Title: Building High Performance Scalable Infrastructure Services through
Analysis-driven Caching on Structured Networks

Speaker: Venu Ramasubramanian 

Abstract:

Caching is widely used in distributed systems to improve performance, balance
load, and tolerate failures.  Traditionally, caching has been performed in an
ad-hoc manner, where nodes opportunistically cache objects that they
encounter, use heuristics for cache management, and provide loose consistency
semantics.  Several studies on caching show that opportunistic caching is
effective only when a few popular objects receive a large portion of the
queries and objects are immutable or change occasionally.  Yet, a large
number of critical applications, including the Domain Name System (DNS) and
the Web, are characterized by heavy-tail popularity distribution and mutable
objects.

In this talk, I will describe a novel form of caching for structured
overlays, where objects are proactively cached based on an analysis of
performance trade-offs.  Analysis-driven caching poses the fundamental
trade-off between performance and overhead as an optimization problem,
determines the optimal caching strategy through numerical algorithms, and
replicates objects using inexpensive distributed protocols.  It enables
structured overlays to achieve an order of magnitude improvement in lookup
performance, from O(log N) to O(1), with minimal network and storage
overhead, and efficiently adapt to large upheavals in the popularity of
objects, which arise during flash-crowds, by automatically spreading the load
across the distributed system.

I will show the effectiveness of analysis-driven caching for building
performance-demanding infrastructure services by describing CoDoNS
(Cooperative Domain Name System), a new scalable name service for the
Internet.  CoDoNS combines self-organizing overlays and analysis-driven
caching to provide high resilience to failures and denial-of-service attacks,
low latencies for DNS queries, and fast propagation of changes to DNS
bindings.

-----------------------------------------------------------------