Motivation
|
The primary motivation for this work is the poor utilization of
resources in data centers. Consolidation is often used to avert
this problem, but ad-hoc mechanisms produce more problems.
Ironically, the consolidated servers still see poor utilization
but fail to meet service levels when bursty load comes.
The goal of this project is to explore the use of dynamic, control-based
allocation of resources (CPU cycles, memory, network bandwidth, storage
bandwidth, etc.) to:
- enable a resource sharing system to achieve application-level QoS
goals, and
- minimize the resource requirements for whole application stacks.
in a virtualized infrastructure.
|
|
Approach
Our approach is to specify the desired behavior of an application-stack in
terms of a set of metrics, monitor these metrics, and control the
configuration of the system so that the desired behavior is achieved. We
explictly monitor behaviour indicators from multiple locations in an
application stack including applications software, systems software layers,
and hardware stack. These indicators are then used as inputs to a control
element, which makes decisions for the settings of actuators -- parameters
that change behavior of a particular hardware or software layer.
Architecture
In our current system, the controller consists of two sub-controllers called
AppController and NodeController. Every control interval,
- App controllers determine the resources to meet goals
- The requests for each VM are sent to node controllers
- Node controllers finally allocates resources
An example set of applications and how the controller interaction happens is
below.
People
- Pradeep Padala
- Kang G. Shin
- Xiaoyun Zhu (HP Labs)
- Mustafa Uysal (HP Labs)
- Zhikui Wang (HP Labs)
- Sharad Singhal (HP Labs)
- Arif Merchant (HP Labs)
Publications
Arif Merchant, Mustafa Uysal, Pradeep Padala, Xiaoyun Zhu, Sharad
Singhal and Kang G. Shin. Maestro: Quality-of-Service in Large Disk Arrays. In the
proceedings of the International Conference on Autonomic Computing (ICAC '11),
Jun 2011. [PDF]. Best paper award!
Pradeep Padala, Mustafa Uysal, Arif Merchant, Xiaoyun Zhu, Sharad Singhal and
Kang G. Shin. Performance Differentiation for Multi-port Arrays:
A Control-Theoretic Approach. In the proceedings of the fourth
International Workshop on Feedback Control Implementation and Design in
Computing Systems and Networks (FEBID'09), Apr 2009. [PDF]
Pradeep Padala, Xiaoyun Zhu, Mustafa Uysal et al. Automatec Control of
Multiple Resources. In the
proceedings of the EuroSys 2009 [PDF] [PPT]
Jin Heo, Xiaoyun Zhu, Pradeep Padala, Zhikui Wang.
Memory Overbooking and Dynamic Control of Xen Virtual Machines
in Consolidated Environments.
IFIP/IEEE Symposium on Integrated Management (IM'09) mini-conference,
Jun 2009. [PDF].
Xiaoyun Zhu, Mustafa Uysal, Zhikui Wang, Pradeep Padala,
Sharad Singhal, Arif Merchant, Kang Shin.
What Does Control Theory Bring to Systems Research?
To appear in ACM SIGOPS Operating Systems Review, Jan 2009.
[PDF].
Xue Liu, Xiaoyun Zhu, Padala Pradeep, Zhikui Wang, Sharad Singhal.
Optimal Multivariate Control for Differentiated Services on a Shared Hosting
Platform. In the proceedings of the 46th IEEE
Conference on Decision and Control (CDC'07), Dec
2007. [PDF]
Pradeep Padala, Xiaoyun Zhu, Mustafa Uysal et al. Adaptive Control of
Virtualized Resources in Utility Computing Environments. In the
proceedings of the EuroSys
2007 [PDF]
Zhikui Wang, Xiaoyun Zhu, Pradeep Padala and Sharad Singhal. Capacity and Performance Overhead in Dynamic Resource Allocation to
Virtual Containers. In the proceedings of the IFIP/IEEE Symposium on
Integrated Management
(IM'2007),
May 2007. [PDF]