Mission-Aware Task Scheduling for Data Center Networks

Today’s modern data centers generate revenue mainly from interactive applications such as web search, social networking, advertisement, and recommendation systems. Despite the diversity of such applications, they have common many characteristics. For instance, studies from Google, Amazon and Microsoft show that when applications are delayed by a fraction of a second, the revenue will be decreased considerably. Delay is not the only factor affecting the revenue. Along with other factors it forces data center operators to use complex task scheduling mechanisms to manage the access of applications to network resources in order to increase the revenue. 

Recently, a variety of schemes have been proposed to perform task scheduling in data centers. However, a well-rounded scheme that can simultaneously support necessary requirements is lacking. These requirements include the support of task scheduling, deadlines, dependency, preemption, incremental deployment, and high scalability. The existing solutions implicitly assume that all of the tasks in a data center yield the same revenue, which may not be true. For instance, performing a web search and loading proper ads for a user are two different tasks and each yields a different amount of revenue. The service provider will generate revenue only when users view or click on the ads. Hence, the tasks should not be treated equally when being scheduled.

We propose a Mission-Aware Task Scheduling scheme, called MATS, which uses different revenue models to quantify each task’s revenue with respect to end users’ satisfaction and resource allocation. When network resources are scarce and not able to complete all of the tasks in time, MATS provides a complete, automated, and global objective to maximize the throughput, user’s satisfaction, and network resource utilization. This project will research the following three tasks:

(1) Development of Proper Revenue Models. Revenue model is the foundation of mission-aware scheduling scheme, since scheduling decisions are made using the concept of revenue. In order to quantify users’ experience, we will develop revenue models using traces from production data centers as well as application logs.
(2) Revenue-based Real-time Scheduling in Data Centers. This problem is NP-hard and we plan to solve it step-by-step. First, a centrally controlled scheduling scheme at the bit-level
will be attempted, followed by centrally controlled scheduling at the packet-level. Finally, scheduling at the packet-level in a distributed manner will be explored.
(3) Performance Evaluation. To evaluate the performance of MATS, we will implement it on a testbed and compare it to existing schemes using a variety of performance metrics.