Switches Know the Exact Amount of Congestion S ERHAT A RSLAN , N ICK M C K EOWN Buffer Sizing Workshop, Stanford University Dec 2-3, 2019
Main Takeaways 1. We should directly measure the cause of congestion. 2. Congestion is buildup of the queue at a bottleneck link. 3. Switches today can provide their buffer occupancy freely. 4. Bottleneck buffer occupancy can be collected by updating a header field on packets. Serhat Arslan, sarslan@stanford.edu
Serhat Arslan, sarslan@stanford.edu
Serhat Arslan, sarslan@stanford.edu
Why donβt switches simply tell everything? Measure the congestion directly at the location of congestion itselfβ¦ Serhat Arslan, sarslan@stanford.edu
Congestion Signal on TCP Header Serhat Arslan, sarslan@stanford.edu
Construct Your Custom Signal π = π 2 , β¦ , π 56 , π₯βππ π π β βππ_πππ£ππ’ ππ π π π( π ) = π₯{ β π π π > ππππππ } π( π ) = β πEπ RTT ECN ππ π ππ π π = πππ( π ) β πEπ π( π ) = Bottleneck! Serhat Arslan, sarslan@stanford.edu
Effectiveness Analysis Ablation Scenario: Serhat Arslan, sarslan@stanford.edu
Effectiveness Analysis Serhat Arslan, sarslan@stanford.edu
Feasibility Analysis Overhead Costs: Latency Throughput NOPQ RST β 0.4% of MTU Enough to collect 8 M Serhat Arslan, sarslan@stanford.edu
Main Takeaways 1. We should directly measure the cause of congestion . 2. Congestion is buildup of the queue at a bottleneck link . 3. Switches today can provide their buffer occupancy freely. 4. Bottleneck buffer occupancy can be collected by updating a header field on packets. Serhat Arslan, sarslan@stanford.edu
Questions? S E R H A T A R S L A N s a r s l a n @ s t a n f o r d . e d u
Recommend
More recommend