elastic - dash

26
ELASTIC Client Side Controller for DASH Yigit UNALLAR

Upload: yigit-unallar

Post on 05-Apr-2017

22 views

Category:

Engineering


0 download

TRANSCRIPT

Page 1: Elastic - DASH

ELASTICClient Side Controller for DASH

Yigit UNALLAR

Page 2: Elastic - DASH

DESIGN DRAWBACKS:

• QoE,• On-off traffic pattern;

• Unfairness,• Underutilization of bottleneck,

Page 3: Elastic - DASH

PROPOSED SOLUTION:

• Feedback Linearization Adaptive Streaming Controller,

• Feedback Control Theory,• Controlled Testbed;

• BW cap. is set,• Delays are set.

Page 4: Elastic - DASH

COMPARISON CRITERIA:

• Fully bottleneck utilization,• Fair bottleneck share, • Fairly share w/ TCP Greedy flows.

Page 5: Elastic - DASH

INTRODUCTION:• Client-side Algorithms,

• On-off pattern at steady state,• Unfairness sharing a bottleneck,

• 3 Popular Services suffer,• Unable to get fair share w/ TCP Greedy Flow,

“Downward Spiral Effect”• Algorithm w/o on-off traffic pattern, “downward spiral

effect”, large player buffers and increased segment size proposed!

Page 6: Elastic - DASH

APPROACH:• ELASTIC proposed,• Performance comparison w/ Conventional, PANDA&

ELASTIC,• Many videos, same algorithm, same bottleneck,• Many videos, TCP greedy flow, same bottleneck,

Page 7: Elastic - DASH

RELATED WORK:

• Architecture;• Video segments of fixed duration τ ,

• Playout buffer full ➔Steady- state,• HTTP requests each τ sec. to maintain,

• On-off patterns occur during download& idle modes,

• 3 Commercial services suffer from unfair share due to on-off

traffic pattern!!

Page 8: Elastic - DASH

ELASTIC:

• Design Requirements:• Dynamically select video level l(t) to achieve QoE,• But what impairs QoE?

• Re-buffering when buffer gets empty,• Frequent quality switches,

• Video level w/ higher bitrate and quality desired!

Page 9: Elastic - DASH

ELASTIC:

• Design Goals:• Low re-buffering ratio,• High obtained video level,• Video level switches, • Fair sharing.

Page 10: Elastic - DASH
Page 11: Elastic - DASH

CONTROLLER:

• Conventional Approach;• Throttle l(t) to match the measured available b(t),• Control q(t) by regulating idle periods,

• However, Elastic proposes;• Throttle l(t) to drive q(t) to a set-point qT,

• No idle state btwn 2 downloads,

• q(t)➔qt , l(t) // b(t)

Page 12: Elastic - DASH

Employ feedback linearization to (1) and steer q(t) to qT, following dynamics imposed;

Page 13: Elastic - DASH

IMPLEMENTATION:

“Where kp=1/100,

ki=1/1000.”

Page 14: Elastic - DASH

TESTBED:

Page 15: Elastic - DASH

SCENARIOS& METRICS:• S1:• Video + TCP,• b = 4Mbps,• ”Dynamic behavior of algorithms.”

• S2:• Many Video + Many TCP,• b = 40Mbps,• “Comparing performances employing Table II metrics.”

Page 16: Elastic - DASH

SCENARIOS& METRICS:

Page 17: Elastic - DASH

RESULTS:• S1:

Page 18: Elastic - DASH

RESULTS:• S1: “TCP flow starts at t=100s.”

“As seen, ELASTIC does not produce an on-off traffic, thus behaving as a TCP flow!”

Page 19: Elastic - DASH

RESULTS:• S1:

Page 20: Elastic - DASH

RESULTS:• S2: “TCP flow starts at t=100s, for each N the fair share can be computed as

F=40Mbps/N.”

Page 21: Elastic - DASH

RESULTS:

• S2: NTCP=0,

Page 22: Elastic - DASH

RESULTS:

• S2: NTCP != 0,

Page 23: Elastic - DASH

RESULTS:

• S2: NTCP != 0,

Page 24: Elastic - DASH

RESULTS:

• S2: NTCP != 0, N<50 =>ratio < 1%

Page 25: Elastic - DASH

CONCLUSIONS:• Identical traffic pattern to any long-lived TCP flow accomplished.

• All algorithms performed well in the absence of TCP flows.

• in S2:• Fair share obtained,• Video continuously reproduced,• Video level switches mitigated.

Page 26: Elastic - DASH

THANK YOU…