Research Topics
[BA] Experimental evaluation of the trade-off between energy-efficiency and performance at the example of video transcoding
The aim of energy efficient computing is to significantly reduce the energy consumption of servers without significantly affecting the their performance. In a cluster or cloud environ- ment, this can be achieved by avoiding conditions in which servers are either overutilised or underutilised. In overutilised servers, due to the existence of bottlenecks in resource utilisa- tion, performance is reduced even through servers are consuming a large amount of energy. Likewise, in underutilised servers, resources are not utilised efficiently even though they are ready for work, consuming power. Ideally, application-level performances should be maintained while platform-level energy consumption is monitored and managed. At the application level, application components can be optimally configured to satisfy user requirements (which are expressed by quantifiable service level agreements) and at platform-level, virtual machines are live migrated to consolidate servers and turn down underutilised servers. The aim of this thesis is to quantify the trade-off between energy-efficiency and perfor- mance. As a use case, the thesis considers the transcoding of videos. Users of transcoders specify transcoding quality, input and output video formats, and transcoding time. Based on these requirements, the application identifies the best transcoder and determines the amount of resources the transcoder requires. With the optimal configuration, the application should be given a transcoding task and the resource and energy consumption of the transcoder should be measured (observed), first in isolation, then while the application executes together with other applications (benchmarks). For the latter case, after a transcoder starts execution of a batch job (for example, transcoding 2000 videos), virtual machines encapsulating different benchmarks will be gradually migrated into the server and observation will be made to quantify:
- The relationship between the transcoder’s energy consumption and its performance (for example, how much energy is required to transcode a Kbit of video)?
- How much overall work has the server accomplished for each Joule of energy (or Watt of power) it consumes? There must be a mechanism to quantify this relationship.
- The stability of the transcoder’s performance as the load of the server gradually increases (it is recommended to develop a graph that shows QoS vs. Energy-efficiency).
Betreuer: René Schöne