Paper 1: Mark Stemm and Randy H. Katz

Measuring and Reducing Energy Consumption of Network Interfaces in Hand-Held devices

IEICE (Institute of Electronics, Information and Communication Engineers) Transactions on Communications, special Issue on Mobile Computing

Presentation slides

Motivation

• Next generation of hand-held devices will require seamless network connectivity.
• They have to meet stringent power requirements.
• The NI (Network Interface) consumes significant power in hand-held devices & portable computers.

Goals

• Measure NI power usage
• They measured power usage of 2 PDAs & 4 NIs
• NI power usage was larger than PDA power usage
• This means that NI power management is required

Definition: Idle state - the device is powered on but not doing anything.

Definition: Sleeping state - most circuits in the device are turned off. It can detect only simple events like an interrupt. It takes a long time to come back up from this state.

Measurement conclusions

• Receive state power consumptions is almost same as in Idle state (for both WaveLAN & Metricomm).
• Send state power consumption is almost same as in Receive state for WaveLAN. But for Metricomm, its considerably higher.
• Hence, protocols should minimize packets sent from the mobile device. (Most applications have this characteristic anyway).

Transport Layer Simulation

• Compared performance (using ns simulation) of 4 protocols.
  • TCP reno (ACK for every packet)
  • TCP reno with delayed ACKs (ACK for every other packet)
  • Reliable UDP with fixed window size (ACK per window of packets)
  • Reliable UDP with unlimited window size (ACK per transfer)
• Assuming that the mobile device is the receiving end, they measured:
  • number of ACKs sent
  • number of data packets sent
  • amount of time taken
• Components of energy consumption:
  • SendRecv - this is the energy spent while sending/receiving a packet
  • Idle - this is the energy that must be spent while in the idle state waiting for packets
• Measurement conclusions:
  • Protocols with fewer ACKs are better (UDP with unlimited window size
  • But other protocols also perform well enough if only SendRecv energy is considered
  • TCP interacts poorly with a wireless network in the presence of losses (mistakes losses for congestion). This results in frequent unnecessary back-off and hence a longer transmission time. Thus the ``Idle'' component of energy consumption shoots up.

Application Level Optimizations

Email

• They propose an e-mail reading program that wakes up periodically & checks for e-mail by contacting a mail-host. The period of wait is called the attention span.
• To evaluate the effectiveness of such a strategy, they collected traces of mail arrival times & sizes of mails at CS Dept, U.C.Berkeley. The trace was fed into a simulation.
• Studied the trade-off between energy consumption & mail staleness with varying attention span.
• A 2 minutes staleness can give 20% energy savings. There is a lot to be gained by tolerating a little staleness.

Web Browsing Application

• Collected trace data as above - measuring times & sizes of web transfers.
• Proposed a scheme in which the device alternates between think time & sleep time. After completion of a transfer, the device is kept on for an interval called attention span. Note: This attention span has an intuitive meaning that is opposite to the attention span in the e-mail application discussed previously.
• To evaluate the trade-off between power-savings & user-perceived latency in the scheme, they carried out simulations by varying the attention span - using the trace data collected.
• Result: Even in metricomm devices (where wakeup time is quite high), this strategy can result in 30% energy savings with almost no user visible latency.