Packet Coalescing and Server Substitution for Energy-Proportional Operation of Network Links and Data Servers.

摘要

Electricity generation for Information and Communications Technology (ICT) contributes over 2% of the human-generated CO2 to the atmosphere. Energy costs are rapidly becoming the major operational expense for ICT and may soon dwarf capital expenses as software and hardware continue to drop in price. In this dissertation, three new approaches to achieving energy-proportional operation of network links and data servers are explored.;Ethernet is the dominant wireline communications technology for Internet connectivity. IEEE 802.3az Energy Efficient Ethernet (EEE) describes a Low Power Idle (LPI) mechanism for allowing Ethernet links to sleep. A method of coalescing packets to consolidate link idle periods is investigated. It is shown that packet coalescing can result in almost fully energy-proportional behavior of an Ethernet link. Simulation is done at both the queuing and protocol levels for a range of traffic models and system configurations. Analytical modeling is used to gain a deeper general insight into packet coalescing.;The architecture of a hybrid web server based on two platforms – a low-power (ARM based) and a high-power (Pentium based) – can be used to achieve step-wise energy-proportional operation and maintain headroom for peak loads. A new method based on Gratuitous ARP for switching between two mirrored platforms is developed, prototyped, and evaluated. Experimental results show that for up to 50 requests per minute, a hybrid server where the Master platform is a 2012 server-grade desktop PC can sleep for 50% of time with no increase in response time.;HTTP can be used for redirection in space – a new method for precise redirection in time is proposed and used to schedule requests to a high-power server in a hybrid server. The scheduling method is modeled as a single server queue with vacations where the vacation duration is fixed and the service distribution is directly a function of the request load. This approach is well suited for delay tolerant applications such as application updates and file back-up. Energy-proportional operation is shown to be achievable in a prototype system.;A first-order estimation with conservative assumptions on the adoption rate of the methods proposed and studied here shows that these methods can collectively enable energy savings in the order of hundreds of million dollars in the US annually.