--- Begin Message ---
ET-I3 Seminar
Wide Area Performance Monitoring Using Aggregate Latency Profiles
Prof. Louiqa Raschid
University of Maryland
College Park
Venue: BY 510; 10AM, March 12th (Friday)
A challenge in supporting Wide Area Applications (WAA) is that of scalable
performance management. In particular, the unpredictable behavior of a
dynamic WAN results in a considerable variability of latencies, which calls
for models to predict the access latency (end-to-end delay) between a client
and a server. Networking research has had significant success in developing
models to predict latency, based on physical topologies and grounded in
network protocols and behavior. However, such solutions are not designed to
scale to the continuous monitoring and maintenance required for end-to-end
support of millions of clients. Individual Latency Profiles (iLPs) were
proposed in the literature to capture latency distributions experienced by
clients when connecting to a server; it is a passive measurement made by
client applications and is gathered on a continuous basis. We propose a
technique for managing millions of iLPs, and aggregating them into aggregate
Latency Profiles (aLPs). We use measures such as mutual information and
correlation to compare the similarity of pairs of iLPs and we propose
Relevance Networks to manage a large collection of iLPs and to group them into
aLPs. The proposed method can enhance network based models, and has the
advantage of not requiring explicit knowledge of topology and network
characteristics. We have conducted experiments demonstrating that aLPs can
benefit from the considerable non-random associations between iLPs, to improve
their ability to predict latencies.
---
--- End Message ---