Effect of IW and Initial RTO changes Ilpo J arvinen, Aki Nyrhinen, - PowerPoint PPT Presentation

Effect of IW and Initial RTO changes Ilpo J¨ arvinen, Aki Nyrhinen, Aaron Yi Ding, Markku Kojo Department of Computer Science University of Helsinki IETF79 / Beijing Nov 11th 2010

Introduction Simulation study to evaluate effects of recently proposed changes: Initial Window change from 3 packets to 10 packets Initial RTO change from 3 seconds to 1 second Focus on (typical) slow/moderate bit-rate wireless links like environments Initially presented IW10 results in the last ICCRG meeting @ Maastricht IETF79 / Beijing Nov 11th 2010 2

Test setup Links (bw/one-way propagation delay) EGDE 160kbps/250ms, BDP = 7 pkts (6.7) HSPA 2Mbps/70ms, BDP = 24 pkts (23.3) LTE 50Mbps/15ms, BDP = 125 pkts No wireless errors, nor allocation / error related delays considered 11ms propagation delay from sender to wireless link Buffer (FIFO) sizes BDP (Bandwidth Delay Product) 2 · BDP 50 Packets (EDGE only) Workload: A burst of 1, 2, 6 or 18 simultaneous downstream TCP flows (total 180kB) competing against a similar later starting burst (another 180kB), 100 replications ns2 TCP SACK in use IETF79 / Beijing Nov 11th 2010 3

Summary of IW10 Effects With small number of TCP flows, IW10 improves performance With larger number of flows, IW10 tends to decrease performance - Regardless of IW, too many flows clearly results in suboptimal performance Fairness for later starting traffic improves with IW10 Fairness within both bursts worse with IW10 IETF79 / Beijing Nov 11th 2010 4

IRTO: LTE (50Mbps/15ms, BDP=125 Packets) No changes No spurious RTOs RTOs with IW10 when # of flows is 6+6 or 18+18 But not in the beginning for the flow that completes last (not for the SYN nor the first packet) ⇒ IRTO has no effect IETF79 / Beijing Nov 11th 2010 5

IRTO: HSPA (2Mbps/70ms, BDP ≈ 23 Packets) Later Starting Burst Observations 5.5 6+6 flows 18+18 flows 5 When overloaded, small improvement Elapsed time of the longest flow (s) 4.5 (median, quartiles, min-max) for the longest cases 4 among later starting 3.5 traffic 3 Opposite effect for the first starting burst 2.5 (the shortest cases 2 delayed) 1.5 No changes due to 1 IRTO1 with 1+1 or IW3 IW3 IW10 IW10 IW3 IW3 IW10 IW10 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 2+2 flows BDP BDP BDP BDP 2BDP 2BDP 2BDP 2BDP IETF79 / Beijing Nov 11th 2010 6

IRTO: EDGE (160kbps/250ms, BDP ≈ 7 Packets) 50 Packets Buffer Observations Later Starting Burst 34 4 1+1 flows, el. time With large buffer, number 2+2 flows, el. time 32 1+1 flows, RTOs of RTOs increase 2+2 flows, RTOs 30 Elapsed time of the longest flow (s) Mostly spurious RTOs 3 (median, quartiles, min-max) 28 ⇒ Completion of the 26 longest flow is delayed RTOs 24 2 The same trend with larger number of flows 22 When IW10 in use, the 20 1 first starting burst is able 18 to take advantage and 16 completes unfairly early 14 0 IW3 IW3 IW10 IW10 IRTO3 IRTO1 IRTO3 IRTO1 IETF79 / Beijing Nov 11th 2010 7

IRTO: EDGE (160kbps/250ms, BDP ≈ 7 Packets) First Starting Burst Later Starting Burst 28 28 1+1 flows 1+1 flows 2+2 flows 2+2 flows 26 26 Elapsed time of the longest flow (s) Elapsed time of the longest flow (s) 24 24 (median, quartiles, min-max) (median, quartiles, min-max) 22 22 20 20 18 18 16 16 14 14 12 12 10 10 IW3 IW3 IW10 IW10 IW3 IW3 IW10 IW10 IW3 IW3 IW10 IW10 IW3 IW3 IW10 IW10 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 IRTO3 IRTO1 BDP BDP BDP BDP 2BDP 2BDP 2BDP 2BDP BDP BDP BDP BDP 2BDP 2BDP 2BDP 2BDP Observations Mostly the same regardless of IRTO IW10+IRTO1 becomes more fair RTO occurred sooner for the later starting burst (a spurious one) IETF79 / Beijing Nov 11th 2010 8

RED Configuration Cfg RED REDok Link EDGE HSPA/LTE EDGE HSPA LTE 0.002 0.002 0.2 0.02 0.001 w q max p 0.1 0.1 0.65 0.65 0.1 3 5 3 3 5 th min 9 20 40 50 125 th max buffer size 2 · BDP 2 · BDP 50 100 400 Large buffers with RED configuration were not tested Not useful because of avg > th max dropper REDok config aimed to highly varying load Thus vastly different from “default configuration” Aggressive enough to respond to slow start Parameters are link characteristics dependent IETF79 / Beijing Nov 11th 2010 9

Single Flow One-way Delay (FIFO, RED and IW3, IW10) 2 1+1 flows, EDGE, IRTO3, IW3, BDP 1+1 flows, EDGE, IRTO3, IW10, BDP 1+1 flows, EDGE, IRTO3, IW3, 2BDP 1.8 1+1 flows, EDGE, IRTO3, IW10, 2BDP 1+1 flows, EDGE, IRTO3, IW3, RED 1+1 flows, EDGE, IRTO3, IW10, RED 1.6 1+1 flows, EDGE, IRTO3, IW3, 50 pkts 1+1 flows, EDGE, IRTO3, IW10, 50 pkts 1.4 One-way delay (s) 1.2 1 0.8 0.6 0.4 0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 CDF IW10 slightly more aggressive RED similar to FIFO behavior (too slow to react) With BDP IW10 hurts itself due to self-congestion Slightly smaller delays except for the highest end IETF79 / Beijing Nov 11th 2010 10

Single Flow One-way Delay (FIFO, REDok and IW3, IW10) 2 1+1 flows, EDGE, IRTO3, IW3, BDP 1+1 flows, EDGE, IRTO3, IW10, BDP 1+1 flows, EDGE, IRTO3, IW3, 2BDP 1.8 1+1 flows, EDGE, IRTO3, IW10, 2BDP 1+1 flows, EDGE, IRTO3, IW3, REDok 1+1 flows, EDGE, IRTO3, IW10, REDok 1.6 1.4 One-way delay (s) 1.2 1 0.8 0.6 0.4 0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 CDF Also REDok fails to control the delay increase IW10 imposes Maximum values with REDok : IW10: 2.80s IW3: 2.06s IETF79 / Beijing Nov 11th 2010 11

6 Flows One-way Delay (FIFO, RED and IW3, IW10) 0.45 6+6 flows, HSPA, IRTO3, IW3, BDP 6+6 flows, HSPA, IRTO3, IW10, BDP 6+6 flows, HSPA, IRTO3, IW3, 2BDP 0.4 6+6 flows, HSPA, IRTO3, IW10, 2BDP 6+6 flows, HSPA, IRTO3, IW3, RED 6+6 flows, HSPA, IRTO3, IW10, RED 0.35 One-way delay (s) 0.3 0.25 0.2 0.15 0.1 0.05 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 CDF Again, RED reacts too slowly IW10 less aggressive due to self-congestion ⇒ more bursty IETF79 / Beijing Nov 11th 2010 12

6 Flows One-way Delay (FIFO, REDok and IW3, IW10) 0.45 6+6 flows, HSPA, IRTO3, IW3, BDP 6+6 flows, HSPA, IRTO3, IW10, BDP 6+6 flows, HSPA, IRTO3, IW3, 2BDP 0.4 6+6 flows, HSPA, IRTO3, IW10, 2BDP 6+6 flows, HSPA, IRTO3, IW3, REDok 6+6 flows, HSPA, IRTO3, IW10, REDok 0.35 One-way delay (s) 0.3 0.25 0.2 0.15 0.1 0.05 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 CDF With REDok, traffic regulation works without heavy tail-drop ⇒ IW10 shows to be significantly more aggressive Maximum values with REDok : IW10: 0.429s IW3: 0.296s IETF79 / Beijing Nov 11th 2010 13

One-Way Delay in Rest of Cases Similar behavior observed: Self-congestion ⇒ IW10 is less aggressive Except for the very highest end (in some of the cases) With low enough load, IW10 is slightly more aggressive IRTO1 only slightly “shifts” curves Only happening when IRTO1 has some effect in the first place Quite insignificant in numbers Actual shape of the delay curves vary per queue size and type, however, those differences are out of scope here IETF79 / Beijing Nov 11th 2010 14

Conclusions Smaller initial RTO performs better when effective e2e RTT smaller than 1 second More controversial when e2e RTT is larger IW10, while improving elapsed times, imposes higher queuing delay than IW3 However, if self-congesting, IW3 is more aggressive in terms of queuing delay AQM (RED) failed to control the increase in the queuing delay IETF79 / Beijing Nov 11th 2010 15

Questions? IETF79 / Beijing Nov 11th 2010 16

Backup slides IETF79 / Beijing Nov 11th 2010 17

RED config (detailed ns2) Queue/RED set bytes_ true Queue/RED set queue_in_bytes_ true Queue/RED set gentle_ false Queue/RED set setbit_ false Queue/RED set use_mark_p_ false Queue/RED set mean_pktsize_ 1500 Queue/RED set idle_pktsize_ 1500 Queue/RED set q_weight_ $wq Queue/RED set thresh_ $minth Queue/RED set maxthresh_ $maxth Queue/RED set linterm_ [expr 1.0/$maxp] Queue/RED set wait_ false IETF79 / Beijing Nov 11th 2010 18

6 Flows Elapsed Times (FIFO, RED, REDok and IW) HSPA, Later Starting Burst 8 1+1 flows 2+2 flows 6+6 flows 7 18+18 flows Elapsed time of the longest flow (s) (median, quartiles, min-max) 6 5 4 3 2 1 IW3IRTO3 IW10IRTO3 IW10IRTO1 BDP 2BDP RED REDok BDP 2BDP RED REDok BDP 2BDP RED REDok IETF79 / Beijing Nov 11th 2010 19

6 Flows Fairness (FIFO, RED, REDok and IW) HSPA 1 Jain’s fairness index between the bursts 0.95 (median, quartiles, min-max) 0.9 0.85 0.8 1+1 flows 2+2 flows 6+6 flows 18+18 flows 0.75 IW3IRTO3 IW10IRTO3 IW10IRTO1 BDP 2BDP RED REDok BDP 2BDP RED REDok BDP 2BDP RED REDok IETF79 / Beijing Nov 11th 2010 20