-1

What does the operating system do in ordder to manage power in current HPC clusters ? What are the functionalities embedded in current HPC clusters in order to save power

kashyapa
  • 337
  • 4
  • 17
  • 1
    I smell homework, which is perfectly fine - just as long as you're open about the fact that it's homework and you give us some background on your understanding of the problem. – Mark Henderson Oct 28 '10 at 19:35
  • THis is not homework. I'm trying to find a research problem in the area of HPC power management. So i want to know the current technologies in this field – kashyapa Oct 29 '10 at 19:25
  • Is there anything I can do to increase it ? – kashyapa Oct 29 '10 at 20:48
  • review your previous question you've asked, if they have good answers please mark them as the answer (checkmark next to the question). – Chris S Nov 02 '10 at 12:25

2 Answers2

1

Not every cluster is the same, and the power management varies as widely as any other feature. Some use very rudimentary techniques to save a little power, others shutdown nodes that aren't used, or other actions that save more significant amounts of power.

If you have questions about a specific model of cluster we may be able to help more, but open ended questions like this will not have a good answer.

Chris S
  • 77,945
  • 11
  • 124
  • 216
  • Do current HPC clusters running Linux Kernel on x86 Platform have Ondemand GOvernors running in the kernel , which can change the P-states based on the CPU utilization ? – kashyapa Oct 29 '10 at 19:27
  • I'm not sure what a "P-states" is, but if you're referring to power management, then yes. Linux and almost all modern OSes have had power management for quite a few years. – Chris S Oct 29 '10 at 20:44
1

Well, if you look at a typical HPC cluster, the compute nodes will be standard x86 servers running Linux. So you might ask, what can Linux do to save power on x86 hardware? Well, there's all these ACPI C-states and whatnot that can be used to put processors into some kind of sleep mode based on inactivity, or clock down the processors and so forth.

Then, some batch schedulers have features allowing idle nodes to be shut down, and then when there is work to do, the scheduler can boot these nodes (via IPMI, ILO, or some other remote management system).

In practice, we haven't really looked into any of that. Why? Well, our clusters run more or less constantly at 90+% load. If there at some point is lots of idle nodes, some clever geezer will see his/her chance and submit a mile long list of jobs to the queue; Problem solved. :)

janneb
  • 3,841
  • 19
  • 22
  • DOes it have Ondemand GOvernors running in the kernel , which can change the P-states based on the CPU utilization ? – kashyapa Oct 29 '10 at 19:26
  • 1
    @kashyapa: Most clusters run just the stock RHEL kernel. If that kernel supports the ACPI power management system of the hardware, sure, the ondemand governor can be used. But why would we? We run at full load 24/7/365, and there is the concern that switching into sleep states will cause jitter which is poison for tightly synchronized parallel applications. – janneb Oct 29 '10 at 21:13