Wednesday, August 14, 2013

SIGCOMM2013: Less Pain, Most of the Gain: Incrementally Deployable ICN

Presenter: Vyas Sekar

Co-authors: Seyed Kaveh Fayazbakhsh , Yin Liny, Amin Tootoonchian , Ali Ghodsiz, Teemu Koponen, Bruce M. Maggsy, K. C. Ng , Vyas Sekar , Scott Shenker

Information Centered Networking: Fetch a piece of content to the user based on what the content is, not where it is. ICN is inherently more secure, and lowers latency and congestion. However, ICN requires router upgrades so that maintain cached content, and upon miss, perform content based routing, instead of IP routing.

There are 2 main dimensions to consider in the design space for ICN: 1) Cache placement (do we cache objects on every router? only at edge routers?) and 2) Request routing: (do we choose the shortest path to origin? or route to the nearest replica?). Simulations using topologies from Rocketfuel, in order to evaluate the different points in the design space. The simulations, which use real CDN request logs and 5% object caching (LRU) show similar performance results (measured as a function of latency, congestion and server load) between 4 tested architectures that span the design space. (<10% difference on average/representative case, 17% on best case). Noticeably, simply adding storage at the edge provides a big performance boost.

The authors propose an edge-based caching architecture is incrementally deployable, which builds on the observation that, both ICN requirements (decoupling names from locations and tying objects to owners) can be achieved by building on existing HTTP extensions. In the proposed scheme, called idICN, a client requests an object based on its ID, which is received by a proxy edge cache, which, if it does not have the object stored, asks for the location of the object from a name resolution system. It then requests the object by address as usual from a reverse proxy, which fetches the object from the origin server, and sends it back to the client's edge proxy along with some metadata to be cached with.

In summary, the authors propose an architecture which, without changes to core internet routing, achieves ICN, dropping latency and congestion (17% performance boost in simulation), and adding security, just by adding changes at the edge of the network.