3 Fred Douglis and Brian Marsh Low power disk management for mobile computers Technical Report MITL TR 53 93 Matsushita Information Technology Laboratory April 1993 responsible for developing the concept of Redundant Arrays of Inexpensive Disks RAID now a 3 billion industry segment Katz s recent research has focused on wireless communications and mobile computing applications 4 D P Hembold D E Long and B Sherrod A Dynamic Disk Spin down Technique for Mobile Computing In Proc MOBICOM 96 November 1996 5 T Imielinski S Viswanathan and B R Badrinath Energy efficient indexing on air In Proceedings ACM SIGMOD 1994 From January 1993 through December 1994 Katz was a program manager and deputy director of the Computing Systems Technology Office of ARPA He was responsible for wiring the White House to the Internet and also assisted the Clinton Administration in formulating policies related to the National Information Infrastructure and wireless technologies 6 Ravi Jain and John Werth Airdisks and airraid modeling and scheduling periodic wireless data broadcast Computer Architecture News 23 4 23 28 September 1995 7 Kester Li Roger Kumpf Paul Horton and Tom Anderson A quantative analysis of disk drive power management in portable computers In Proceedings 1994 Winter USENIX pages 279 291 San Franscisco California Winter 1994 8 Brian Marsh Systems issues for mobile computing Technical Report MITL TR 50 93 Matsushita Information Technology Laboratory Feburary 1993 9 ns LBNL Network Simulator http wwwnrg ee lbl gov ns 1996 10 W M Smith and P S Ghang A Low Power Median Access Control Protocol for Portable Multi Media Systems In Proc Third Workshop on Mobile Multimedia Communications MoMuC 3 Sept 1996 11 Stanley Zdonik Michael Franklin Rafael Alonso and Swarup Acharya Are Disks in the Air just pie in the sky In IEEE Workshop on Mobile Computing Systems and Applications pages 12 19 Santa Cruz California December 1994 He is a member of the ACM and a senior member of the IEEE Mark Stemm ACM S 1995 IEEE S 95 is a Ph D student in Computer Science at the University of California at Berkeley His research interests are in the areas of wireless and wide area computer networks mobile computing and operating systems He received his B S degree in Computer Science with University and Departmental honors from Carnegie Mellon University in 1994 and his M S degree in Computer Science from the University of California at Berkeley in 1996 On the WWW his URL is http www cs berkeley edu stemm His e mail address is stemm cs berkeley edu Professor Randy H Katz is a leading researcher in computer system design and implementation His research experience has spanned numerous disciplines He has written over 120 technical publications on CAD database management multiprocessor architectures high performance storage systems and video server architectures He was 7 Page Response Time sec impact on the number of packets sent and received by the mobile device the actual power difference is minimal This is because the energy consumed simply by keeping the network interface on during the transfer contributes the most to the final energy cost In the presence of a high packet error rate however current TCP sender implementations overreact to packet losses mistaking them for congestion This slows down the transfer rate which increases the amount of time that the transfer takes and the amount of energy consumption by the network interface 3 2 915MHz WaveLAN 2 4GHz WaveLAN 1 Simulations show that for email our optimizations can reduce the energy consumption to the minimum possible the energy required to receive messages For web browsing fast sleep idle transitions allow significant power savings with no impact on user visible latency Even for interfaces with longer sleep idle transitions however significant power savings can be achieved with less aggressive management of the network interface 0 100 200 300 400 500 Attention Span sec Figure 9 Response time as a function of attention span for 915 Mhz and 2 4 Ghz Wavelan Page Response Time sec 8 6 1 Recommendations for Future Networks Interfaces and Protocols 6 Current generation transport and link level protocols may need some tuning to minimize the power cost of network interfaces Any protocol that leaves a mobile receiver idle unnecessarily such as TCP s backoff in the presence of wireless losses wastes power Even when the protocol is performing correctly inefficient link layer scheduling may be the problem a link layer that allocates 2 Mb on a contention basis for 10 mobiles causes each of them to consume 10 times as much power 100 times as much power total as a base station that uses a TDMA scheme to coordinate delivery of data to receivers Recent work has proposed more intelligent link layer schemes to handle this problem 10 The valuable lesson is that network interfaces can consume a significant fraction of the power budget of PDAs and this requires smart software and applications to make sure that battery lifetime is not needlessly shortened Metricom 4 2 0 100 200 300 400 500 Attention Span sec Figure 10 Response time as a function of attention span for Metricom that increasing attention spans lead to significant reductions in energy costs with no user visible increase in latency Figure 10 shows the response time as a function of attention span for the Metricom NI This illustrates the effect of a large sleep wakeup transition For shorter attention spans the 5 second delay as the interface is powered on has a uservisible latency For larger attention spans however the latency to retrieve the web page dominates 7 Acknowledgments Thanks to Paul Gauthier and Daishi Harada who worked on earlier versions of this work Thanks also go to Bruce Mah from UC Berkeley for providing us with the WWW traces This work is supported by DARPA contract DAAB07 95 CD154 and grants from the California MICRO Program Hughes Aircraft Corporation Metricom and AT T 6 Conclusions Recommendations Our measurements of PDA and Network Interface power and energy consumptions show that Network Interfaces consume a significant fraction of the total power on a PDA Additional measurements for sending and receiving packets of various sizes indicate that the power consumed when the interface is on and idle is virtually identical to the cost of receiving packets For some interfaces the cost of sending packets can be significant when compared to the cost of being idle but application and transport level considerations