TL;DR
Enterprises want to save energy by letting idle desktops sleep, but IT departments resist solutions requiring special hardware or complex management. GreenUp offers a software-only, decentralized approach where any awake machine can act as a "sleep proxy" for sleeping machines on the same subnet, ensuring availability without dedicated servers.
What Problem Does This Paper Address?
Large enterprises face a significant energy drain from idle desktop computers. While putting these machines to sleep can save millions, IT departments are hesitant to adopt existing solutions. These solutions often require specialized hardware, disruptive virtualization, or dedicated proxy servers for each subnet, all of which are costly and difficult to manage. The core challenge is to provide seamless remote access and availability to sleeping machines without introducing significant IT overhead or single points of failure.
Methods
GreenUp employs a novel, decentralized approach to manage sleeping machines:
- Decentralized Management: Instead of a dedicated server, any awake machine on the subnet can act as a "sleep proxy" for sleeping machines. This distributes the load and eliminates single points of failure.
- Distributed Probing: To ensure no sleeping machine goes unmanaged, awake proxies periodically probe a random subset of machines. This is done efficiently to avoid overwhelming the network or individual machines.
- Subnet State Coordination: To maintain a consistent view of machine states across the subnet, GreenUp uses a broadcast-based mechanism. This allows machines to share their status and for managers to track their "managees."
- Guardians: To prevent an "apocalypse" (where all proxies fall asleep), a small, dynamically chosen set of machines are designated as "guardians." These machines are prevented from sleeping automatically, ensuring at least a minimum level of proxy availability.
- Wake-on-LAN (WoL) Integration: When a sleeping machine is accessed, its proxy uses WoL magic packets to wake it up. The system is designed to handle potential WoL failures and delays.
Datasets
The paper uses data from a seven-month live deployment of GreenUp on over 100 machines within Microsoft Research. They also reference data from a previous study by Reich et al. [23] for analyzing sleep behavior.
Key Insights and Ideas
- The "Sleep Proxy" is Key, but Needs Decentralization: The core idea of a proxy handling traffic for sleeping machines is sound, but the enterprise IT barrier is the need for dedicated, managed servers. GreenUp's innovation is making any machine a proxy.
- Availability Trumps Consistency: In this context, it's far more critical that a machine wakes up when needed than for all proxies to have a perfectly synchronized, real-time view of every machine's state. This allows for simpler, more robust designs.
- Correlated Sleep is a Real Problem: Machines don't always sleep independently. Power outages or scheduled maintenance can cause many machines to sleep simultaneously, necessitating a mechanism like "guardians" to prevent total system failure.
- WoL is Mostly Reliable, but Has Quirks: While Wake-on-LAN is the primary wakeup mechanism, the paper identifies rare but critical failure modes, particularly when machines are in the process of sleeping. This led to adjustments in probing timeouts.
- Energy Savings are Significant, but User Behavior is Key: GreenUp enables machines to sleep for about 31% of the time, leading to substantial energy savings. However, the actual savings depend on whether users allow their machines to sleep in the first place.
Key Results
- GreenUp successfully provides high availability for sleeping machines, with a wake-up success rate of over 99%.
- The system is highly efficient, with CPU utilization on a manager machine increasing by less than 1% when managing up to 100 machines.
- The guardian mechanism effectively ensures at least three proxies are awake on a subnet, preventing apocalypse.
- The system is designed to be easily deployable, requiring no special hardware.
- In their deployment, participants spent approximately 31% of their time asleep, translating to significant potential energy savings.
- GreenUp wakes machines quickly enough to satisfy user patience in about 85% of access attempts.
Limitations and Future Work
- Encrypted Traffic: GreenUp currently struggles with fully encrypted traffic (e.g., IPSec). Future work involves developing techniques to handle this, potentially by identifying specific traffic patterns or requiring client-side modifications.
- WoL Reliability: While generally reliable, rare WoL failures still occur. The team plans to build a mechanism to automatically test participants for persistent WoL issues.
- Guardian Selection: Currently, guardians are chosen randomly. Future work aims to predict which machines are likely to be awake anyway and select them as guardians to further reduce the energy cost.
- User Behavior Influence: The paper notes that IT mandates often dictate sleep policies. Future work could explore how GreenUp might encourage users to adopt more aggressive sleep policies in environments without such mandates.
- Server Cluster Applications: The techniques developed for GreenUp are being explored for adapting server cluster sizes dynamically to load, aiming for energy efficiency in server environments.
Summary updated