Utility Computing - University of California, Berkeley

Utility Computing Amin Vahdat Internet Systems and Storage Group http://issg.cs.duke.edu [email protected] OPUS: An Overlay Utility Service Peering Overlay node App demand Dynamically allocate resources to competing services Based on changing application and network characteristics, SLAs Create topology based app requirements Bandiwdth, latency, loss rate, cost ($) sensitivity

Current Questions What are the proper SLAs for specifying utility for competing applications Allocate resources among competing applications based on changing network/application conditions Overlays that dynamically trade performance for cost Two-metric network optimization (NP-complete with centralized information) Scale to 10,000 nodes

Service overlays versus application overlays ModelNet Graph with edges augmented to include BW, latency, loss rates, etc. Edge Nodes Router Core Step 1: specify target widearea topology 100Mb Switch

Step 2: map topology to router core Graph partitioning Accuracy? Congestion? Gb Switch Step 3: run real code on edge nodes Subjected to specified network chars Open Questions and Status

How to introduce noise into emulation environment? Scalability versus accuracy How far can system scale? Target: 10 node core (gigabit interconnect), 100 edge nodes Current status Reproduced CFS results from SOSP 2001 paper Using RON network connectivity information Figures 6, 7, 8 4000-node gnutella network