2003 Kernel Summit: Resource management

[Posted July 22, 2003 by corbet]

This article is part of LWN's 2003 Kernel Developers' Summit coverage.

Ken Rosendal talked briefly about resource and workload management. He started with some definitions of terms: resource management is control of resources available to applications which affect performance: memory, CPUs, network usage, block I/O usage, etc. A workload is a set of processes working together to accomplish a single task. Workload associations can be dynamic; a relational database manager is part of a particular workload while processing that workload's requests, for example. Workload management is essentially the use of resource management to achieve a user's goals for a particular workload.

User goals can be expressed in a number of ways. Some users are interested in pure resource utilization; a particular workload, for example, should get 30% of the processor allocated to it. Strict percentage-based management schemes are most often used with batch jobs. Other management schemes are more concerned with response times; interactive job goals are often specified in these terms. Finally, some workload goals are given in terms of transaction delays.

True workload management brings some challenges. The user must be able to specify what a given workload is, for starters. Is it a certain application? Perhaps a database system is part of a given workload when it is handling transactions of a certain type. Once that has been done, the user can provide the goals for that workload. Management at this level requires application support; the system must be told when certain types of transactions begin and end, or when "user interaction events" begin and end.

No proposals for actually implementing any of this were given. There was some talk about whether most workload management should be handled in user-space, perhaps through the use of virtual machines like User-mode Linux.