Multipath TCP

Speaker Info:

Costin Raiciu (Lecturer, University Politehnica of Bucharest)
Christoph Paasch (PhD Student, Universite Catholique de Louvain)

Abstract:
Networks have become multipath: mobile devices have multiple radio interfaces, datacenters have redundant paths and multihoming is the norm for big server farms. Meanwhile, TCP is still only single-path and this makes it very difficult for endpoints to take advantage of network-level redundancy . There are numerous examples of problems caused by this mismatch: for instance, TCP connections are reset when a mobile device exits WiFi coverage despite the availability of 3G connections; and in datacenters random load-balancing often overloads a subset of links leaving parts of the network underutilized.

In this talk we will present Multipath TCP, an extension of TCP that is currently being standardized at the IETF. Multipath TCP enables unmodified applications that today use TCP to function over multiple paths and achieve better performance and better robustness. Multipath TCP requires no changes to the networks: it works just fine over the existing infrastructure.

We will discuss MPTCP's key design issues and the main deployment scenarios where we envision MPTCP can make a difference today. We will also highlight the main challenges we encountered while implementing MPTCP in the LInux Kernel, present a few experimental results and (conditions permitting) will do a live demo

Slides: https://docs.google.com/presentation/...

• Category

  Science & Technology

[Orbofs]

That guy in plaid needs to STOP ASKING QUESTIONS EVERY DAMN MINUTE.

[Olivier Bonaventure]

From a protocol design viewpoint, SCTP is cleaner, but SCTP does not go through the various types of middleboxes that exist in today's Internet and it requires changes to the applications. Mutlipath TCP works with existing applications in today's Internet.

[svommams566]

Is youtube.com/watch?v=iqx1cShqH6­I the video you have in mind?

[svommams566]

The spanning tree is operating at a lower layer. What will happen at the MPTCP layer will depend on what the spanning tree protocol will do at the lower layer. It may be the spanning tree protocol simply disables redundant link, in that case you obviously won't get any performance benefit from those links. You may get better performance by using IP rather than Ethernet, since IP has better support for alternative paths.

[svommams566]

Three reasons: Aggregated Ethernet will still send all packets from a TCP connection over just one of the aggregated links, because TCP doesn't deal well with reordering. If you are using different carriers you need to use different IPs for packets you send due to filtering of spoofed IP addresses. And to receive packets on both interfaces, they must be send to your different IP addresses.

[Tim Hosking]

Interesting subject, but I lost patience when the audience were allowed to hijack the session with inaudible questions which should have been left until the end.

[Ravikumar.T naidu]

I am unable to understand why all this is needed at all ? why cant we have the tcp stack run on aggregated ethernet ?

[doug65536]

Another unintelligible Google talk. Speakers should know that saying 5 words per second defeats the entire purpose of speaking at all.

[adhitya akbar]

I thinks it should turning up multiple interface...

[adhitya akbar]

is there any effects with link-aggregation with multiple interface? o.o

[Christoph Paasch]

This per-packet load-balancing is almost never used, because it introduces reordering of the packets. If you have reordering, TCP's performance will suffer a lot because 3 duplicate ACKs trigger a fast retransmit and the congestion window will be divided by two.

[hyperthreaded]

I don't get it. TCP is "single-path"? Isn't TCP done in the endpoints only? The routers don't know about TCP, and if you have some dynamic, load-based routing scheme in place in your network, each packet of the connection can use a different path.