Skilled in the Art of Being Idle: Reducing Energy Waste in Networked Systems

Summary

This paper explores methods to reduce power usage in idle systems in the most effective way possible.  Some of the problems for systems to enter low-power sleep modes  when idle is that doing so causes the system to lose network connectivity and prevents it from running tasks scheduled to run when the system is in low utilization.

After collecting and analyzing data from 250 enterprise users, they were able to find out the value of proxying, which applications would benefit from proxying and how comprehensive it would need to be in order to be effective.  

Their studies showed that machines are idle about 50% of the time, and this results in 60 TWh/year of wasted electricity, which amounts to 6 billion dollars.  When comparing idle time traffic between home and office environments, they found that average number of packets at home is 1 and at the office is 3.  Based on these they found that any solutions that woke the host up to handle network traffic wouldn't be efficient.

Based on the above data, they propose a proxy architecture that supports different idle-time behaviors.  When designing the proxy, they classified the traffic along 2 dimensions, the first being the need to proxy the traffic with the following options:

• don't wake protocols: generate sustained and periodic traffic and should be dealt with without waking the host
• don't ignore protocols: require attention to make sure correct order of operations
• policy-dependent traffic: how it's handled depends on tradeoff between sophistication of functionality and energy savings.

and the second being the complexity required to proxy the traffic with the following options:

• ignorable (drop): can be safely ignored
• handled via mechanical response: easy to construct required response
• require specialized processing: usually requires state maintenance

The authors designed the following 4 proxies:

• proxy_1: ignores all traffic listed as ignorable & wakes machine to handle all other incoming traffic
• proxy_2: ignores all traffic listed as ignorable, issues responses for protocol traffic to be handled with mechanical responses & wakes the machine for all other incoming traffic
• proxy_3: (1) and (2) as in proxy_2, but wakes up all traffic belong to any of telnet, ssh, VNC, SMB, and NetBIOS & drops any other incoming traffic.
• proxy_4: identical to proxy_3 but also wake up for regular network backups, anti-virus software updates, FTP traffic and Intel specific updates.

They find that proxy_1 is inadequate for home and office and that proxy_3 achieves the most sleep in all scenarios.

Criticism & Questions

I think this was a very interesting paper.  The methodology seemed sound and their evaluation  thorough, especially their classification of idle-time traffic.  I will be curious to see how their work goes forward and if it ever gets implemented in real-world machines.