SRR: The Smoothed Round Robin Scheduler

1
SRR The Smoothed Round Robin Scheduler

• Paul Southerington
• ECE 742
• 5 May 2005

2
Overview

• Design Goals
• Why is it significant?
• How does it work?

3
General Goals of Scheduler

• Conflicting Design Goals
• Fairness
• SRR Attempts to provide long-term and short-term
  fairness
• Bounded Delay
• Scalability
• Complexity
• SRR Claims O(1) Complexity
• Formal Proofs in original paper

4
Overview

• So how does it work?
• Weight Spread Sequences
• Weight Matrix
• Scheduler Operation

5
The Weight Spread Sequence

• Spreads flow service time across entire round
• Values map to columns of matrix
• k is the order of the WSS
• Chosen based on flow weights
• Defined recursively
• Sk Sk-1, k, Sk-1
• S1 1

6
The Weight Spread Sequence (2)

• Length 2k - 1
• Sample k1 through k4
• S1 1
• S2 1 2 1
• S3 1 2 1 3 1 2 1
• S4 1 2 1 3 1 2 1 4 1 2 1 3 1 2 1

7
The Weight Matrix

• One row for each traffic flow
• Number of Columns k log2(wmax)
• Same as order of WSS
• Each row is a Weight Vector
• Series of binary coefficients
• Represent each flow weight in binary
• Each bit comprises one element of matrix

8
Sample Weight Matrix

• Weights w15, w22, w32, w43, w51

9
Scheduler Operation

• Scan WSS from left-to-right
• For each entry in WSS
• Serve traffic from flows in matching column
• Flows served sequentially (top-to-bottom)
• Move to next entry in WSS
• Start Over

10
Sample Weight Matrix

• Weights w15, w22, w32, w43, w51
• WSS S3 1 2 1 3 1 2 1

• Service order f1, f2, f3, f4, f1, f1, f4, f5,
  f1, f2, f3, f4, f1

11
Sample Service Curve
12
SRR Implementation (1)

• Weight Matrix and WSS
• Size may be fixed
• WSS may be precomputed
• Can tradeoff granularity for max speed
• With 32nd-order WSS
• 4 Tbps with 1 kbps resolution
• 8 Tbps with 2 kbps resolution

13
SRR Implementation (2)

• Three Functions
• Add_flow()
• Executed whenever a new flow arrives
• Adds new entry to bottom of doubly-linked list
  (column)
• Worst-case O(k) operation, not O(1)!
• Del_flow()
• Executed when a flow is no longer backlogged
• Basically Add_flow() in reverse

14
SRR Implementation (3)

• Three Functions (contd.)
• Schedule()
• Operates as described in previous example
• Also handles deficit value
• Add_flow and Del_flow called from here

15
Future Improvements

• Improvements to Add_flow and Del_flow
• Try to improve worst-case O(k) performance
• Parallel operation
• WSS Compression
• Remove WSS elements for empty columns
• May not be possible to implement efficiently

16
Conclusions

• Elegant solution to conflicting design goals
• May not really perform at O(1)
• Still some room for improvement