idash
TRANSCRIPT
• HTTP-based delivery for VoD,• Congestion in Cache Feeder/ Access Link,• Connection Characteristics vary over time, • Dynamically adapt the requested video quality, • Scalable Video Coding for DASH,
Introduction
• HTTP/TCP over RTP/UDP,✓ Immune to firewall and NAT traversal issues, ✓ Substantially relieve the load on the video server by
using HTTP cache,
• DASH,✓ Content in various versions, ✓ To alleviate the congestion, switching occurs, ✓ Segmenting each version in chunks, ✓ Multiple Representation of each video(H.264/AVC),✓ All reps in one file (SVC)
Introduction
• Minimizing the transmitted data,
• Relieving the server,
How to select every moment what is in the cache?
Related Work
• Most algorithms rely on,✓ LRU,
✓ LFU,
✓ Combination thereof,
• In this work, ✓ Chunks to be consumed predicted,
o User playing nth chunk will play n+kth chunk of the same video k time instants later!!
Related Work
• Serving a great number of users w/ a single bit stream,
• Coping w/ congestion by applying on-the-fly adaptation,
• OP refers to a valid sub-stream at a certain quality level and a corresponding bit rate,
How to encode multiple quality layers?
• CGS or MGS,o CGS, the number of OPs is the same as the number of
layers,
o MGS, switching on a NAL unit basis, much higher number of possible Ops,
Efficient Encoding w/ SVC
• So many layers => so many Ops => coding overhead!!,
• Encode a reduced number of quality layers w/ MGS
Efficient Encoding w/ SVC
Different OPs are obtained by dropping enhancement layer packets!
• Dropping Q2(blue) in T2& T3 levels,
• Dropping the Q2 of all temporal levels and pictures from Q1(red) of T2& T3 levels,
Efficient Encoding w/ SVC
OPS must be selected in such a way that users experience smooth quality degradation!
• Sims are based on real data(Observation of deployed VoD service),
• 1 month,• More than 500 movies offered,• Average 3400 reqs per day,
Model
• Caching Algorithms,
➢ LRU,➢ Most recently requested chunks are kept,
➢ CC,➢ Number of guaranteed hits of chunks, LRU and LFU
combined,
➢ SVC might score more cache-hit-ratio as opposed to AVC,
Model
• Congestion in cache feeder?➢ @ peak hours, cache misses and updates are
higher!➢ Current infra. Not sufficient, congestion occurs,
Data received with delay, thus ramping down,Waiting timer set after first ramp-down, if equal,
ramps up,
• Congestion in access links?
Model
• Congestion in cache feeder?
• Congestion in access links?
Markov Chain, to describe transition probabilities, utilized!
Model
• Cache-hit-ratio,➢ Percentage that objects can be served from the
cache,• Not downgrading clients,
➢ Percentage…,• D/R
➢ Desired/ Received rate,
Performance Targets
ResultsCase 1: DASH Clients requesting too many data in peak hours,
Case 2: Cross-traffic due to DASH clients sharing resources,
•Videos comprised of 90 mins and chunks of 10 secs length,
ResultsCase 1: DASH Clients requesting too many data in peak hours,
CHR,7 % up for low cache cap,
ND,lower coding efficiency,
D/R,Reutilization of already cached video content,
Case 2: Cross-traffic due to DASH clients sharing resources,
ResultsCase 1: DASH Clients requesting too many data in peak hours,
Case 2: Cross-traffic due to DASH clients sharing resources,
Different versions of the req videosspoil the cache!