Contents
Copyright ©Cornell University |
High Performance Systems Software
People
-
Faculty
-
Researchers
-
Graduate Students
Research
Program
Analysis
Algorithmic breakthroughs made by our project have enabled us to design
fast and practical algorithms for several fundamental program analysis
problems.
-
Fractal symbolic analysis
is a new analysis technique for use in restructuring compilers to verify
legality of program transformations. It combines the power of symbolic
program analysis with the tractability of dependence analysis, and it permits
restructuring
of far more complex codes than current technology can handle.
-
We have developed an intermediate representation
called the Abstract Matrix Form (AMF)
and we have used it to extract matrix operations from low-level code. This
led us to a new approach to optimizing programs
in languages like MATLAB.
-
Optimal algorithm for computing control dependence:
-
This algorithm computes a data structure called
APT
which answers queries for control dependence successors and predecessors
in
time proportional to output size. Preprocessing time is
O(|E|).
This is an improvement over previous algorithms which took O(|E||V|)
space and time (|E| and |V| are the number of edges and nodes in the control
flow graph).
-
A sample implementation of the APT data structure for computing control
dependence, as described in the TOPLAS paper of Pingali and Bilardi can
be found here . It implements
cd, conds and cdeq queries.
-
Optimal algorithms for determining control regions:
-
The problem is to compute all nodes in a control flow graph that have the
same control dependence predecessors as a given node. We have developed
two very different algorithms for this problem. One of the algorithms
requires computing the postdominator tree of the control flow graph, and
requires O(|E|) space and time. The other
algorithm
is based on cycle-equivalence, and it too requires O(|E|)
time. These are improvements over previous algorithms which required O(|E||V|)
time.
-
Optimal algorithm for weak control dependence:
-
For program verification, Clarke and Podgurski have proposed a variation
of control dependence called weak control dependence. Their algorithm
for computing this relation takes
O(|E|3) time.
Our algorithm computes this relation in
O(|E|)
time.
-
Generalized control dependence:
-
We have defined a generalized control dependence
that subsumes both standard control dependence and weak control dependence,
and shown how to compute generalized control dependence predecessors, successors
and regions optimally.
-
This
paper describes how to compute interprocedural dominators and interprocedural
control dependence in O(|E||V|) time.
-
O(|E|) time algorithm to compute the
SSA form of programs:
-
This algorithm performs-function
placement for computing the Static Single-Assignment (SSA) form of programs
in O(|E|) time. On the SPEC benchmarks, this algorithm runs
5 times faster than other O(|E|) time algorithms for this
problem.
-
Memory
Hierarchy Optimizations
The performance of most codes on fast machines is limited by the
so-called memory wall. Processors run much faster than memory, so if a
code touches large amounts of data, the processor spends most of its cycles
waiting for memory requests to complete rather than in doing computations.
Many of these codes can benefit from restructuring to improve locality
of reference. It is tedious to do this restructuring by hand, so automatic
restructuring technology for accomplishing this goal is highly advantageous.
Our group is a world-leader in the area of compiler technology
for memory hierarchy optimization. Compiler product-lines of companies
like Intel, Hewlett-Packard and Silicon Graphics incorporate technology
developed by our group.
-
Imperfectly-nested loop transformations: Imperfectly-nested loops are loops
that have statements nested in some but not all of the loops in the loop
nest.
-
Automatic generation of divide-and-conquer codes from iterative codes
-
Data-centric Blocking for Memory Hierarchies:
-
Transformations for perfectly nested loops:
-
We have invented a loop transformation framework
based on integer non-singular matrices. This framework subsumes loop permutation,
skewing, scaling and reversal.
-
We implemented the LAMBDA loop transformation toolkit
based on this framework. A paper on LAMBDA won the
best paper award at ASPLOS V.
-
We used LAMBDA to enhance parallelism and locality of reference in perfectly
nested loops. A detailed description of the experiments with LAMBDA can
be found in our TOCS paper.
-
The theory behind LAMBDA is described in Wei Li's PhD thesis.
-
Hewlett-Packard has adopted the technology
in LAMBDA for performing loop transformations in its entire compiler product
line.
-
Intel has recently licensed the LAMBDA
toolkit for use in Merced compilers.
-
Transformations for imperfectly nested loops:
-
We have extended the theory behind LAMBDA to handle imperfectly nested
loop transformations like distribution and fusion. This approach has been
implemented in the MU toolkit.
-
A paper on MU was nominated for the Best Student Paper
award at Supercomputing '96.
Next-generation
Generic Programming
We are implementing a novel generic programming system
for simplifying the job of writing computational science codes that use
sparse matrices. This system uses two API's; a high-level one for the algorithm
designer and low-level one for the sparse matrix format implementor. Restructuring
compiler technology is used to translate programs written in the high-level
API into those that use the low-level API.
One view of this system is that it is an aspect-oriented
system which uses restructuring compiler technology to weave sparse format
aspects
into the functional specification of the algorithm.
-
The API's in the generic programming system is described in this
paper.
-
An elegant restructuring compiler technology for translating between these
API's is described in this paper. This
is the most up-to-date description of the compiler technology.
-
This paper introduces the data-centric
code generation model that is the foundation of the low-level API, and
shows how it can be viewed abstractly as the optimization of certain relational
queries generated from the user program.
-
A paper in SIAM '97 discusses how sparse matrix formats
can be described to the compiler for use in optimization and code generation.
-
A paper in Supercomputing '97 shows how the relational
approach can be used to generate
parallel sparse matrix code.
-
A unified approach to compiling both dense and sparse matrix programs can
be found here.
-
Paul Stodghill's PhD thesis contains a detailed
description of compiling sparse codes for sequential machines.
-
Vladimir Kotlyar's PhD thesis contains details
on how to extend the sequential techniques to parallel codes, and to codes
that are imperfectly-nested and have dependencies.
-
Slides from Kotlyar's job talk summarize
the work that he did for his thesis.
Irregular
Applications
Our project has developed parallel algorithms for several unstructured
and semi-structured applications including Structured Adaptive Mesh Refinement
(SAMR) and circuit simulation. SAMR is used to simulate physical phenomena
like shock wave propagation for which efficient simulation requires a grid
whose coarseness varies with time. We have also participated in the development
of a parallel MATLAB environment called MultiMATLAB.
-
SAMR:
-
A paper in ICS '97 describes compiler and runtime
support for SAMR.
-
A more applications-oriented paper can be found
here.
-
MultiMATLAB:
-
Vijay Menon has worked with Anne Trefethen in the implementation of a parallel
MATLAB system called MultiMATLAB.
-
Circuit Simulation:
-
We have developed a new technique for compiled zero
delay logic simulation which partitions the circuit into fanout free regions
(FFRs), transforms each region into a linear sized BDD (binary decision
diagram), and converts each BDD into executable code. On standard benchmarks,
we observed a performance improvement of upto 67%.
Program
Representation
The dependence flow graph(DFG) is an intermediate program representation
that unifies control flow and dataflow information. It has an operational
semantics based on the dataflow model of computation. It can be used for
performing standard dataflow analyses, and for program debugging using
slicing.
-
An introduction to the dependence flow graph (DFG) and its operational
semantics can be found in our POPL '91 paper.
-
Algorithms for contructing the DFG can be found in our
JPDC
paper.
-
We have also explored an extension of the DFG that
incorporates distance and direction information.
-
Micah Beck's PhD thesis has a detailed description of the construction
of DFGs.
-
Richard Johnson's PhD thesis contains a detailed
discussion of the DFG and a related data structure called the Quick Propagation
Graph (QPG), and their use for program analyses.
-
IBM's VLIW compiler
uses the DFG as its internal representation.
Compiling
for Parallel Machines
Our project implemented one of the first compilers that generated message-passing
code from sequential shared-memory programs with data distribution directives.
We introduced several concepts like runtime resolution and the owner-computes
rule that have become part of the standard terminology in this area.
-
A survey paper on parallel languages can be found
here.
-
Automatic alignment:
-
We have developed an algorithm to compute data and
computation alignment automatically by reducing the alignment problem to
the standard problem of finding a basis for the null space of a matrix.
-
Code Generation Techniques:
-
A paper in PLDI'89 described a code generation strategy
based on runtime resolution and the owner-computes rule.
-
We implemented a compiler, based on this approach, to translate programs
in the dataflow language Id Nouveau into message-passing code for the Intel
iPSC/2. A summary can be found in our TPDS paper.
-
A performance analysis of the code generated by our compiler for the SIMPLE
benchmark from Los Alamos can be found
here.
-
Anne Rogers' PhD thesis has a complete description.
Instruction-level
Parallelism
Our project has developed algorithms for software pipelining. We have also
collaborated with Stamatis Vassiliadis from IBM Glendale Labs in the design
of superscalar processor that implements register renaming, dynamic speculation
and precise interrupts in hardware.
We have developed a software pipelining scheme called bidirectional
slack scheduling which generates loop code with minimal register pressure
without sacrificing the loop's minimum execution time.
We have collaborated with Stamatis Vassiliadis of IBM Glendale Labs in
the design of a superscalar processor that implements
register renaming, dynamic speculation and precise interrupts in hardware.
A detailed discussion of the superscalar processor and related software
issues can be found in Mayan Moudgill's PhD thesis.
Functional
Languages with Logic Variables
The addition of logic variables to functional languages gives the programmer
novel and powerful tools such as incremental definition of data structures
through constraint intersection. Id Nouveau is one such language
for which we have given a formal semantic account using the notion of closure
operators on a Scott domain. We have also used logic variables to model
demand propagation in the implementation of lazy functional languages,
and we have explored the use of accumulators, which are generalized
logic variables.
-
Language Constructs:
-
In joint work with Arvind and Rishiyur Nikhil, a weak form of logic variable
called
I-structures was added to the dataflow language
Id. This work was the origin of the language Id Nouveau.
-
With K. Ekanadham of IBM Hawthorne, we generalized the idea of a logic
variable to obtain a construct called the
accumulator.
-
Semantics of Id Nouveau:
-
Lazy Evaluation:
-
We have shown that demand propagation in lazy functional
languages can be modeled using logic variables. This makes it possible
to expose the process of demand propagation to the compiler of a lazy language,
permitting optimization of demand propagation code.
Software
The following packages are available for downloading.
-
A sample implementation of the APT data structure for computing control
dependence, as described in the TOPLAS paper of Pingali and Bilardi can
be found here. It implements cd,
conds and cdeq queries.
List of Papers
-
Nikolay Mateev, Keshav Pingali, Paul Stodghill,
and Vladimir Kotlyar.
Next-generation Generic Programming
and its Application to Sparse Matrix Computations.
In International Conference on Supercomputing, 2000.
-
Nawaaz Ahmed, Nikolay Mateev, and Keshav Pingali.
Synthesizing Transformations for
Locality Enhancement of Imperfectly-nested Loop Nests.
In International Conference on Supercomputing, 2000.
-
Nikolay Mateev, Vijay Menon, and Keshav Pingali.
Fractal
Symbolic Analysis for Program Transformations.
Technical Report TR2000-1781, Cornell Computer Science Department,
2000.
-
Nawaaz Ahmed, Nikolay Mateev, and Keshav
Pingali.
Tiling
Imperfectly-nested Loop Nests.
Technical Report TR2000-1782, Cornell Computer Science Department,
2000.
-
Nawaaz Ahmed, Nikolay Mateev, Keshav Pingali,
and Paul Stodghill.
Compiling Imperfectly-nested Sparse
Matrix Codes with Dependences.
Technical Report TR2000-1788, Cornell Computer Science Department,
2000.
-
James Ezick, Gianfranco Bilardi, and Keshav
Pingali.
Efficient Computation of Interprocedural
Control Dependence.
Submitted for publication, 2000.
-
Gianfranco Bilardi and Keshav Pingali.
The Static Single Assignment Form and its Computation.
Submitted for publication, 2000.
-
Nawaaz Ahmed and Keshav Pingali.
Automatic Generation of Block-recursive
Codes.
Submitted for publication, 2000.
-
Nikolay Mateev, Vijay Menon, and Keshav Pingali.
Left to Right and vice versa: Applying Fractal
Symbolic Analysis to Restructuring Linear Algebra Codes.
Submitted for publication, 2000.
-
Vijay Menon and Keshav Pingali.
A Case for Source-Level Transformations in
MATLAB.
In The 2nd Conference on Domain-Specific Languages, 1999.
-
Vijay Menon and Keshav Pingali.
High-Level Semantic Optimization of
Numerical Codes.
In International Conference on Supercomputing, 1999.
-
Induprakas Kodukula, Keshav Pingali,
Robert Cox, and Dror Maydan.
An experimental evaluation of tiling
and shackling for memory hierarchy management.
In International Conference on Supercomputing, 1999.
-
Keshav Pingali.
Parallel and Vector Programming Languages.
In Wiley Encyclopedia of Electrical and Electronics Engineering,
vol.15, 1999.
-
Vladimir Kotlyar.
Relational Query Processing
Approach to Compiling Sparse Matrix Codes.
slides from a talk, 1999.
-
Vladimir Kotlyar.
Relational Algebraic Techniques
for the Synthesis of Sparse Matrix Programs.
PhD thesis, Cornell University, 1999.
-
Induprakas Kodukula.
Data-centric Compilation.
Ph.D. Thesis, Cornell Universiry.
-
Vladimir Kotlyar, Keshav Pingali, and Paul Stodghill.
Compiling Parallel Code for Sparse Matrix
Applications.
In Supercomputing, November 1997.
-
Vijay Menon and Anne E. Trefethen.
MultiMATLAB: Integrating MATLAB with High-Performance
Parallel Computing.
In Supercomputing, November 1997.
-
Induprakas Kodukula, Nawaaz Ahmed, and Keshav Pingali.
Data-centric Multi-level Blocking.
In Programming Language Design and Implementation, June 1997.
-
Keshav Pingali and Gianfranco Bilardi.
Optimal Control Dependence and the Roman
Chariots Problem.
ACM Transactions on Programming Languages and Systems (TOPLAS),
19(3), May 1997.
-
Nikos Chrisochoides, Induprakas Kodukula, and Keshav
Pingali.
Compiler and run-time support for semi-structured
applications.
In International Conference on Supercomputing, 1997.
-
Nikos Chrisochoides, Induprakas Kodukula, and Keshav
Pingali.
Compiler support for easing the programmer's
burden.
In Workshop on Structured Adaptive Mesh Refinement Grid Methods,
1997.
-
Nikos Chrisochoides, Induprakas Kodukula, and Keshav
Pingali.
Data Movement and Control Substrate for
network-based Parallel Scientific Computing.
In CANPC '97, 1997.
-
Vladimir Kotlyar, Keshav Pingali, and Paul Stodghill.
Compiling Parallel Sparse Code for User-Defined
Data Structures.
In SIAM Conference on Parallel Processing for Scientific Computing,
volume 8, 1997.
-
Vladimir Kotlyar, Keshav Pingali, and Paul Stodghill.
A Relational Approach to Sparse Matrix
Compilation.
In EuroPar, 1997.
-
Vladimir Kotlyar, Keshav Pingali, and Paul Stodghill.
Unified
framework for sparse and dense SPMD code generation.
Technical Report 97-1625, Cornell Computer Science Department, 1997.
-
Paul Stodghill.
A
Relational Approach to the Automatic Generation of Sequential Sparse Matrix
Codes.
PhD thesis, Cornell University, 1997.
-
Induprakas Kodukula and Keshav Pingali.
Transformations of Imperfectly Nested Loops.
In Supercomputing, November 1996.
-
Gianfranco Bilardi and Keshav Pingali.
A Framework for Generalized Control Dependence.
In Programming Language Design and Implementation. SIGPLAN,
June 1996.
-
Sudeep Gupta and Keshav Pingali.
Fast
Compiled Logic Simulation Using Linear BDDs.
Technical Report TR95-1522, Cornell Computer Science Department, 1995.
-
Vladimir Kotlyar, Keshav Pingali, and Paul Stodghill.
Automatic Parallelization of the Conjugate
Gradient Algorithm.
In Languages and Compilers for Parallel Computers, volume 8,
1995.
-
Keshav Pingali and Gianfranco Bilardi.
APT: A Data Structure for Optimal Control
Dependence Computation.
In Programming Language Design and Implementation. SIGPLAN,
1995.
-
Wei Li and Keshav Pingali.
The
LAMBDA Loop Transformation Toolkit.
Technical Report TR94-1431, Cornell Computer Science Department, 1994.
-
Mayan Moudgill.
Implementing
and Exploiting Static Speculation on Multiple Instruction Issue Processors.
PhD thesis, Cornell University, August 1994.
-
Vladimir Kotlyar, Induprakas Kodukula, Keshav Pingali,
and Paul Stodghill.
Solving Alignment Using Elementary Linear
Algebra.
In Languages and Compilers for Parallel Computers, volume 7,
1994.
-
Richard Johnson, David Pearson, and Keshav Pingali.
The Program Structure Tree: Computing Control
Regions in Linear Time.
In Programming Language Design and Implementation. SIGPLAN,
1994.
-
Richard Johnson.
Efficient
Program Analysis Using Dependence Flow Graphs.
PhD thesis, Cornell University, 1994.
-
Wei Li.
Compiling
for NUMA Parallel Machines.
PhD thesis, Cornell University, 1994.
-
Richard Huff.
Lifetime-Sensitive Modulo Scheduling.
In Programming Language Design and Implementation. SIGPLAN,
1993.
-
Richard Johnson and Keshav Pingali.
Dependence Based Program Analysis.
In Programming Language Design and Implementation. SIGPLAN,
1993.
-
Mayan Moudgill, Keshav Pingali, and Stamatis Vassiliadis.
Register Renaming and Dynamic Speculation:
an Alternative Approach.
In MICRO, 1993.
-
Radha Jagadeesan and Keshav Pingali.
Abstract
Semantics for a Higher-Order Functional Language with Logic Variables.
In Principles of Programming Languages, volume 19, 1992.
Available as Cornell CS TR91-1220.
-
Wei Li and Keshav Pingali.
Access Normalization: Loop Restructuring
for NUMA Compilers.
ACM Transactions on Computer Systems, 11(4), November 1993.
-
Wei Li and Keshav Pingali.
Access Normalization: Loop Restructuring
for NUMA Computers.
In Architectural Support for Programming Languages and Operating
Systems, volume 5, 1992.
An extended version appeared in ACM Transactions on Computer Systems.
-
Richard Johnson, Wei Li, and Keshav Pingali.
An Executable Representation of Distance
and Direction.
In Languages and Compilers for Parallel Computers, volume 4,
1991.
-
Radha Jagadeesan, Keshav Pingali, and Prakash Panagaden.
A
Fully Abstract Semantics for a Functional Langauge with Logic Variables.
ACM Transactions on Programming Languages and Systems, 13(4),
October 1991.
Available as Cornell CS TR89-969.
-
Radha Jagadeesan.
Investigations
into Abstractions and Concurrency.
PhD thesis, August 1991.
-
Wei Li and Keshav Pingali.
A Singular Loop Transformation Framework
Based on Nonsingular Matrices.
International Journal of Parallel Processing, 22(2), April 1991.
-
Micah Beck, Richard Johnson, and Keshav Pingali.
From Control Flow to Data Flow.
Journal of Parallel and Distributed Computing, 12, 1991.
-
Keshav Pingali, Micah Beck, Richard Johnson, Mayan
Moudgill, and Paul Stodghill.
Dependence Flow Graphs: An Algebraic Approach
to Program Dependencies.
In Principles of Programming Languages, volume 18. SIGPLAN,
1991.
-
Anne Rogers and Keshav Pingali.
Compiling for Distributed Memory Machines.
IEEE Transactions on Parallel and Distributed Systems, 5(3),
March 1994.
-
Keshav Pingali and Kattamuri Ekanadham.
Accumulators:
New Logic Variable Abstractions for Functional Languages.
In Theoretical Computer Science, volume 81, 1991.
Available as Cornell CS TR91-1231.
-
Arvind, Rishiyur Nikhil, and Keshav Pingali.
I-structures:
Data Structures for Parallel Computing.
ACM Transactions on Programming Languages and Systems, 11(4),
October 1989.
Available as Cornell CS TR87-810.
-
Keshav Pingali and Kattamuri Ekanadham.
Accumulators:
New Logic Variable Abstractions for Functional Languages.
Theoretical Computer Science, 81, 1991.
Available as Cornell CS TR91-1231.
-
Anne Rogers.
Compiling
for Locality of Reference.
PhD thesis, Cornell University, 1991.
-
Anne Rogers and Keshav Pingali.
Process
Decomposition Through Locality of Reference.
In Programming Language Design and Implementation, June 1989.
-
Keshav Pingali and Anne Rogers.
Compiler
Parallelization of SIMPLE for a Distributed Memory Machine.
In International Conference on Parallel Processing, August 1989.
-
Keshav Pingali.
Lazy
Evaluation and the Logic Variable.
Addison-Wesley, 1987.
Available as Cornell CS TR88-952.
|