fairness issues in red gateways
DESCRIPTION
Fairness Issues in RED Gateways. Srinivas R. Avasarala [email protected] CS Dept., Purdue University. Motivations. Study the aims and approaches of RED and the variants proposed, SRED and FRED, to achieve fairness. Provide Implementations of RED variants on Network Simulator ns. - PowerPoint PPT PresentationTRANSCRIPT
Motivations
• Study the aims and approaches of RED and the variants proposed, SRED and FRED, to achieve fairness.
• Provide Implementations of RED variants on Network Simulator ns.
• Analyze and compare performances
RED Overview
• Detects congestion using avg. queue size
• Drop probability is a function of the avg. queue size. [min_th, max_th, p_max]
• No bias against bursty traffic
• No global synchronization
• A flow’s drop rate is proportional to its share of the bandwidth through the gateway
RED Drawbacks
• There is no TCP-awareness• Dropping packets from flows in proportion
to their bandwidth doesn’t result in fair sharing. e.g Fast link vs. a Slow link
• Does not consider the number of flows• Discrimination against:
– Large RTT flows w.r.t. Small RTT flows– Adaptive flows w.r.t. Non-adaptive flows
SRED: Stabilized RED
• Avg. Q size is not used for drop probability
• Does statistical estimation of #active flows
• Drop probability is: f (instantaneous Q size, #active flows)
• Stabilizes buffer occupation at a level independent of the number of active flows
• Identifies misbehaving flows
SRED Approach
• Hit(t) = 0 if no hit; 1 if hit• Hit frequency P at an instant t is defined as:
P(t) = (1 – α)*P(t – 1) + α * Hit(t), 0 < α < 1• α ~ p/M, p is overwrite prob, M is flow table size
• psred(q) = pmax if B/3 <= q < B,
= pmax/4 if B/6 <= q < B/3,
= 0 if 0 <= q < B/6
• pzap = psred(q) * min (1, 1/(256 * P(t))2)
FRED: Flow RED
• Uses per-active-flow accounting
• Each flow’s loss rate depends on the flow’s buffer use
• Protects adaptive flows
• Isolates non-adaptive greedy flows
• Uses the RED approach with modifications
FRED Approach
• Maintain a flow-state table, one entry per active flow with fields: qleni and strikei
• Parameters: minq, maxq, avgcq• Identify non-adaptive flows as:
– qleni >= maxq
– avg >= maxth & qleni > 2 * avgcq– qleni >= avgcq & strike > 1
• Drop only from robust flows: qleni > minq
My Conclusions
• SRED’s performance depends on the correctness of its statistical estimations
• SRED’s equations use parameters that require appropriate tuning to achieve results
• FRED is very effective in achieving all its goals
• FRED’s reliance on per-flow-state prevents it from getting a wide acceptance