**********************************************************

Title: Caching Algorithms for Network Buffers

Speaker: Sundar Iyer, Cisco Systems
1:00pm, Talk, CIS X Aud

Abstract:

Networking systems have significant memory requirements in the data
plane.  Packets must be buffered, queued, routed, counted, monitored
etc. A typical high performance line card has 40+ Gbit/s of traffic
traversing, which can represent around 80M packets/sec. In total, there
may be as many as 10-30 read and write transactions for each packet.

The challenge is that there is no memory technology which is optimum for
these requirements. SRAM is fast, but expensive per Mbit. DRAM is low
cost, but very slow. The current solution is to have the packet
processor support many wide memory interfaces with multiple memory
technologies and over speed the memory sub-system to overcome memory
conflicts. This is expensive, increases complexity, power and space
requirements.

At Nemo (and now Cisco), we have been developing memory sub-systems to
alleviate this problem for different networking applications. The main
idea is to use a caching hierarchy, similar to that seen in Computer
Architecture. However the challenges are different. In a sense the
problem is easier because networking systems have well defined data
structures as compared to those seen in Computer Architecture. In
another sense it is harder because networking caches cannot afford
misses, since it may result in packet losses.

In this talk, I will focus only on packet buffering for networking
systems. I will describe some of the peculiarities of buffering
requirements for different market segments, and describe three different
but related algorithms, which utilize the constraints on memory access
patterns, to build memory sub-systems which are both efficient and lower
cost.

Bio:

Sundar Iyer is currently a technical leader in the Internet Switching
Business Unit at Cisco Systems. He defended his PhD in Computer Science
from Stanford in 2003. In the fall of 2003, Sundar co-founded Nemo
Systems, where he was the CTO and principal architect. Nemo, which is
now part of Cisco, specialized in memory algorithms for networking
sub-systems. Previously, Sundar was a founding member and senior
architect at SwitchOn Networks (since acquired by PMC-Sierra in 2001)
where he helped develop algorithms for deep packet classification.
Sundar has a Master's degree from Stanford in 2000, and a Bachelor's
degree in Computer Science from IIT Bombay in 1998.