Simulation Results for QoS, pDCF, VDCF, Backoff/Retry January 2001 Simulation Results for QoS, pDCF, VDCF, Backoff/Retry Greg Chesson Aman Singla Atheros Communications Greg Chesson, et al, Atheros Communications

Outline Backoff/Retry/Access Methods Backoff: January 2001 Outline Backoff/Retry/Access Methods Backoff: constant (non-increasing) exponential (doubling) hybrid (doubling only after first retry) Access Methods: DCF, pDCF QoS Scenarios Pair of 1 Mbyte/s Mpeg streams plus background load Sim Group: 10/20/30-node scenario at 11 Mb/s Greg Chesson, et al, Atheros Communications

Backoff Scenarios January 2001 Node 1 Simulate 3 backoff methods No backoff increase on retry Exponential backoff Hybrid (increase backoff after first retry) Simulate both pDCF and DCF 36 Mb/s PHY, CW=15,PP=.12 For each combination of backoff and MAC Simulate with 2 thru 12 nodes 60 total runs No upper layer protocol or application Heavy offered load Fully backlogged queues Plot Goodput (aggregate bandwidth) Collisions Channel idle time Latency/Jitter Node 2 AP Node 0 Node 12 Greg Chesson, et al, Atheros Communications

January 2001 Backoff Results constant backoff => high collision rate, reduced bandwidth, more congestion Exponential backoff => gradual degradation Hybrid backoff => close to exponential performance These conclusions should not be controversial as they support the conventional wisdom. Greg Chesson, et al, Atheros Communications

January 2001 MAC Results Goodput: pDCF slightly better than DCF (~4%) because aggressive access technique improves channel efficiency at the expense of more variable channel access wait times. Latency/Jitter: DCF better than pDCF because TxOPs are more evenly distributed between stations (even for a 2-station scenario!). Explanation: study expected MAC access delay times per station. Greg Chesson, et al, Atheros Communications

Goodput January 2001 pDCF DCF Channel hybrid saturation pDCF DCF Steep dropoff If constant backoff Greg Chesson, et al, Atheros Communications

Collision time for constant backoff January 2001 Collision time for constant backoff hybrid Exponential backoff Greg Chesson, et al, Atheros Communications

1 station 2 stations n stations January 2001 pDCF 1 station 2 stations n stations DCF 1 stations 2 stations pDCF shows ~3X latency/jitter for 2 stations, why? Greg Chesson, et al, Atheros Communications

Plot % time that another station gets 1, 2, 3, … TxOPs before you. January 2001 Plot % time that another station gets 1, 2, 3, … TxOPs before you. That is, how many times do you lose before winning a TxOP? Ideal curve for 2 stations centers about 1. stations alternate accesses would defer (lose) only once per transmission Majority of DCF(CW=15) samples lie within ideal curve pDCF (PP=.12) show 50% of samples are back-to-back. Thus, pDCF stations see greater latency variance (jitter). Sometimes you lose 4 or more times. Greg Chesson, et al, Atheros Communications

Ideal curve for 5 stations centers about 4. January 2001 Ideal curve for 5 stations centers about 4. Neither pDCF nor DCF are ideal, but only 20% of DCF transmission are back-to-back from The same station compared to 30% for pDCF. Difference increase with more stations. Jitter increases with more stations. Greg Chesson, et al, Atheros Communications

Conclusions Exponential backoff essential for robustness January 2001 Conclusions Exponential backoff essential for robustness DCF(CW) equals pDCF(PP=2/(CW+1)) pDCF achieves slightly more bandwidth pDCF introduces much more jitter pDCF(adaptive PP) equals DCF(adaptive CW) Set CW=2/PP – 1 After adaptation: same bandwidth/jitter differences No compelling reason to change basic access method Greg Chesson, et al, Atheros Communications

Qos Simulations Video scenario: 2 1-Mbyte/sec streams plus load January 2001 Qos Simulations Video scenario: 2 1-Mbyte/sec streams plus load AP sends to 2 stations 4 additional stations generating high load TCP load (tcp data and tcp acks forward thru AP) unconstrained UDP load (AP acts as infinite sink) Plot bandwidth, latency, packet drops for DCF AP with packet scheduler (only) VDCF Greg Chesson, et al, Atheros Communications

Video splits bandwidth at first. Then tcp streams start. January 2001 Video splits bandwidth at first. Then tcp streams start. Packets are dropped from truncated fifos (IFQ drops). No packets are dropped for excessive MAC retransmission. Behavior is worse with UDP background load (next slide). Greg Chesson, et al, Atheros Communications

Many more dropped packets. Caused by high latencies (next slide). January 2001 Many more dropped packets. Caused by high latencies (next slide). Greg Chesson, et al, Atheros Communications

Latency unsatisfactory for the application. January 2001 Latency unsatisfactory for the application. UDP flows have more latency because the small UDP packets Are usually waiting for the larger video frames. Greg Chesson, et al, Atheros Communications

AP Scheduling Add priority packet scheduling to AP January 2001 AP Scheduling Add priority packet scheduling to AP Video packets jump to head of queue(-5) Assume 5 packets committed to hardware Video can bypass all but 5 Greg Chesson, et al, Atheros Communications

Bandwidth plot for AP priority queue-only configuration. January 2001 Bandwidth plot for AP priority queue-only configuration. No other MAC QoS needed for this scenario. Demonstrates capability of AP queuing to control traffic. Majority of traffic passes through AP in many (but not all) environments. Greg Chesson, et al, Atheros Communications

Latency with priority queue-only configuration. January 2001 Latency with priority queue-only configuration. Greg Chesson, et al, Atheros Communications

AP queue-only configuration is less effective with UDP load. January 2001 AP queue-only configuration is less effective with UDP load. Greg Chesson, et al, Atheros Communications

Adding VDCF TCP load Increases bandwidth slightly Improves latency January 2001 Adding VDCF TCP load Increases bandwidth slightly Improves latency UDP (raw) load Provides differentiated bandwidth Greg Chesson, et al, Atheros Communications

January 2001 Greg Chesson, et al, Atheros Communications

Video stream latency with VDCF with TCP load January 2001 Video stream latency with VDCF with TCP load Greg Chesson, et al, Atheros Communications

Video streams latency with VDCF and UDP load. January 2001 Video streams latency with VDCF and UDP load. Greg Chesson, et al, Atheros Communications

Sim Group Scenario 30 active stations on 11 Mb/s phy January 2001 Sim Group Scenario 30 active stations on 11 Mb/s phy 3 traffic classes: High, Medium, Low Objective: 2/2/2 differentiated bandwidth Plot goodput, latency, drops for DCF-only baseline VDCF with two parameter sets 32 active stations Add two stations in a 4th traffic class: X-High Use PIFS instead of DIFS Greg Chesson, et al, Atheros Communications

10 active stations 20 stations 30 stations January 2001 Baseline DCF Goodput 10 active stations 20 stations 30 stations 3 Mb/s load 6 MB/s load 9 Mb/s load (meltdown) Greg Chesson, et al, Atheros Communications

Plot both offered load and DCF Goodput. January 2001 Plot both offered load and DCF Goodput. Goodput closely overlays the random exponential load generation except in overload. Generated load goodput Greg Chesson, et al, Atheros Communications

Same scenario with conservative VDCF parameters. January 2001 Same scenario with conservative VDCF parameters. Observe nearly 2X differentiated bandwidth for 10 and 20 station loads only. Greg Chesson, et al, Atheros Communications

No packet drops for high-priority class. January 2001 Aggressive VDCF No packet drops for high-priority class. IFQ drops constrained to lower classes during overload. Greg Chesson, et al, Atheros Communications

January 2001 Latency experienced by a selected high-priority stream (first 60 seconds) Greg Chesson, et al, Atheros Communications

VDCF can differentiate latency using CO. January 2001 VDCF can differentiate latency using CO. Greg Chesson, et al, Atheros Communications

PIFS Add two sources using PIFS rather than DIFS January 2001 PIFS Add two sources using PIFS rather than DIFS Each generates 1504 byte packets every 30ms Greg Chesson, et al, Atheros Communications

X-High class has low latency even in overload January 2001 X-High class has low latency even in overload (need zoomed-in view to see) Greg Chesson, et al, Atheros Communications

X-High class sees preferential Access at all times. January 2001 X-High class sees preferential Access at all times. Greg Chesson, et al, Atheros Communications

Conclusions VDCF provides differentiated bandwidth January 2001 Conclusions VDCF provides differentiated bandwidth using conservative parameters Even in overload situation CO can influence latency differentiation between classes Constant CO/CW settings for each 90 second simulation: Default (conservative) settings are useful adaptation can improve performance Real-time adaptation probably unnecessary with VDCF PIFS experiment demonstrates value for AP For polling, other preferential access Greg Chesson, et al, Atheros Communications