Transcript CMU job talk - Carnegie Mellon University

Globally-Distributed Software Development:
The Bell Labs Collaboratory
James Herbsleb
Bell Labs
Lucent Technologies
2701 Lucent Lane
Lisle, IL 60532 USA
+1 630 713 1869
[email protected]
Collaboratory Research Team
• University of
Michigan
– Tom Finholt
– Mark Handel
• Carnegie Mellon
University
– Alberto Espinosa
• Bell Labs Research
–
–
–
–
–
–
–
–
David Atkins
David Boyer
James Herbsleb
Stacie Hibino
Audris Mockus
Dewayne Perry
Larry Votta
Graham Wills
2
Collaboratory Project
Tools
TeamPortal
Models of
Development
How to distribute work
across global sites.
Best
Practices
Planning Travel
xxxxxxxxxxxxxxxxx
xxxxxx
Establishing Liaisons
xxxxxx xxxxx xxxxxxxxxxxxxx
xxxxx xxxx
xxxxxxxxxxxxxxxx
Rear View Mirror
Building Trust
xxxxxxx xxxxx
xxxxxxxxxx
CalendarBot
Communication Etiquette
xxxxxxxxxxxxxxxxxxx
xxxxxxxx xxxxxx
Preventing Delay
Design
Experience Browser
Research
Team
New Products
Naperville
Swindon
Malmesbury Dublin
Chippenham
xxxxxxxx xxx xxxxxx
xxxxxxxxxxx
xxxxxx xxxxxxx
Code
Using Commercial Tools
Test
xxxxxxxxx
xxxxxxx xxxx xxxxxxx
xxxxxx xxxxxx
Global Development
Solutions
Hilversum
Huizen
Nuremberg
Paris
Columbus
Brussels
Bangalore
Empirical Studies
3
Multi-site Delay
Work
Days
30
Modification Request (MR) interval
Last Modification - First Modification
All changes July 1997 to July,1999
multi site
single site
18.1
20
12.7
10
0
4.9
Network Element A
6.9
Network Element B
4
Multi-site
H1
Multi-site work just
takes longer
H2
Multi-site MRs are
larger, take longer
H3
Multi-site MRs are
more diffuse, take
longer
H4
Multi-site MRs
involve more
people, take longer
Number of People
Work Interval
Size
Diffusion
Modeling Interval
Variable
MR interval
Number of people
Diffusion
Size
Time
Severity
Fix
Multi-site
Measure used in models
Log of number of days, first
delta to last delta
Log of number of people
Log of number of modules
touched by change
Log of number of delta
Date
Is high severity
Is fix
Set of sites of all actors has
more than one element
6
Graphical model of work interval for Network Element A
Multi-site
H1
Multi-site work just
takes longer
H2
Multi-site MRs are
larger, take longer
H3
Multi-site MRs are
more diffuse, take
longer
H4
Multi-site MRs
involve more
people, take longer
199.7
0.27
Number of People
154.1
0.24
Work Interval
35.9
0.12
Size
Diffusion
148.9
0.25
Graphical model of work interval for Network Element A
Multi-site
199.7
0.27
Number of People
154.1
0.24
Work Interval
35.9
0.12
Size
148.9
0.25
Diffusion
Graphical model of work interval for Network Element A (left) and B (right)
Multi-site
Multi-site
199.7
0.27
2009.7
0.55
Number of People
Number of People
154.1
0.24
566.8
0.25
Work Interval
35.9
0.12
Size
Diffusion
Work Interval
701.7
0.34
148.9
0.25
Size
96.2
-0.13
Diffusion
Graphical model of work interval for Network Element A (left) and B (right)
MR Interval
Distance Requires More People?
• MR is assigned to “owner” who recruits
others
• Multi-site requires more people?
– Who is the right expert?
– Can MR owner get “right person” to do the
work?
– “Partial” expertise, several people
– Correct errors
11
Collaboratory Project
Tools
TeamPortal
Models of
Development
How to distribute work
across global sites.
Best
Practices
Planning Travel
xxxxxxxxxxxxxxxxx
xxxxxx
Establishing Liaisons
xxxxxx xxxxx xxxxxxxxxxxxxx
xxxxx xxxx
xxxxxxxxxxxxxxxx
Rear View Mirror
Building Trust
xxxxxxx xxxxx
xxxxxxxxxx
CalendarBot
Communication Etiquette
xxxxxxxxxxxxxxxxxxx
xxxxxxxx xxxxxx
Preventing Delay
Design
Experience Browser
Research
Team
New Products
Naperville
Swindon
Malmesbury Dublin
Chippenham
xxxxxxxx xxx xxxxxx
xxxxxxxxxxx
xxxxxx xxxxxxx
Code
Using Commercial Tools
Test
xxxxxxxxx
xxxxxxx xxxx xxxxxxx
xxxxxx xxxxxx
Global Development
Solutions
Hilversum
Huizen
Nuremberg
Paris
Columbus
Brussels
Bangalore
Empirical Studies
12
How to Identify Experts?
• Use change history
– Developers use spontaneously
– Previous tool that used several sources of data to
identify expert (McDonald & Ackerman)
• Requirements
–
–
–
–
–
Identify experts quickly and easily
Do not overburden a few individuals
Allow the user to find alternatives
No descriptions of expertise
No additional data collection, e.g., social networks
13
Quantitative Basis for Expertise
• Why quantify expertise?
– Compare potential experts
– Identify distributions of expertise, i.e., broad, as
well as narrowly specialized
• Experience atoms (EAs)
– Smallest meaningful unit of experience – delta
– Has several dimensions, e.g.,
• part of the product
• tools and technology used
• Person, team, organization that made change
14
Experience Browser
15
Collaboratory Project
Tools
TeamPortal
Models of
Development
How to distribute work
across global sites.
Best
Practices
Planning Travel
xxxxxxxxxxxxxxxxx
xxxxxx
Establishing Liaisons
xxxxxx xxxxx xxxxxxxxxxxxxx
xxxxx xxxx
xxxxxxxxxxxxxxxx
Rear View Mirror
Building Trust
xxxxxxx xxxxx
xxxxxxxxxx
CalendarBot
Communication Etiquette
xxxxxxxxxxxxxxxxxxx
xxxxxxxx xxxxxx
Preventing Delay
Design
Experience Browser
Research
Team
New Products
Naperville
Swindon
Malmesbury Dublin
Chippenham
xxxxxxxx xxx xxxxxx
xxxxxxxxxxx
xxxxxx xxxxxxx
Code
Using Commercial Tools
Test
xxxxxxxxx
xxxxxxx xxxx xxxxxxx
xxxxxx xxxxxx
Global Development
Solutions
Hilversum
Huizen
Nuremberg
Paris
Columbus
Brussels
Bangalore
Empirical Studies
16
Extended Bridge Use
• Observed open bridge for flexible
collaboration
• Example: switch in Mexico City has a
problem
– kept bridge open for 5 weeks
– composition of team very dynamic
– greatly simplified communication
– seemed to speed problem resolution
17
Observation to Speculation
• Do users’ practices mirror circuit technology?
• Flexible, “usually-open” voice channels
– billing could be based on bandwidth used
– conversations represented visually
– virtual warroom
• With voice over IP, “talking over distance” and
“call” will not be synonymous
– little need for “meetings,” “calls” and similar
discrete events
18
OpenChannel
From Speculation to Service Design
OpenChannels
TradeNet
TradeNet
estimation
Audris Mockus
New Jersey
Mark Handel
Palo Alto
David Atkins
Palo Alto
Anjum Khan
London
Sue Cummins
London
Don Fong
Chicago
Ron Davis
New Jersey
Ron Davis
New Jersey
John Spillane
Palo Alto
Jan Goff
Palo Alto
Don Fong
Chicago
Anjum Khan
Palo Alto
Jim Hill
London
Participate
Monitor
Excel
Participate
Excel
Monitors (listen only)
TradeNet:proposal1:hdw
TradeNet:proposal 1
Monitor
TradeNet:proposal 2
Replay
Monitor
Replay
Replay
Participate
Become a
participant
Change from
participate to
monitor
Participants
(listen and speak)
Shared documents
Replay the
last few
minutes
19
We need to finish our preparations for the review!
ConnectIcon
• Antidote for phone
tag
• Send presence and
contact ability to
anyone
Current spec.
http://www-spr.research.bell-labs.co
ConnectIcon
Hi Jim,
We need to talk about the review tomorrow!
Ann Kelly
To check my availability and get my contact information, please click this link:
ConnectIcon from Ann Kelly
ConnectIcon
3 days ago
Currently
in use
23 hours ago
Busy
10
20 minutes
hours ago
ago
26 hours ago
Collaboratory Results
• Substantial use of tools
• Effective practices identified and introduced
• Little or no use of organizational models
• Overall effect difficult to assess quantitatively,
given organizational changes
• Qualitative data suggest positive effects
• Insights into communication and coordination
23
Research Directions
• Collaboration technologies
– Central role of informal, unscheduled communication
– How to create the “virtual 30 meters”
• Conway’s law – software architecture and
organizational structure
– Conway postulated homomorphism
– Understanding relationship is critical for bringing
architecture research into real world
– What architectures are feasible for given
organizational structure?
– What organizational structures are feasible for a given
architecture?
24