Some Personal Observations from 15 Years of Mobile Computing

1
Some Personal Observationsfrom 15 Years of
Mobile Computing

• M. Satyanarayanan
• School of Computer Science
• Carnegie Mellon University

2
Credits(in order of collaboration)
Also Dan Siewiorek, Peter Steenkiste, David
Garlan, David OHallaron, Jeannette Wing, Mike
Kozuch
3
Systems Context
4
Where is the Beef?
Or Tofu, for vegetarians like me?
5
Early-90s Dream of Mobile Computing
6
Phenomenal Hardware Progress
7
Explosion of Activity

• Proliferation of marketplace for mobile computers
• Increasing demand for wireless communication
• Numerous experimental and commercial systems and
  applications
• Many conferences and workshops
• Lots of smart people investing time and energy
• But ????

8
Some Nagging Doubts

• Are we just riding the technology curve?
• Is there anything deep about mobile computing?
• Are there any real intellectual challenges?
• Do we have any valuable insights?

9
Core Challenges of Mobility

• Less security robustness
• theft, destruction more likely
• greater exposure to subversion
• Multi-modal Interaction
• hands and eyes occupied
• speech/gesture recognition
• augmented reality
• Scarce user attention
• focus of attention elsewhere
• lower human performance
• higher error rate

• Resource poverty
• vs. static elements of same era
• weight, power, size constraints
• Communication uncertainty
• bandwidth / latency variation
• intermittent connectivity
• may cost real money
• Finite energy source
• actions may be slowed or deferred
• communication costs energy

10
How Did We Get Here?(and where are we going?)
11
Evolutionary Path
Smart spaces Invisibility Uneven
conditioning Localized scalability
12
Smart Spaces

• Convergence of IT and building technology
• allows each world to sense and influence the
  other
• smart space influenced by presence of user(s)
• computing space influenced by surroundings
• Individual objects may also be smart

13
Invisibility

• Weisers ideal complete disappearance of
  technology
• Approximation minimal user distraction
• system meets user expectations continuously
• rarely presents surprises
• interactions almost at a subconcious level
• Caveat anticipation may avoid later unpleasant
  surprise
• e.g. pain in a normally-unnoticed body part
• proactivity may sometimes be valuable

14
Masking Uneven Conditioning

• Nonuniform penetration of pervasive computing
  technology
• at least for many decades
• org structure, economics, business models, etc.
• Smart and dumb environments of varying
  degrees will coexist
• large dynamic range can be jarring to a user
• hurts goal of minimizing user distraction
• Can personal computing space compensate?
• disconnected operation is a trivial example of
  this capability
• complete masking impossible
• but reduced variability may be feasible

15
Localized Scalability

• Environments saturated with computing
• intense interactions with users personal
  computing space
• severe bandwidth, energy and distraction
  implications
• further complicated by multi-user environments
• Scalability becomes key problem
• Historically, scalability has ignored physical
  separation
• only logical connectivity matters
• recipe for disaster in pervasive computing!
• More viable attenuate distant or irrelevant
  interactions
• learn from nature inverse square laws!
• contradicts death of distance ethos of Internet

16
Some Deep Conceptual Issues(originally posed at
PODC 1995 invited talk)
17
Effectiveness of Caching

• The traditional metric of cache performance is
  miss ratio
• Reliance on miss ratio assumes
• all misses equally important
• all misses require equal service time
• ordering and spacing of misses irrelevant
• Mobile computing violates these assumptions
• disconnected user impeded by critical file miss
• miss on big files more expensive
• disk spin down for power saving
• Examples of alternative metrics
• references until first cache miss
• references until a critical miss

18

• Some open questions
• What are the right set of metrics?
• Under what circumstances does one use each?
• How does one efficiently monitor these metrics?
• What are the implications for caching algorithms?

19
Multi-Fidelity Computation

• Classical notion of algorithm
• precise output specification
• resources consumed are dependent variables
• time, memory, energy
• Adaptation for mobility suggests a different
  viewpoint
• Do the best you can using no more than X units
  of resource
• output correctness criteria no longer fixed
• multiple notions of correct each is a level of
  fidelity
• reversal of roles of output spec and resource
  consumption
• Compelling approach where human interaction
  involved
• especially augmented reality applications
• quick dirty answer now better than perfect
  answer later
• deferred work often goes away!
• humans are noisy sources of input, tolerant sinks
  of output

20

• Allow system to find sweet spots
• Give the best result you can cheaply
• knee in fidelity-resource usage curve
• especially valuable with steep performance cliffs
• PhD work of Dushyanth Narayanan (2002)
• Some open questions
• What is a good fidelity metric?
• How do we compare multi-fidelity algorithms?
• Can multi-fidelity computations be composed?
• Given computation P, how do we find close
  computation Q?

21
Consistency at High Latency

• Weak connectivity makes cache coherence
  expensive
• communication latency
• state restoration after intermittent failures
• Coda solution
• use callbacks instead of demand validations
• raise granularity of cache coherence
• trade off precision of invalidation for speed of
  validation
• The notion of callback can be generalized
• client caches data satisfying predicate P
• server remembers cheap predicate Q
• Q ? P, but reverse need not be true
• Q re-evaluated by server on each update
• if Q false, notify client that P might be false
• Q is a semantic callback for P

22

• Some open questions
• When are semantic callbacks most useful?
• What forms can P and Q take? Code?
• How does one derive Q from P quickly?
• Can cost-benefit tradeoff be made adaptive?

23

• Value in speed-of-light limited systems?
• mobile networks foreshadow problems of future
  distributed systems
• (e.g. current interest in Delay Tolerant
  Networking)
• communication latency ?? computational speed
• coast-to-coast RPC gt 30 msec at speed of light
• 30 msec is over 3 million instructions on a
  (very slow!) 100MIP processor!
• Example tele-operated Mars rover (10 minutes one
  way at c!)
• use Q as semantic validator conditional
  execution of commands
• no different from pre-telegraph era!

24
Lose Not HopeSome tough problems eventually get
solved
25
Strict Consistency at Low Bandwidth

• Very difficult challenge to offer good
  performance
• especially challenging for databases
• most approaches weaken consistency
• (both in the file system and database worlds)
• Recent work by Niraj Tolia offers a better
  approach
• key insight even a stale replica can be useful
• Cedar optimistically exploits them to reduce data
  transmission volume
• large replicas ok as laptop disks grow bigger
• eliminates heroic efforts at replica control
• occasional (lazy) replacement of client replica
  is fine
• periodic transmission of updates by server to
  clients also ok
• no weakening of consistency or transactional
  semantics

26
Protocol Overview
select shown update goes directly to server
Use client resources to reduce network
transmission
27
Cedar Implementation
Completely transparent to databases and
applications
28
Mobile Sales Benchmark(inspired by TPC-App)
29
To Carry or to Find?
30
Phenomenal Hardware Progress
31
Meanwhile . . . .
32
(No Transcript)
33
(No Transcript)
34
Just for Comparison . . . .
35
How will almost-free hardwarechange our world?

• Can lead to hands-free mobility!

36
Carry vs. Find
Air
Light
Water
Telephone (pre-2000)
Telephone (post-2000)
Personal Computing (now)
Personal Computing (future)
37
Key Observation

• Anticipated worst case scenario is determining
  factor
• Find ? resource available at all visited
  locations
• Carry ? reasonable doubt at least at one
  location
• Ubiquity can substitute for portability
• Can we do this for personal computing?
• ? goal of Internet Suspend/Resume effort

38
Two Approaches to Mobility

• Think wallet, not Swiss Army Knife!