flow arrival rate
li
k Length of kth packet of flow i
ti
k Arrival time of flow i
Ti
k Duration
C Output link speed
F Estimated rate of the accepted traffic

16. Pseudocode
Don’t worry
It’s just an initial prototype

17. And finally…

18. Evaluation
• Competitors
– FIFO (first in first out, naïve)
– RED (random early detection)
– FRED (flow random early drop, drop based
on flow state)
– DRR (deficit round robin, an efficient
implementation of weighted fair queueing)
– CSFQ

19. What Should We Evaluate
• Total throughput
• Single flow throughput in congested
network
– TCP
– UDP
• Throughput under Large Latencies

20. Total Throughput of TCP/UDP Flows

21. Multiple Congested Links
• Each UDP sends at twice its fair rate

22. Relative Throughput of A TCP Flow

23. Relative Throughput of A UDP Flow

24. Throughputs with Large Latency
• Propagation delay: 100ms
Algorithm Mean Std. dev
DRR 6080 64
CSFQ 5761 220
FRED 4974 190
RED 628 80
FIFO 378 69

25. Throughputs with Large Latency
• Propagation delay: 100ms
Algorithm Mean Std. dev
DRR 6080 64
CSFQ 5761 220
FRED 4974 190
RED 628 80
FIFO 378 69

26. Summary
• CSFQ
– a hybrid scheme that asymptotically converges
to max-min fairness
• Two neat ideas
– a decomposition of edge and core routers
– a simple technique to estimate the fair rate in
a core router
• Can achieve a significant degree of
fairness
– comparable or superior to FRED

27. Drawbacks & Discussion
• Effect of large latencies
• All routers in an island need to be
simultaneously upgraded
• Dynamic packet state needs to be
implemented somehow
– Using IP fragmentation header fields?
• Assume benign routers
– What if buggy or malicious routers exist?

28. Thanks!