ACM Tech Talk – November 2008 Edition
ACM November 2008 edition had two talks:
Dr Ravi Gudi from Honeywell research gave a talk on Decomposition Paradigms for Large Scale Systems. Dr Ravi Gudi has studied at IIT Bombay and is currently working with the research team at Honeywell Technology Solutions. Using the example of building a chemical refinery and a railway workyard, he explained how to design and manage a large scale system by breaking it down into smaller parts using various techniques. Two of the techniques for decomposition that he talked about are model co-ordination and goal co-ordination. He talked about optimisation problems such as placements of sensors in an area to collect data and transmit it back effectively. The other problem he talked about is communication and feedback while accounting for the delay in transmission. There has been some interesting research in this area at CMU. Some of these fundamental problems see applications in social networks, scheduling in multi-core systems and building Internet scale distributed systems such as CDNs. For example the cascade algorithm invented at CMU can be applied for ranking blogs by number of unique breaking stories. It tells you which blogs should you follow to get the maximum amount of information with the least amount of effort. The same Cascades algorithm can also be used for getting optimal placment of sensors at least cost. Communication in Large scale systems is also facilitated by techniques such as the gossip protocol. Amazon uses the gossip protocol for communication between it’s machines in it’s AWS datacenters. However this communication method overloaded Amazon’s servers and led to a outage.
Prakash Venkatraman, Senior Architect at Oracle talked about using Signature matching and rule-based matching systems for diagnosing faults in systems (root cause analysis). Taking the example of Oracle Support system, he explained how by looking at the error logs and mointoring different parameters, you can build and deduce rules which can be used for fault analysis. Using Bayesian probability and using statistical co-relation techniques such as that used in filtering spam, the whole process of doing fault analysis can be automated to make the work of humans easier.