Improving Linux resource control using CKRM

Author(s)

---------

Shailabh Nagar

Rik Van Riel

Hubertus Franke

Chandra Seetharaman

Vivek Kashyap

Haoqiang Zheng

Affiliations

------------

IBM T.J. Watson Research Center

Red Hat, Inc.

IBM Linux Technology Center

Proposal

--------

With the 2.6 kernel having addressed a number of scalability bottlenecks, manageability is one of the few remaining challenges for Linux as an enterprise operating system. The current kernel primarily enforces physical resource consumption only through per-task or per-address space limits. The mapping of tasks and address spaces to workloads is too dynamic to allow workload managers to efficiently allocate resources for the latter using controls on the former.

Last year, we had presented an approach to address the problem using the Class-based Kernel Resource Management (CKRM) framework and related schedulers for CPU, memory, I/O and inbound network. CKRM allows users (or workload management middleware) to dynamically define kernel recognized groups of tasks called classes. Each class can define shares for each resource whose scheduler has been made class-aware. Resource usage monitoring can also be done on a per-class basis. Using a classification engine module to manage the workload-class mapping, workload managers can now control and monitor workloads better. Even without classification engines, users can manually define classes to restrict resource consumption of arbitrary groups of less important jobs such as disk backup.

The CKRM framework and API have undergone a complete revision since last year. We now uses a virtual filesystem called rcfs (resource control filesystem) to create classes, manage shares and obtain usage statistics. Shares have upper and lower bounds called guarantees and limits. Classes can be defined hierarchically to allow users/sysadmins to manage their resource allocations using classes but without root privileges. Finally, inbound network control is fully integrated into the CKRM framework. All these changes have implications on the design of the class-aware CPU, memory, I/O and network schedulers.

In the presentation, we propose to describe the design and implementation of the new framework, API and any of the resource schedulers that have been ported over to the new API. Performance evaluation of the framework and the schedulers will also be included. CKRM is being proposed as a 2.7 kernel feature and will be of interest to all kernel developers.

More details on CKRM can be found at http://ckrm.sourceforge.net