the user in experimental computer systems research peter a. dinda gokhan memik, robert dick bin lin,...

The User In Experimental Computer Systems Research

Peter A. DindaGokhan Memik, Robert Dick

Bin Lin, Arindam Mallik, Ashish Gupta, Sam Rossoff

Department of Electrical Engineering and Computer ScienceNorthwestern University

http://presciencelab.org

2

Experimental Computer Systems Researchers Should…

• Incorporate user studies into the evaluation of systems

• Incorporate direct user feedback into the design of systems

3


• Incorporate user studies into the evaluation of systems– No such thing as the typical user– Really measure user satisfaction

• Incorporate direct user feedback into the design of systems– No such thing as the typical user– Measure and leverage user variation

4

Outline• Prescription• Experiences with user studies and direct user

feedback– User comfort with resource borrowing– User-driven scheduling of interactive VMs– User satisfaction with CPU frequency– User-driven frequency scaling– User-driven control of distributed virtualized envs.– Prospects for speculative remote display

• Principles for client/server context• General advice

5

Experiences in Detail• Concepts : ExpCS 2007 @ FCRC• Specific Projects

– User comfort with resource borrowing• HPDC 2004, NWU-CS-04-28

– User-driven scheduling of interactive VMs• Grid 2004, SC 2005, VTDC 2006, NWU-EECS-06-07

– User satisfaction with CPU frequency• CAL 2006, SIGMETRICS 2007, NWU-EECS-06-11

– User-driven frequency scaling (/process-driven voltage scaling)• CAL 2006, SIGMETRICS 2007, NWU-EECS-06-11

– User-driven control of distributed virtualized envs.• Portion of Bin Lin’s thesis, see also ICAC 2007

– Prospects for speculative remote display• NWU-EECS-06-08

6

User Comfort With Resource Borrowing

• Systems that use “spare” resources on desktops for other computation– *@Home, Condor on desktops, etc.

• How much can they borrow before discomforting user?– Inverse: How much must desktop

replacement system give?

7

User Comfort With Resource Borrowing

• Developed system for controlled resource borrowing given a profile– CPU contention, disk BW contention, physical

memory pages– User presses “irritation button” to stop

• User study – 38 participants– Four apps

• Word, Powerpoint, Web browsing, Game

– Ramp, Step, Placebo profiles– Double blinded

8

Example Result

Massive Variation inUser Response

9

User-driven Scheduling of Interactive VMs

• Virtual machine-based desktop replacement model– VM runs on backend server– User connects with remote display– VM is scheduled according to periodic real-

time model• Allows straightforward mixing of batch and

interactive VMs + isolation properties

• What should interactive VM’s schedule be?

10

User-driven Scheduling of Interactive VMs

• VSched scheduler on server

• User interface on client

Non-centering joystick allows user to set schedule

$10 interface in studyCheaper interfaces possible

Onscreen display indicates price of current schedule

Also indicates whenschedule cannot be admitted

11

User-driven Scheduling of Interactive VMs• User study

– 18 participants– 4 applications

• Word, Powerpoint, Internet browsing, Game

– Survey response + measurement• Deception scheme to control bias in survey response

• Results– Almost all could find a setting that was comfortable– Almost all could find a setting that was comfortable and

believed to be of lowest cost– Lowest cost highly variable, as expected given previous

results– <1 minute convergence typical

• Interface captures individual user tradeoffs– Fewer cycles for tolerant users– More cycles for others

12

User Satisfaction With CPU frequency

• Modern processors can lower frequency to reduce power consumption– Software control: DVFS - conservative

• How satisfied are users of different applications at different clock frequencies?

• User Study– 8 users– 3 frequencies + Windows DVFS– 3 apps

• Presentation, Animation, Game– Rate comfort on 1 to 10 scale– Double-blinded

13

Example Results

• Dramatic variation in user satisfaction for fixed frequencies– And for DVFS

Presentation

Game

14

User-driven Frequency Scaling

• Developed system to dynamically customize frequency to user– User presses “irritation button” as input– 2 very simple learning algorithms

• User study – 20 participants– Three apps

• Powerpoint, Animation, Game

– Comparison with Windows DVFS

– Double blinded

15

Example Results (Measured System Power)

Users

% g

ain

over

Win

dow

s D

VF

S

Users % g

ain

over

Win

dow

s D

VF

S

Powerpoint

Game

16




17

Principles for the Client/Server Context

• User variation– Considerable variation in user satisfaction with any

given operating point– No such thing as a typical user

• User-specified performance– Have user tell system software how satisfied he is– No decoupling of user response from user and

OS-level measurements• Think global feedback

• Thin, simple user-system interface– One bit is a lot of information compared to zero

• Learning to decrease interaction rate– Model the individual user

18




19

General Advice for Evaluating Systems with User Studies

• Consult an HCI or psychology expert– User studies are different but not impossible– At least consult the literature

• Engage your IRB early– These are “social science”-based studies– Easier the second time around

• Accept small study size– Parameter sweeps, hundreds of traces impossible– Internet volunteerism not especially effective – Use non-user studies to augment if possible– Robust statistics

20


• Accept that random sample unlikely– Selection bias estimation, if possible– Report all your data, not just summaries

• Histogram instead of curve fit

• Measure the noise floor / placebo effect– Vital to determine how much of user satisfaction is

actually under your control

• Double-blind to greatest extent possible– Investigator bias and subject bias

21


• Correlate system-level measurements with user responses to validate the latter– Consider deception when this is impossible

• Eliminate user-visible extraneous information during any study– What the user knows can hurt you

• Example: disk light

22

General Advice for Incorporating Direct User Feedback

• Out-of-band devices work best– Avoid cognitive context switch

• Use as little input as possible– One bit is much more information than zero– Utility of input may not be clear to user

• Output as little information as possible• Minimize input rate through learning• Bridge explicit feedback to implicit feedback

when possible

23


• Incorporate user studies into the evaluation of systems– No such thing as the typical user– Really measure user satisfaction

• Incorporate direct user feedback into the design of systems– No such thing as the typical user– Measure and leverage user variation

24

For MoreInformation

• Peter Dinda– http://pdinda.org

• Prescience Lab– http://presciencelab.org

25

User-driven Control of Distributed Virtual Environments

• Area of current exploration (part of Lin’s thesis)

• Idea: Can we frame these problems as games that naïve or expert users/admins can solve?

• Initial results interesting, but still too early too tell– Scaling– Dimensionality– Categorical dimensions– …

26

Typical Design Models• Optimize User

Satisfaction Subject to Constraints

• Systems software’s decisions have dramatic effect on user experience

• But how does systems software know how well it is doing?

Systems Software

Application(s)

Satisfaction with System/AppCombination

IndividualUser

Interface Considerations

Resource Managementand SchedulingConsiderations

Core API

27



• One option: let the application tell it!

• But how does the application know?

Systems Software

Application(s)


IndividualUser



Core API Policy API

28




• Assume typical user and apply general rules derived from him/her– And figure out how to

translate to the policy APISystems Software

Application(s)


TypicalUser



Core API Policy API

<500 ms latencyand <100 ms jitter

29




• Or formalize tradeoffs– And figure out how to

translate to the policy API

Systems Software

Application(s)


TypicalUser



Core API Policy API

Utility Function LatencySat

isfa

ctio

n

30



• Another option: generalize over applications and infer user experience

Systems Software

Application(s)


TypicalUser


Core API

LatencySat

isfa

ctio

nInterface Considerations

Inferred Latency

Good/Bad?

31



• Another option: Get the utility function right from the individual user– Assuming he/she

knows it…

Systems Software

Application(s)


IndividualUser


Core API


Systems Software

Application(s)


PolicyInterface

“What’s a utility function?”

“What is your utility function?” or“Which of these profiles are you most like?”

32



• Another option: Expose the system software to the user in its glory details– Works great for us!

Systems Software

Application(s)


IndividualUser


Core API


Systems Software

Application(s)


PolicyInterface

“What the…”

33

Typical Evaluation Approaches

• Workloads– User workload model/generator

• How to account for user variation?• How to evaluate as closed system?• How to validate?

– User traces• Context dependent• How to evaluate as closed system?

34

Typical Evaluation Approaches

• Metrics– Can system meet performance objectives

given through policy interface?• What should the objectives be?

– Can system optimize over some combination of utility functions?

• What should the utility functions be?

35

New Model for Characterization and Evaluation

• User studies to characterize user response– Examine the range of user satisfaction for

some perceivable quantity or combination of quantities

– Capture the variation, not only the mean– Variation = opportunity

• User studies for evaluating systems– Directly measure user satisfaction with

your system

36

New Model: Direct User Feedback

• Optimize User Satisfaction Subject to Constraints

• User conveys satisfaction (or dissatisfaction) through a simple user interface

Systems Software

Application(s)


IndividualUser


Core API


Systems Software

Application(s)


SatisfactionFeedback

37

New Model: Direct User Feedback

• Optimize User Satisfaction Subject to Constraints

• User has some direct control over systems-level decision making through a simple interface

Systems Software

Application(s)


IndividualUser


Core API


Systems Software

Application(s)


Some ControlOver DecisionMaking

38


• Virtuoso project (see virtuoso.cs.northwestern.edu)– User “rents” collection of virtual machines

• Virtuoso front-end looks like computer vendor– Providers stand up resources on which VMs can run

or communicate– Virtuoso provides adaptation mechanisms

• VM migration• Overlay topology and routing (VNet)• CPU reservations (VSched)• Network reservations (optical with VReserve)• Transparent network services (VTL)

– Virtuoso provides inference mechanisms• Application traffic and topology (VTTIF)• Network bandwidth and latency (Wren)

39


• Optimization problem: Given the inferred demands and supply, choose a configuration made possible by the adaptation mechanisms that maximizes a measure of application performance within constraints– Formalizations– NP-Hard problem in general– Approximation bound is not great either– Heuristic solutions

• Can the user or a system administrator solve these problems given the right interface?– Can a naïve human do it?

the user in experimental computer systems research peter a. dinda gokhan memik, robert dick bin lin,...

Documents

user studies

user response slide

server user interface

typical user measure

backend server user

leverage user variation

interactive vms schedule

resource borrowing systems