1 / 12

ShapeShifter: Scalable, Adaptive End-System Multicast

ShapeShifter: Scalable, Adaptive End-System Multicast. John Byers, Jeffrey Considine, Nicholas Eskelinen, Stanislav Rost, Dmitriy Zavin Listed alphabetically. Problem. Problem: efficient delivery of popular bulk content Existing Approaches: Single-source unicast

betty
Download Presentation

ShapeShifter: Scalable, Adaptive End-System Multicast

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. ShapeShifter: Scalable, Adaptive End-System Multicast John Byers, Jeffrey Considine, Nicholas Eskelinen, Stanislav Rost, Dmitriy Zavin Listed alphabetically

  2. Problem • Problem: efficient delivery of popular bulk content • Existing Approaches: • Single-source unicast • Forward caching/Content Delivery Networks • Reliable IP Multicast • End-system multicast • Our approach: • Improving end-system multicast through the use of forward error correction and better topologies

  3. Network-supported (IP) Multicast • Optimal solution: duplicates and disseminates data only when necessary • Relies on network support: in the real world, IP Multicast lacks deployment • Scalability concerns: per-group accounting and topology managementdo not scale due to limited router resources • Reliability: many proposals, few solutions

  4. End-System Multicast • Does not rely on network support: builds and manages a virtual, overlay topology of unicast links on top of the network’s physical topology • Flexibility: optimization of the tree according to a richer set of metrics (perhaps specified by the application), ability to change topology on-demand • Improved scalability: end-systems are richer than routers in terms of dedicated resources • Problems: increased network resource requirements compared to IP Multicast, non-optimal mapping of the virtual topology onto physical topology

  5. Related Work • Narada/End-System Multicast: Build and maintain a mesh of low-latency unicast links and use its minimal spanning tree for distribution. Also showed costs relative to IP Multicast are not excessive. • Overcast: A core group of well-placed nodes uses end-system multicast to distribute bulk content internally, in order to eventually provide it to clients. A node chooses a parent based on bandwidth through the candidate nodes using the number of network hops as a tie breaker.

  6. Improvements in ShapeShifter • Erasure codes: improved overlay management, more connected graph structure, increased adaptivity • Scalable group management: a node need only be aware of a small portion of the graph but achieves coverage through continuous discovery • Measurements: metrics crucial to optimization of the overlay graph, such as shared-link congestion (refer to Khaled’s presentation)

  7. Forward Error Correcting Codes • FEC codes: a well-known solution to dealing with packet loss without using feedback – instead of retransmitting packets, redundant packets are sent combining the original packets to recover from losses. e.g. x, y, x+y, a, b, a+b+x • Efficient codes: instead of the traditional slow Reed-Solomon codes, we use a variant of Tornado codes. This allows fast decoding while only requiring a small number of extra packets. • Strategy: the original server sends out FEC packets along the end-system multicast graph (à la Digital Fountain).

  8. Recoding • Problem • Correlation: client nodes may have a high degree of correlation in information received due to common sources • Duplication: given correlation, duplicate packets received from client nodes can be ineffectual • Solution • Recoding: nodes blend received packets to generate new, high utility symbols for other nodes • Beyond trees: recoding allows non-tree topologies since duplication is avoided

  9. Uncorrelated Loss Compensation • A neighbor node with greater throughput may supplement the flow of content to another node, circumventing the problematic physical link. Loss Rate5% Loss Rate30%

  10. Download from Multiple Nodes in Parallel Well-connected newcomer scenario 1 MB/s 1 MB/s 1 MB/s 1 MB/s Non-uniform dissemination of content

  11. Resilience To Partitioning Collaborative Reconstruction Partition avoidance through discovery and awareness of multiple candidates

  12. Future Work • Implementation underway • More analysis of • Codes and correlation • Graph management • Security issues • Testing • WAN, simulated and real • Mobile environments • Extensions to streaming media

More Related