Click here for full text:
Controllable Fair Queuing for Service Utilities
Karlsson, Magnus; Karamanolis, Christos; Chasse, Jeff
HPL-2005-21
Keyword(s): fair queuing; closed-loop systems; utility computing
Abstract: Computing and storage utilities must control resource usage to meet contractual performance targets for hosted customers under dynamic conditions, including flash crowds and unexpected resource failures. This paper explores properties of proportional share resource schedulers that are necessary for stability and responsiveness under feedback control. It shows that the fairness properties commonly defined for proportional share schedulers using Weighted Fair Queuing (WFQ) are not preserved across changes to the relative weights of competing request flows. As a result, conventional WFQ schedulers are not controllable by a resource controller that adapts by adjusting the weights. The paper defines controllable fairness properties, presents an algorithm to adjust any WFQ schedule when the weights change, and proves that the algorithm results in controllable-fair WFQ schedulers. The analytic results are confirmed by experimental evaluation using a three-tier Web service and a prototype controllable-fair scheduler called C- SFQ(D). C-SFQ(D) extends depth-controlled Start-time Fair Queuing, which supports interposed proportional sharing for complex network services with no internal support for differentiated service quality. The prototype includes an adaptive control system that adjusts the flow weights on the C-SFQ(D) scheduler to meet latency and throughput targets under a variety of conditions. Experimental results demonstrate the importance of controllable-fair scheduling for feedback control of network services. Notes:
12 Pages
Back to Index
| 
Printable version
