|M.Sc Student||Nave Eyal-Itzhak|
|Subject||TCP Window Based Dynamic Voltage and Frequency Scaling|
(DVFS) for Low Power Communication Network
Controller System on Chip (SoC)
|Department||Department of Electrical and Computer Engineering||Supervisor||PROFESSOR EMERITUS Ran Ginosar|
|Full Thesis text|
The U.S. annual electricity consumption of network equipment is estimated to be 46.3 TWh in 2012 (with ~15% annual growth rate). The U.S. annual cost of this network electricity consumption reaches billions of dollars (1 TWh costs ~$100 million). DVFS is a well-known energy reduction technique balancing performance and power consumption, already implemented in microprocessors. In network controllers, the decision whether to enable high performance at the cost of power should not rely solely on the workload, i.e. the amount of data needed to be transmitted, but also on the ability of the network to deliver it. Flooding data to a congested network is as efficient as high (speed) gear and fully pressing the gas pedal for a car stuck in a traffic jam.
This work presents an approach for power reduction in network controller SOC transmitting data using the TCP protocol. This approach uses the TCP window management capability to sense congested networks for deciding which DVFS power mode to use, avoiding inefficient usage of high power. This research compares the energy saving of TCP Window based DVFS (TWD) and Packet Buffer based DVFS (PBD) in order to check which policy saves more energy from existing network controllers which do not use DVFS at all. In addition, the effect of various traffic load levels and network congestion levels on the energy consumption of the two DVFS policies is examined as well.
A dedicated simulation environment was developed to test this approach and compare the energy consumptions. The simulation environment is highly configurable, thereby laying the infrastructure for future research experiments of energy consumption of DVFS mechanisms for network systems. In this research, the developed simulation environment is used to compare the energy consumption of TWD and PBD in various traffic load levels and various network congestion levels. Simulation results show that across various traffic load levels, TWD achieves an average of 10% more energy saving than PBD. In addition, across various network congestion levels, TWD nearly doubles the energy saving achieved by PBD, saving up to 52% of the energy in the highest congestion level, compared to only 26% energy saving achieved using PBD. If TWD would have been fully adopted in 2012, it would have saved $145 million annually in the U.S. when compared to using PBD, and $1.476 billion annually when compared to not using DVFS at all (~32% reduction). World global savings are 3-5 times larger.