Biswajit.HD
Member
- Joined
- 5 Aug 2011
- Messages
- 2,282
- Reaction score
- 281
Data center managers are well versed in distributing power efficiently to physical servers. But the proliferation of virtualization, with multiple virtual machines and applications running on a single piece of hardware, has made this task a lot more complicated.
That's why a Duke University researcher has partnered with Microsoft researchers to design a system that monitors the power needs of individual virtual machines and distribute power based on application priorities.
Currently, IT shops use tools to over-subscribe power distribution, provisioning less power to applications than they could theoretically use, figuring that applications typically won't hit their peak power load.
But while that method works out well on physical servers, it falls short with virtual machines, Harold Lim of Duke said while giving a talk at the annual Usenix technical conference in Portland, Ore.
[MORE FROM USENIX: Want to stop cybercrime? Follow the money]
It's easy to cap the power to a whole server, but hard to cap the power of individual virtual machines and differentiate between applications. What's needed is an application-aware power distribution system that has visibility into the virtual machine layer, he said.
Lim, along with Aman Kansal and Jie Liu of Microsoft Research, designed a virtualized power shifting (VPS) system that budgets power with these considerations in mind.
VPS dynamically shifts power among various distributed components to efficiently utilize the total available power budget, as workloads and power availability vary," they write. "Power is distributed among application components in the correct proportions to achieve the best performance. The system respects application boundaries and differentiates performance based on priorities."
The system, they say, is more granular than existing technologies both in terms of examining individual VMs and providing customization in the method of power distribution. One challenge is making sure that throttling one set of applications doesn't affect another.
"In contrast to existing techniques that use only one power control knob, typically frequency scaling, VPS uses multiple power control knobs and selects the optimal combinations of power settings to optimize performance within the available power budget," the researchers write.
VPS shifts power dynamically as workloads change, and as power availability changes. The system should be able to handle a sudden drop in the power budget, Lim said.
But it's still a work in progress. For example, the question of how to automatically shut down servers and move virtual machines from one box to another is unresolved. But given that live migration is already offered with hypervisors by the likes of VMware and Microsoft, it seems likely the capability could be integrated into future power management systems.
That's why a Duke University researcher has partnered with Microsoft researchers to design a system that monitors the power needs of individual virtual machines and distribute power based on application priorities.
Currently, IT shops use tools to over-subscribe power distribution, provisioning less power to applications than they could theoretically use, figuring that applications typically won't hit their peak power load.
But while that method works out well on physical servers, it falls short with virtual machines, Harold Lim of Duke said while giving a talk at the annual Usenix technical conference in Portland, Ore.
[MORE FROM USENIX: Want to stop cybercrime? Follow the money]
It's easy to cap the power to a whole server, but hard to cap the power of individual virtual machines and differentiate between applications. What's needed is an application-aware power distribution system that has visibility into the virtual machine layer, he said.
Lim, along with Aman Kansal and Jie Liu of Microsoft Research, designed a virtualized power shifting (VPS) system that budgets power with these considerations in mind.
VPS dynamically shifts power among various distributed components to efficiently utilize the total available power budget, as workloads and power availability vary," they write. "Power is distributed among application components in the correct proportions to achieve the best performance. The system respects application boundaries and differentiates performance based on priorities."
The system, they say, is more granular than existing technologies both in terms of examining individual VMs and providing customization in the method of power distribution. One challenge is making sure that throttling one set of applications doesn't affect another.
"In contrast to existing techniques that use only one power control knob, typically frequency scaling, VPS uses multiple power control knobs and selects the optimal combinations of power settings to optimize performance within the available power budget," the researchers write.
VPS shifts power dynamically as workloads change, and as power availability changes. The system should be able to handle a sudden drop in the power budget, Lim said.
But it's still a work in progress. For example, the question of how to automatically shut down servers and move virtual machines from one box to another is unresolved. But given that live migration is already offered with hypervisors by the likes of VMware and Microsoft, it seems likely the capability could be integrated into future power management systems.