vCache: Supporting Cost-Efficient Adaptive Bitrate Streaming

Motivation

Agenda

• Adaptive bitrate (ABR) streaming
• vCache
• System Design
• Workflow​

reference: https://ieeexplore.ieee.org/document/7999149

Adaptive bitrate (ABR) streaming

Reality

• limited bandwidth capacity
• unstable network conditions
• diverse viewing devices

Result

• growing demand for video traffic

• low quality of viewing experiences

ABR

ABR Streaming

Each video is transcoded into multiple representations and then cached in streaming servers.

ABR Streaming

A media presentation description (MPD) file is required to manifest the available representations for a video.

ABR Streaming

When starting a video session, the video player

1. obtains the MPD file
2. selects the best possible quality representations according to the current network conditions and device capacity

Problem -- Operational Costs↑

Transcoding

• compute-intensive
• consumes tremendous resources

Caching multiple representations

• requires more storage space
  • a few of video chunks are frequently requested

vCache

Physically Cached

the video chunk is cached ➜ can be directly read from storage

Virtually Cached

only the metadata of the video chunk is cached

ABR streaming with vCache

The MPD files are cached in streaming servers to manifest the available representations of each video.

ABR streaming with vCache

When starting a new session, the video player

1. obtains the MPD file
2. requests video chunks of the appropriate representations

ABR streaming with vCache

The streaming server will

read the requested video chunks in vCache

(dynamically cached / transcoded on the fly).

Difference

1. not be pre-transcoded

2. vCache manages video chunks dynamically

What can vCache do

1. dynamically provision resources

    transcoding delays <= acceptable range

2. strike a tradeoff between storage and computing costs

    ➜ reduce the operational costs of ABR streaming

System Design

Framework

Resource

virtualization

• vCache is based on network function virtualization (NFV)
• provides the capacity of underlying hardware
  • virtual computing
  • storage
  • network resources

Request interface

receives a request ➜ locates the video chunk

• physically cached
  • directly read and delivered
• Otherwise
  • initiate a transcoding request
    • transcode the video chunk of the source version into the user-requested representation
    • the transcoded video chunk will be rendered to the streaming server

Cache management

• analyzes the request information

• dynamically determines whether a video chunk should be physically or virtually cached

• Goal: reduce the overall cost

Transcoding

management

• A virtually cached video chunk will be transcoded on the fly when being requested ➜ delays

• dynamically provisions resources 

  • transcoding workload is time varying

• ensure that transcoding delays <= an acceptable range

Workflow

vCache receives a request for a video chunk

➜ looks up the requested video chunk

➜ reads and delivered / sends a transcoding request

The dispatcher equally dispatches transcoding requests among the active virtual machines (VMs) for load balancing.

A transcoding server will transcode the requested video chunk on the fly when receiving a request.