Summary: TRAID

This paper essentially proposes a content routing design based on
name-based routing. Conventional content routing distributes content
delivery but does not effectively distribute content discovery. The paper
argues that there are two basic problems with conventional content
routing. Firstly, their design does not scale well because of long round
trip incurred due to a centralized name server whenever there is a cache
miss. Secondly, the proprietary nature of their schemes makes them
undesirable for global use.

The paper assumes that content routing problem is similar to routing
problem. The main idea of the paper is to have a network based scheme for
content discovery through content routers. Content routers provide content
server information to their clients just like the way IP routers provide
routing service for their clients. Clients ask for specific content and
replicated servers can be viewed as alternate routes through which
information can be accessed. Content routers distribute, maintain, and
make use of information about content reachability. Thus these content
routers participate in both IP routing as well as name-based routing.

Content lookup is accomplished by INRP protocol. Clients initiate a
request to the nearby local content router by providing server portion of
the URL. Content routers maintain the set of name-to-next hop mappings.
Content routing performs routing by name, similar to BGP, using Name Based
Routing Protocol (NBRP). Like in BGP, name based routing advertisement
essentially contains content name, next hop towards a content server
address or a content router through content can be accessed. Routing
advertisement may also include server load but the paper does not provide
detailed discussion as to how this can be made scalable. The key benefits
of this approach include low-latency and its ability to make INRP and NBRP
open protocols.

The paper proposes various schemes to address the scalability of NBRP
scheme. First, to handle large number of names, NBRP uses explicit
aggregation - where it combines collections of name suffixes that map to
same routing information into routing aggregates. Second, it makes use of
version number to detect changes in the contents and send only .Delta.
aggregate instead of sending large aggregates each time. Third, all names
in the routing table are treated identically. (accomplished by mirroring
and indirection). Fourth, the number of routing entries is comparable with
best possible with IP routing.

One important challenge with NBRP is that addition of new names. Addition
of new names is much more common than in addition of new BGP prefixes. The
paper argues that this is not a problem with some preliminary explanation
but I thought it is worth to have a detailed discussion on this.

Simulation results are encouraging as well. There is an overhead of 0.5
milliseconds for going through a single hop of content router. Further,
they show that INRP provides better performance than contacting root name
servers. INRP reduces the latency by 86 to 95% and also eliminates the
variability in latency. Finally, I found that throughput declines
significantly as number of routing peers increases. (1050 updates/sec with
one peer to 370 updates/sec for six peers).

Overall, the idea seemed pretty good. The motivation of the approach is
very clear i.e. to avoid latency to the DNS root server. The paper also
provided sufficient discussion about the benefits of this scheme over
existing content discovery schemes. However, the deployment is still a
concern because of need to upgrade routers to support content routing
approach.
Received on Thu Nov 30 2006 - 10:40:23 EST