1. Stochastic Fair Blue An Algorithm For Enforcing Fairness Wu-chang Feng (OGI/OHSU) Dilip Kandlur (IBM) Debanjan Saha (Tellium) Kang Shin (University of Michigan) April 26, 2001

2. Outline Motivation Background –Packet scheduling –Buffer management Bloom Filters and Blue Stochastic Fair Blue Conclusion

3. Motivation Internet relies on TCP congestion control Proliferation of non-responsive applications Network mechanisms –Packet scheduling –Buffer management

4. Packet scheduling approaches WFQ, W2FQ [Bennett96], Virtual Clock[Zhang90], SCFQ [Golestani94], STFQ [Goyal96] –per-flow packet scheduling and queue management  accuracy, correctness  overhead, partitioned buffers SFQ [McKenney90], CBQ [Floyd94] –approximate fairness using hashes or flow classification  low overhead  misclassification RED with per-active flow queuing [Suter98]  lower overhead than WFQ  overhead still O(N)

5. Buffer management approaches RED with penalty box [Floyd97] –Analyze drop history  no per-flow state  accuracy, complexity, large #s of bad flows Flow RED [Lin97] –Examine per-flow queue occupancy  no per-flow state  accuracy vs. queue size, large #s of bad flows Core-Stateless FQ [Stoica98] –Rate labeling at edge with priority dropping in core  core router overhead, correctness  packet overhead, coordination required, deployment

6. Stochastic Fair Blue Buffer management algorithm Single FIFO queue Combines –Bloom filters –Blue queue management

7. Bloom Filters Used in –Spell checkers (by dictionary word) –Browser caches (by URL) –Web caches (by URL) Apply multiple independent hash functions on input dictionary Lookup or locate objects based on their hashing signature

8. Blue De-couple congestion management from queue length Rely only on longer-term queue and link history Salient features –Low packet loss –High link utilization –Low queuing delay http://www.thefengs.com/wuchang/blue

9. Blue AB Sources Sinks L Mbs Rate = L Mbs Queue drops and/or ECN marks at steady rate Rate = Exactly what will keep sources at L Mbs Sinks generate DupACKs/ECN

10. Example Blue Algorithm Single dropping/marking probability –Increase upon packet loss –Decrease when link underutilized –Freeze value upon changing Upon packet loss: if ((now - last_update) > freeze_time) then P mark = P mark + delta last_update = now Upon link idle: if ((now - last_update) > freeze_time) then P mark = P mark - delta last_update = now

11. Stochastic Fair Blue (SFB) Single FIFO queue Multiple independent hash functions applied to each packet Packets update multiple accounting bins Blue performed on accounting bins Observation –Non-responsive flows drive P to 1.0 in all bins –TCP flows have some bins with normal P –P min = 1.0, rate-limit –P min < 1.0, mark with probability P min

12. SFB h L-1 h1h1 P=1.0 Non-responsive Flow P min =1.0 TCP Flow P=0.2 P=0.3 P min =0.2 h0h0 0 1 2 N-1 N L virtual bins out of L*N actual bins

13. SFB Evaluation 400 TCP flows 1 non-responsive flow sending at 45 Mbs Evaluation –200KB, 2-level SFB with 23 bins per level (529 virtual bins) –200KB RED queue –400KB SFQ with 46 RED queues 100 Mbs 45 Mbs 100 Mbs

14. SFB Evaluation REDSFQ+RED Loss rates Non-responsive = 10.32 Mbs TCP Flows = 3.07 Mbs Loss rates Non-responsive = 43.94 Mbs TCP Flows = 2.53 Mbs

15. SFB Evaluation Loss rates Non-responsive = 44.84 Mbs TCP Flows = 0.01 Mbs SFB

16. SFB and Misclassification SFB deteriorates with increasing non-responsive flows Non-responsive flows pollute bins in each level Probability of misclassification –p = [1 - (1 - 1/N) M ] L –Given M, optimize L and N subject to L*N=C

17. SFB and Misclassification 8 non-responsive flows4 non-responsive flows

18. SFB with Moving Hash Functions SFB –Virtual buckets from spatial replication of bins Moving hash functions –Virtual buckets temporally Advantages –Handles misclassification –Handles reformed flows

19. SFB with Moving Hash Functions

20. Conclusion Stochastic Fair Blue –Combine Blue queue management with Bloom filters  no per-flow state, small amount of buffers, tunable to # of bad flows, amenable to hardware implementations  short-term accuracy Current status –2 known vendor implementations –Linux implementation Future work –Hardware implementation using COTS programmable network processor