Wide area traffic: the failure of Poisson modeling
IEEE/ACM Transactions on Networking (TON)
Modelling extremal events: for insurance and finance
Modelling extremal events: for insurance and finance
Exploiting process lifetime distributions for dynamic load balancing
ACM Transactions on Computer Systems (TOCS)
Self-similarity in World Wide Web traffic: evidence and possible causes
IEEE/ACM Transactions on Networking (TON)
SIGMETRICS '98/PERFORMANCE '98 Proceedings of the 1998 ACM SIGMETRICS joint international conference on Measurement and modeling of computer systems
Heavy-tailed probability distributions in the World Wide Web
A practical guide to heavy tails
On choosing a task assignment policy for a distributed server system
Journal of Parallel and Distributed Computing - Special issue on software support for distributed computing
Analytic modeling of load balancing policies for tasks with heavy-tailed distributions
Proceedings of the 2nd international workshop on Software and performance
Task assignment with unknown duration
Journal of the ACM (JACM)
EQUILOAD: a load balancing policy for clustered web servers
Performance Evaluation
Dynamic Load Balancing on Web-Server Systems
IEEE Internet Computing
Design and Evaluation of Effective Load Sharing in Distributed Real-Time Systems
IEEE Transactions on Parallel and Distributed Systems
HPDC '00 Proceedings of the 9th IEEE International Symposium on High Performance Distributed Computing
Dynamic Scheduling of Web Server Cluster
ICPADS '02 Proceedings of the 9th International Conference on Parallel and Distributed Systems
A least flow-time first load sharing approach for distributed server farm
Journal of Parallel and Distributed Computing
Deferred Assignment Scheduling in Cluster-Based Servers
Cluster Computing
Optimal state-free, size-aware dispatching for heterogeneous M/G/-type systems
Performance Evaluation - Performance 2005
Analysis of size interval task assignment policies
ACM SIGMETRICS Performance Evaluation Review
ICPADS '08 Proceedings of the 2008 14th IEEE International Conference on Parallel and Distributed Systems
Performance Evaluation
Hi-index | 0.01 |
Server farms, consisting of a collection of hosts and a front-end router that dispatches incoming jobs to hosts, are now commonplace. It is well known that when job service requirements (job sizes) are highly variable, then the Size-Interval task assignment policy is an excellent rule for assigning jobs to hosts, since it provides isolation for short jobs by directing short jobs to one host's queue and long jobs to another host's queue. What is not understood is how to classify a “short” job versus a “long” job. For a long time it was believed that the size cutoff separating “short” jobs from “long” ones should be chosen to balance the load at the hosts in the server farm. However, recent literature has provided empirical evidence that load balancing is not always optimal for minimizing mean response time. This article provides the first analytical criteria for when it is preferable to unbalance load between two hosts using Size-Interval task assignment and in which direction the load should be unbalanced. Some very simple sufficient criteria are provided under which we prove that the short job host should be underloaded, and likewise for the long job host. These criteria are then used to prove that the direction of load imbalance depends on moment index properties related to the job size distribution. For example, under the Bounded Pareto (BP) job size distribution with parameter α and a sufficiently high upper bound (the BP is well known to be a good model of empirical computer system workloads), we show that α determines the direction of load imbalance. For α1, the long job host should be underloaded. Many other job size distributions are considered as well. We end by showing that load unbalancing can have a dramatic impact on performance, reducing mean response time by an order of magnitude compared to load balancing in many common cases.