Summary: SRM Framework

Summary: SRM Framework

The goal of the paper is present a new SRM framework for scalable and
reliable multicast.

SRM is designed to meet minimal definition of reliable multicast by
delivering data to all members in the group. It defers rest of the
reliability issues to the specific applications. SRM is highly based on the
group delivery model (similar to IP multicast) and enhances the multicast
group by maximizing information and data sharing among all the members.
Since sender based control is not scalable, SRM makes use of receiver-based
control as the building block for reliable multicast. Primarily, SRM model
is based on two assumptions: 1) data has unique and persistent names and 2)
application naming conventions allow imposing hierarchy over name space.

Whenever SRM member generates data, the data is multicast to the group. Each
group member is responsible for detecting the loss and requesting
retransmission. Each member multicasts low-rate periodic session messages to
communicate the highest sequence number for the current page. These session
packet also contain timestamps that are used to estimate host-to-host
distances needed by the repair algorithm. SRM's loss recovery algorithm
provides the foundation for reliable delivery. According to loss recovery
algorithm, member who detects a loss wait a random time and then multicast
their repair request along with data name to suppress requests from other
members sharing the same loss. Each host uses two timers backed-off timer
and repair timer as part of request algorithm. As a security enhancement,
each packet communicated could be accompanied by a tag that not only
authenticates the source of the data but also ensures integrity.

Due to probabilistic nature of the lost recovery algorithms, two or many
hosts can generate request for the same data roughly at the same time (this
leads to duplicates). The goal of request/repair time algorithms is to keep
the number of duplicates low. An important observation is that performance
of request/repair algorithms for loss recovery depends on the underlying
network topology. Simulations suggest that chain topology, deterministic
suppression (suppress requests/repairs further down the chain) works well
whereas for start topology, randomization works effectively. For large tree
topologies, combination of deterministic suppression and probabilistic
suppression works effectively.

The final sections of the paper focus on developing adaptive adjustments of
random time algorithms. Adaptive algorithm adjusts the timer parameters as
function of both delay and number of duplicate requests/repairs in the
recent loss recovery exchanges. So, the idea here is to adjust the timer
parameters based on the results from previous rounds. The simulation results
also confirm that adaptive algorithms quickly reduce the average number of
repairs reaching steady state quickly.

Adaptive algorithms also make use of local recovery algorithms instead of
just relying on the SRM's global loss recovery algorithm that sends
request/repair packets to every one in the group even if packet is dropped
on a link to a single member. This local recovery becomes necessary
condition as the multicast group grows. The key observation is that member
should send a packet with local scope when recent losses have been confined
to a single loss neighborhood. If no repair is received before back off
timer expires, then next request can be sent with global scope. The paper
also provides preliminary information three types of scoping local recovery:
administrative scoping, separate multicast groups, and TTL-based scoping.

Finally, the paper summarizes the key challenges in the future work related
to scalable session messages and local recovery.
Received on Tue Oct 31 2006 - 06:22:02 EST