Keshav Pingali

1
The Hunt for Right Abstractions

• Keshav Pingali
• Department of Computer Science
• University of Texas, Austin

2
Dewars profile

• Bachelor of Technology (1978)
• I.I.T.Kanpur
• S.M. / E.E./Sc.D. (1983,1986)
• M.I.T., Cambridge, MA
• Faculty member at Cornell University (1986-2006)
• Department of Computer Science
• Department of Electrical and Computer Engineering
• Faculty member at UT, Austin (2006-)
• Department of Computer Science
• Director, Center for Grid and Distributed
  Computing, Institute for Computational
  Engineering and Sciences (ICES)
• Favorite hobby
• writing newspaper columns on current political
  events
• Favorite beer
• Corona
• Favorite quote
• Everything should be as simple as possible, but
  not any simpler. Einstein

3
High-level theme of talk

• My definitions
• PL design
• finding abstractions to express algorithms
• PL implementation research
• finding out if these are good abstractions
• Questions
• What are abstractions?
• Why are abstractions useful?
• What does abstraction have to do with programming
  languages?
• What are good abstractions in the context of
  programming languages?

4
The Power of Abstraction

• Bridges of Konigsberg
• town in Prussia
• now named Kaliningrad in Russia
• Problem
• Is there a walk that crosses each bridge exactly
  once?
• Citizens of Konigsberg in the 17th and 18th
  centuries tried to solve this problem

5
Solution by Euler

• Key insight connectivity between land masses is
  what is important, not the actual distances or
  the orientations of the bridges
• Create an abstraction graph
• One node for each land mass
• Edge between two nodes if there is a bridge
  connecting the two land masses
• Graph has nodes of odd degree, so there is no
  walk with desired property
• Led to field we now call topology

6
Abstraction

• Definition
• the act or process of leaving out of
  consideration one or more properties of a complex
  object so as to focus on others
• (e.g.) Euler left out distances and orientations
• a general concept formed by extracting common
  features from specific examples
• (e.g.) general concept of graph allows you to
  solve similar problems

Bridges of Madison County (Iowa)
7
Bad abstractions

• Abstraction is bad if it has thrown away some
  essential feature of the problem
• topologist is someone who does not know the
  difference between a doughnut and coffee-cup
• What is essential depends on the use you intend
  to make of the abstraction

Naked Blue IV Henri Matisse (1952)
8
Abstractions in PL

• My opinion most important advances in PL have
  introduced new abstractions
• Examples
• Procedures (1950?)
• abstraction parameterized code module
  (l-abstraction)
• abstracted away implementation code
• Processor architecture (IBM 360)
• abstraction machine language
• abstracted away implementation of machine
  language
• FORTRAN I (1957)
• abstraction high-level programming language
• abstracted away machine language

9
Abstractions in PL (contd.)

• Examples (contd.)
• Structured programming (1967)
• abstraction structured control-flow constructs
  like if-then-else, while-loops, for-loops etc.
• abstracted away conditional jumps (machine
  language relic)
• Object-oriented programming (1970-)
• abstraction abstract data type
• abstracted away representation of data type
• Automatic storage management (1960-)
• abstraction objects
• abstracted away machine addresses (pointers)

10
Bad abstractions in PL?

• Abstractions that are
• difficult to reason about
• hard to implement efficiently
• Example functional languages for parallel
  programming
• Abstract away the notion of storage only values
  and functions on values
• Elegant parallel execution models reduction,
  dataflow
• Big problem data structure manipulation can be
  very inefficient if you view data structures as
  values
• Unfortunately, parallelism in algorithms is
  mostly data parallelism ?
• Notion of storage might be an essential feature
  of program execution that should not be
  abstracted away by the programming language
• Caveat functional languages are perfectly good
  abstractions for other uses
• Semantics of programming languages

11
Recap

• My views
• PL design
• finding abstractions to express algorithms
• PL implementation research
• finding out if these are good abstractions
• OK, so how do you get started in PL
  implementation research?

12
Getting started

• Take relevant courses
• Modern compiler design
• there is lots more to compilers than the Dragon
  book
• you will learn powerful algorithms and mechanisms
  for implementing abstractions
• Algorithms
• Machine learning
• used increasingly in PL implementation
• Statistics and inference
• learn good experimental protocols
• we do not do enough of this in Computer Science

13

• Read papers from major PL conferences
• PLDI Programming Language Design and
  Implementation
• POPL Principles of Programming Languages
• OOPSLA Object-oriented Programming Systems,
  Languages and Architectures
• ASPLOS Architectural Support for Programming
  Languages and Operating Systems
• PPoPP Principles and Practice of Parallel
  Programming

14

• Try to replicate results from papers that
  interest you
• standard apprenticeship technique in other
  sciences
• we do not do enough of this in Computer Science
• dont hesitate to contact authors if something is
  not clear or you cannot replicate their results
  (but be polite)
• Build on other peoples software
• starting from scratch is too time-consuming
• not even feasible for most research projects
  nowadays
• always be on the look-out for tools and systems
  you can steal (Open64 compiler from Intel, Java
  compilers and JVMs like Jikes,)

15

• Do NOT try to tweak other peoples heuristics in
  minor ways
• Most problems in PL implementation are
  NP-complete or undecidable at some level of
  generality
• Heuristics are the name of the game
• Not difficult to add bells and whistles to other
  peoples heuristics
• You go breadth-first, I go depth-first
• Bulk of papers in conferences are of this nature
• Life is too short to play second fiddle
• The scenery never changes if you are not the lead
  husky.
• Hunt for abstractions
• write lots of code and think about what might
  simplify your job
• talk to lots of people who write lots of code and
  think about what might simplify their job
• Implementing abstractions is critical
• only way to determine if an abstraction is right
• use your prototype and give it to your friends to
  use
• Encourage and learn from feedback
• Few abstractions are perfect the first time
  around

16
The Hunt For Right Abstractions