7
7-11.6.08 @Snagstuhl Undiscombobulating Un- Flows. Jon Crowcroft, University of Cambridge Currently IMDEA Networks. [email protected] http://www.cl.cam.ac.uk/~jac22

Undiscombobulating Un-Flows

Embed Size (px)

DESCRIPTION

Undiscombobulating Un-Flows. Jon Crowcroft, University of Cambridge Currently IMDEA Networks. [email protected] http://www.cl.cam.ac.uk/~jac22. Cdn v. swarm – matrix?. What’s a Flow?. The idea of flow is being undone We’re splitting flows over multipath - PowerPoint PPT Presentation

Citation preview

Page 1: Undiscombobulating Un-Flows

7-11.6.08 @Snagstuhl

Undiscombobulating Un-Flows.

Jon Crowcroft, University of Cambridge

Currently IMDEA [email protected]

http://www.cl.cam.ac.uk/~jac22

Page 2: Undiscombobulating Un-Flows

Cdn v. swarm – matrix?

Page 3: Undiscombobulating Un-Flows

What’s a Flow?

The idea of flow is being undone We’re splitting flows over multipath We’re sourcing traffic in swarms At a point in the net, how do you tell?

Guilt by association used to be possible C.f. ASM – Any Source Multicast Flow ID was group address + Dst Port Need to recover this identity Eg. so net (re-)feedback works Need to work out what appropriate rate is.

Page 4: Undiscombobulating Un-Flows

Data Driven Demands

Next step in evolution of CDN Is to make the swarm a building block Still have higher level identifiers Meta-data, pub/sub/interest etc But no lower level id

Does statistical mux/drop in queue + Rate adjustment at edge Still add up to a sensible behaviour?

And what is the rate adjustment Given there’s more than one edge?

Page 5: Undiscombobulating Un-Flows

Lets dream we had ASM still

Could assign marks to a swarm Based on group Each receiver could assign weight to feedback And each source could adjust aggregate rate to

group to meet fair/efficient goal Question 1: Does this new “flow unit” still

aggregate the right way? Question 2: does it still improve over

multipath Question 3: mech. Do mechanisms to

assign flows to multipath subpaths still apply to this new type of flow unit

Page 6: Undiscombobulating Un-Flows

Some more small nits

How to do reliability (recall srm) We do now have fec and netcoding

as first class building blocks…so maybe ok

But in the presence of packet drop as a CC feedback signal…

…fec level to add isn’t totally obvious

Page 7: Undiscombobulating Un-Flows

Questions?

Just to make matters worse, All the flows are short lived.