Energieeffizienz in verteilten Systemen:
Modellierung und Simulation
Helmut Hlavacs
University of ViennaDepartment of Distributed and Multimedia Systems
FTW Kolloquium 18.5.2010
Energy Efficient ICT
• COST Action IC804 Energy efficiency in large scale distributed systems
• Supported by the European Commission• http://www.cost804.org/• Member States: 17 (+3 pending)• Member Insitutions: ~40
• Chair: Jean-Marc Pierson, IRIT, Toulouse• Vice-Chair, Grant Holder: Helmut Hlavacs, Univ. of Vienna
FTW Kolloquium 18.5.2010
Themen
WG1: Ongoing evaluation of components
WG2: Modeling energy efficiency
WG3: Adaptive actions
WG4: Characterization of performance-energy saving trade-off
WG5: Dissemination
FTW Kolloquium 18.5.2010
Distributed Systems
• Networked computing entities• At network edges• Interact with each other• Heterogeneous• Small or large scale• Communication via NW protocols or middleware
abstraction• Distributed algorithms (-> software!)
FTW Kolloquium 18.5.2010
Optimizing Distributed Systems and Algorithms
• Behavior driven by interaction between nodes and communication systems
• Complex, many parameters• Emergent behavior through local information
• Performance evaluation and optimization– Implement and run (e.g., PlanetLab)– Simulation– Need math. models to understand the performance
FTW Kolloquium 18.5.2010
Saving Power in Distributed Systems
• Optimize single nodes– Advanced Configuration and Power Interface (ACPI)
• C (idle): suspend to RAM/disk, WakeOnLAN• P (operational: frequency, voltage pairs) states
– Multicores: dectivate single cores– Specialize nodes (e.g. nettops vs. GPU)
• In distributed systems– Optimize parameters– Hardware consolidation
FTW Kolloquium 18.5.2010
Idle Consumption
Energy Star
FTW Kolloquium 18.5.2010
Consumption depending on CPU Load
Google 2007
FTW Kolloquium 18.5.2010
Hardware Consolidation
Requires a model of workload, energy consumption and efficiency, network bandwidth, system performance, QoS, …
FTW Kolloquium 18.5.2010
Residential ICT
• World wide (2009): over a Billion PCs• EU-25 (2007)
– 2005: ~105 Mio desktop, 24 Mio laptops and 104 Mio monitors (47 Mio flat panel) installed in households
– 2006: broadband 60 Mio subscriber lines in the EU-25,
– End devices in homes contribute a large share of electricity consumption growth in the EU
• UK (2006): residential office equipment ~7 TWh (or 6% of total residential consumption).
• UK (2007): 21% of work PCs never switched off (1.5 TWh)• USA (2007): 16 TWh by office/home PCs• USA (2008): 74 TWh consumed by Internet equipment
FTW Kolloquium 18.5.2010
Example: File Sharing
• Millions of PCs in households world wide• Long running• Consume large quantities of energy
• Can we make file sharing energy efficient?
FTW Kolloquium 18.5.2010
File Popularity vs. Rank
• Zipf‘s Law but with exponential tail
Dan, Carlsson,
2010
FTW Kolloquium 18.5.2010
Energy Efficient File Downloading
• BitTorrent fluid model (Qiu, Srikant 2004)
• Distributed proxies (Hlavacs et al. 2008)
• BitTorrent with proxy (Anastasi et al. 2010)
• Green BitTorrent (Blackburn, Christensen 2009)
FTW Kolloquium 18.5.2010
BitTorrent
• Most prominent P2P file sharing protocol• Good for popular files• Clients download pieces from a complete source• Start sending missing pieces to each other• Policy agains free riders (choking)• The available bandwidth can be saturated
FTW Kolloquium 18.5.2010
BitTorrent
Seeder: a peer that has the whole file
Leecher: a peer that has only part of the file
FTW Kolloquium 18.5.2010
BitTorrent Fluid Model
• Qiu, Srikant, Modeling and Performance Analysis of BitTorrent-Like Peer-to-Peer Networks, SigComm 2004
• Peers are like buckets where data flows into• Data flows with max up/downlink bandwidth• Good realistic model
• Can we use it to investigate energy efficiency?
FTW Kolloquium 18.5.2010
Model Parameters
• Parameters• x(t)…number of leechers• y(t)…number of seeders• …arrival rate of new leechers• …uploading bandwith of a peer• c…downloading bandwidth of a peer• …abort rate of peers (set to zero in our case)• …rate at which seeders leave the system (=1/)• …effectiveness
FTW Kolloquium 18.5.2010
FTW Kolloquium 18.5.2010
Fluid Model
• Bandwidth that is downloaded into peers
• Bandwidth that is uploaded from peers and seeders
• Rate of leechers turning into seeders
FTW Kolloquium 18.5.2010
Solution of the Differential Equations
Mean download time
The only parameter we can influence in the distributed algorithm is the nice time =1/
System power consumption
FTW Kolloquium 18.5.2010
Which Nice Time is Optimal?
FTW Kolloquium 18.5.2010
The Optimal Nice Time
FTW Kolloquium 18.5.2010
Distributed Hardware Consolidation
• H. Hlavacs, R. Weidlich, K.A. Hummel, A. Houyou, A. Berl, H. de Meer, Distributed Energy Efficiency in Future Home Environments, Annals of Telecommunications 63:7-8, Sept.-Oct. 2008.
• Covers the case for unpopular files• No sharing possible if files are hosted by only one seeder -> use
consolidation• Concentrate parallel downloads on distributed proxies, then move
files to the owners• Proxies are chosen on the fly
– Once a peer wants to download something but does not find a proxy itself, it becomes a proxy
FTW Kolloquium 18.5.2010
Parallel Downloads• Concentrate downloads on a small number of nodes (e.g. might
have larger up/downlink bandwidth)• Can work only if the downstream goodput is less than the available
upstream bandwidth
FTW Kolloquium 18.5.2010
Number of Running PCs
No consolidation
With consolidation
FTW Kolloquium 18.5.2010
Experimental Results
FTW Kolloquium 18.5.2010
Local Proxy• G. Anastasi, I. Giannetti, A. Passarella, A BitTorrent proxy for Green Internet
file sharing: Design and experimental evaluation, Computer Communications 33 (2010) 794–802
• Concentrate downloads on dedicated local proxy
• Critique
• Requires manual management and maintenance
• Bad scaling of local growth (bandwidth)
FTW Kolloquium 18.5.2010
Green BitTorrent
• J. Blackburn, K. Christensen, A simulation study of a new Green BitTorrent, Proceedings First International Workshop on Green Communications (GreenComm 2009), Dresden, Germany, June 2009
• Hibernate seeders that currently do not have any uploads• If number of seeders drops below a limit• -> Wake up sleeping seeders per WakeOnLAN
FTW Kolloquium 18.5.2010
Conclusion
• In large scale distributed systems energy can be saved by
– Local techniques (hardware, OS optimization)– Optimizing parameters of distributed algorithms– Hardware consolidation
• BUT: we have to understand why this works
• -> create models that provide insight