Falcon: A Graph Manipulation Language for Heterogeneous Systems - - PowerPoint PPT Presentation
Falcon: A Graph Manipulation Language for Heterogeneous Systems Unnikrishnan C, IISc, Bangalore Rupesh Nasre, IIT, Madras Y N Srikant, IISc, Bangalore January 20 , 2016 Introduction Experimental Results Introduction Conclusion
Unnikrishnan C, IISc, Bangalore Rupesh Nasre, IIT, Madras Y N Srikant, IISc, Bangalore January 20 , 2016
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs 1 Falcon
is a Domain Specific Language (DSL) for writing Graph algorithms.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 1 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs 1 Falcon
is a Domain Specific Language (DSL) for writing Graph algorithms.
2 Falcon
i) extends C programming language. ii) provides additional data types for Graph processing. iii) constructs for writing explicitly parallel graph algorithms.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 1 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs 1 Falcon
is a Domain Specific Language (DSL) for writing Graph algorithms.
2 Falcon
i) extends C programming language. ii) provides additional data types for Graph processing. iii) constructs for writing explicitly parallel graph algorithms.
3 Support for heterogeneous backends(CPU and GPU). Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 1 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs 1 Falcon
is a Domain Specific Language (DSL) for writing Graph algorithms.
2 Falcon
i) extends C programming language. ii) provides additional data types for Graph processing. iii) constructs for writing explicitly parallel graph algorithms.
3 Support for heterogeneous backends(CPU and GPU). 4 Supports parallel execution of different algorithms on multiple
devices.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 1 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs 1 Falcon
is a Domain Specific Language (DSL) for writing Graph algorithms.
2 Falcon
i) extends C programming language. ii) provides additional data types for Graph processing. iii) constructs for writing explicitly parallel graph algorithms.
3 Support for heterogeneous backends(CPU and GPU). 4 Supports parallel execution of different algorithms on multiple
devices.
5 Supports partitioning of Graph objects and execution of a
single algorithm using multiple devices. Used when graph
6 Supports mutation of Graph object. 7 Allows viewing Graph in different way(say collection of
triangles).
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 1 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Language constructs for parallelization and Synchronization in Falcon
single(t1) {stmt block1} else {stmt block2} The thread that gets a lock on item t1 executes stmt block1 and other threads execute stmt block2. single(coll) {stmt block1} else {stmt block2} The thread that gets a lock on all elements in the collection executes stmt block1 and others execute stmt block2.
Table 1. single statement(Synchronization) in Falcon
Data Type Iterator Description Graph points iterate over all points in graph Graph edges iterate over all edges in graph Graph pptyname iterate over all elements in new ppty. Point nbrs iterate over all neighboring points Point
iterate over dst point of outgoing edges (Directed Graph) Edge nbrs iterate over neighbor edges Set item iterate over all items in Set Collection item iterate over all items in Collection
Table 2. Iterators for foreach(parallelization) statement in Falcon parallel sections- for Multiple parallel regions on different devices.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 2 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
After Last session of HiPEAC, you want to reach home through shortest path??!!
1
int <GPU> changed = 0; // Variable on GPU
2
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
3
foreach (t In p.outnbrs)
4
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
5
}
6
main(int argc, char *argv[]) {
7
Graph hgraph; // graph on CPU
8
hgraph.addPointProperty(dist, int);
9
hgraph.getType() <GPU>graph; // graph on GPU
10
hgraph.read(argv[1]) // read graph on CPU
11
graph = hgraph; // copy graph to GPU
12
foreach (t In graph.points)t.dist=MAX INT;//INFinity
13
graph.points[0].dist = 0; // source has dist 0
14
while( 1 ){
15
changed = 0;
16
foreach (t In graph.points) relaxgraph(t,graph);
17
if (changed == 0) break; //reached fix point
18
}
19
for (int i = 0; i <graph.npoints; ++i)
20
printf(”i=%d dist=%d\n”, i, graph.points[i].dist);
21
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 400 6 400 6 PRG CDG FRA DEL INF INF INF 100 75 INF 100 75 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 3 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides Introduction Language-Data Types and Constructs
Falcon Compiler Code Genaration (Synchronization and parallelization constructs)
Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection On Collection On Collection Convert to OpenMP parallel sections Input Falcon DSL Code Check for Falcon Synchro- nization / Parallel Constructs single statement parallel sections statement foreach statement On Collection Convert to OpenMP parallel sections Single One item Convert to Compare And Swap(CAS) based code Single One item Convert to Compare And Swap(CAS) based code Single Collection Convert to code with barrier for entire paralle region Single Collection Convert to code with barrier for entire paralle region If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU:Convert to CUDA kernel call with Thrust library CPU: Convert to parallel code us- ing Galois Worklist If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma If outermost foreach statement GPU: Convert to CUDA kernel call CPU: Convert to OpenMP pragma yes no no yes
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 4 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
4 Totem- for comparing Performance on CPU, GPU and
heterogeneous execution.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
4 Totem- for comparing Performance on CPU, GPU and
heterogeneous execution.
5 Galois and Green-Marl for comparing Performance on CPU.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
4 Totem- for comparing Performance on CPU, GPU and
heterogeneous execution.
5 Galois and Green-Marl for comparing Performance on CPU. 6 LonestarGPU for comparing Performance on GPU.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
4 Totem- for comparing Performance on CPU, GPU and
heterogeneous execution.
5 Galois and Green-Marl for comparing Performance on CPU. 6 LonestarGPU for comparing Performance on GPU. 7 We were able to get performance close to and some times better
than above systems.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
1 Using Falcon compiler we wrote algorithms like BFS, SSSP and
Boruvka’s-MST.
2 We wrote dynamic algorithms like Survey Propagation(SP),
Delaunay Mesh refinement(DMR) and Dynamic-SSSP in Falcon.
3 Performance of GPU codes were compared with
i) LonestarGPU-(ISS group at the University of Texas at Austin) i) Galois- (ISS group at the University of Texas at Austin) ii)Green-Marl- DSL (PP Laboratory, Stanford University) iV)Totem( NetSysLab, University of British Columbia)
4 Totem- for comparing Performance on CPU, GPU and
heterogeneous execution.
5 Galois and Green-Marl for comparing Performance on CPU. 6 LonestarGPU for comparing Performance on GPU. 7 We were able to get performance close to and some times better
than above systems.
8 Tested on 12-core CPU and 4-GPU machine(1 Kepler and 3 Tesla ).
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 5 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution (Points,Edges) rand1(16M,64M),rand2(32M,128M) rmat1(10M,100M),rmat2(20M,200M) road1(14M,34M) road2(23M,58M)
rand1 rand2 rmat1 rmat2 road1 road2 2 4 6 8
1 1 1 1 1 1 1.15 1.36 1.8 6.69 1.3 1.95 0.71 0.75 1.19 4.09 0.62 1.13
Speedup
LonestarGPU Falcon-GPU Totem-GPU
Speedup of SSSP on GPU
r a n d 1 r a n d 2 r m a t 1 r m a t 2 r
d 1 r
d 2 1 2 3 4
1 1 1 1 1 1 1.77 1.92 1.36 2.53 1.5 2.02
Speedup
LonestarGPU Falcon-MultiGPU
Speedup- SSSP, BFS and MST in parallel on 3 GPUs
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 6 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
rand1 rand2 rmat1 rmat2 road1 road2 5 10
8.62 11.95 7.28 8.39 4.9 5.57 7.23 9.81 6.98 8.39 4.9 5.63 7.64 9.84 4.66 4.95 0.1 0.1 5.8 6.3 0.1 0.1
Speedup
Galois-12 Falcon-12 Totem-12 GreenMarl-12
(a)
SSSP speedup over Galois Single
rand1 rand2 rmat1 rmat2 road1 road2 5 10 15
8.57 8.24 6.67 7.17 6.39 5.89 13.87 12.72 8.42 8.64 6.27 6.35 8.14 8.22 6.53 7.01 0.86 0.65 5.73 6.18 0.6 0.5
Speedup
Galois-12 Falcon-12 Totem-12 GreenMarl-12
(b)
BFS speedup over Galois single
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 7 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution 1 single on collection and barrier on GPU. Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 8 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution 1 single on collection and barrier on GPU. 2 Shucai Xiao and Wu chun Feng. Inter-Block GPU
Communication via Fast Barrier Synchronization.IPDPS 2010.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 8 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution 1 single on collection and barrier on GPU. 2 Shucai Xiao and Wu chun Feng. Inter-Block GPU
Communication via Fast Barrier Synchronization.IPDPS 2010.
3 Importance of good runtime for Collection/Worklist. Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 8 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution 1 single on collection and barrier on GPU. 2 Shucai Xiao and Wu chun Feng. Inter-Block GPU
Communication via Fast Barrier Synchronization.IPDPS 2010.
3 Importance of good runtime for Collection/Worklist. 4 Ulrich Meyer and Peter Sanders. Delta-Stepping: A Parallel
Single Source Shortest Path Algorithm.EAS 1998.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 8 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
r . 5 M r 1 M r 2 M r 1 M 0.5 1 1.5 2
1 1 1 1 0.92 0.93 0.97 0.97 LonestarGPU Falcon-GPU
(a) DMR speedup
Lones- tarGPU
r . 5 M r 1 M r 2 M 10 20
11 11.03 11.01 11.69 12.95 20.91 Galois-12 Falcon-12
(b) DMR speedup
Galois single
r a n d 1 r a n d 2 r m a t 1 r m a t 2 r
d 1 r
d 2 10 20
4.02 5.04 5.93 6.58 8.32 10.72 1.34 1.75 1.45 1.63 10.7 14.16 18.1 16.1 11.49 11.26 0.61 0.56 Falcon-GPU LonestarGPU Falcon-CPU
(c) Dynamic-SSSP- Self relative speedup
Morph Algorithm Results-Speedup DMR(Delaunay Mesh Refinement) and Dynamic-SSSP
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 9 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
rand64 rmat50 rmat60 5 10
8.24 9.04 10.18 7.53 8.45 7.96 8.65 7.13 7.92 9.6 8.7 10.45
Speedup Falcon-sssp Totem-sssp Falcon-bfs Totem-bfs speedup on two GPUs rand128 rmat80 5 10
6.03 5.8 6.45 5.52 4.15 4.83 4.27 4.96
Speedup Speedup on two GPUs and one CPU Heterogeneous Execution-SSSP and BFS speedup
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 10 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
Partitioned Execution of one Algorithm on Multiple Devices
1
fun1(Point ori, Point incom){
2
if(orig.dist >incom.dist)
3
4
}
5
relaxgraph(Point p, HGraph hgraph){
6
foreach(t in p.outnbrs)
7
MIN(t.dist,p.dist+hgraph.getWeight(p,t), hgraph.changed[0]);
8
}
9
main(int argc, char *argv[]) {
10
HGraph hgraph;
11
hgraph.addPointProperty(dist, int);
12
hgraph.addProperty(changed, int);
13
hgraph.read(argv[1]);
14
hgraph.makePartition(1,1,ORDERED);
15
hgraph.updateFunction(fun1);
16
foreach(t In hgraph.points) t.dist=1234567890;
17
hgraph.points[0].dist=0;
18
while( 1 ){
19
hgraph.changed[0]=0;
20
foreach(t In hgraph.points)relaxgraph(t,hgraph);
21
if(hgraph.changed[0]==0)break;
22
hgraph.updatePartition();
23
}
24
for(int i = 0;i <hgraph.npoints; i++)
25
printf(“%d”, hgraph.points[i].dist);
26
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 Part1(PRG, CDG) 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100 , 75 , INF 0 , 100 , 75 , 500 0 , 100 , 75 , 500 Part2(FRA,DEL) 0 , INF, INF , INF 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100, 75 , 675 0 , 100 , 75 , 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 11 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
Partitioned Execution of one Algorithm on Multiple Devices
1
fun1(Point ori, Point incom){
2
if(orig.dist >incom.dist)
3
4
}
5
relaxgraph(Point p, HGraph hgraph){
6
foreach(t in p.outnbrs)
7
MIN(t.dist,p.dist+hgraph.getWeight(p,t), hgraph.changed[0]);
8
}
9
main(int argc, char *argv[]) {
10
HGraph hgraph;
11
hgraph.addPointProperty(dist, int);
12
hgraph.addProperty(changed, int);
13
hgraph.read(argv[1]);
14
hgraph.makePartition(1,1,ORDERED);
15
hgraph.updateFunction(fun1);
16
foreach(t In hgraph.points) t.dist=1234567890;
17
hgraph.points[0].dist=0;
18
while( 1 ){
19
hgraph.changed[0]=0;
20
foreach(t In hgraph.points)relaxgraph(t,hgraph);
21
if(hgraph.changed[0]==0)break;
22
hgraph.updatePartition();
23
}
24
for(int i = 0;i <hgraph.npoints; i++)
25
printf(“%d”, hgraph.points[i].dist);
26
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 Part1(PRG, CDG) 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100 , 75 , INF 0 , 100 , 75 , 500 0 , 100 , 75 , 500 Part2(FRA,DEL) 0 , INF, INF , INF 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100, 75 , 675 0 , 100 , 75 , 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 11 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
Partitioned Execution of one Algorithm on Multiple Devices
1
fun1(Point ori, Point incom){
2
if(orig.dist >incom.dist)
3
4
}
5
relaxgraph(Point p, HGraph hgraph){
6
foreach(t in p.outnbrs)
7
MIN(t.dist,p.dist+hgraph.getWeight(p,t), hgraph.changed[0]);
8
}
9
main(int argc, char *argv[]) {
10
HGraph hgraph;
11
hgraph.addPointProperty(dist, int);
12
hgraph.addProperty(changed, int);
13
hgraph.read(argv[1]);
14
hgraph.makePartition(1,1,ORDERED);
15
hgraph.updateFunction(fun1);
16
foreach(t In hgraph.points) t.dist=1234567890;
17
hgraph.points[0].dist=0;
18
while( 1 ){
19
hgraph.changed[0]=0;
20
foreach(t In hgraph.points)relaxgraph(t,hgraph);
21
if(hgraph.changed[0]==0)break;
22
hgraph.updatePartition();
23
}
24
for(int i = 0;i <hgraph.npoints; i++)
25
printf(“%d”, hgraph.points[i].dist);
26
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 Part1(PRG, CDG) 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100 , 75 , INF 0 , 100 , 75 , 500 0 , 100 , 75 , 500 Part2(FRA,DEL) 0 , INF, INF , INF 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100, 75 , 675 0 , 100 , 75 , 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 11 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
Partitioned Execution of one Algorithm on Multiple Devices
1
fun1(Point ori, Point incom){
2
if(orig.dist >incom.dist)
3
4
}
5
relaxgraph(Point p, HGraph hgraph){
6
foreach(t in p.outnbrs)
7
MIN(t.dist,p.dist+hgraph.getWeight(p,t), hgraph.changed[0]);
8
}
9
main(int argc, char *argv[]) {
10
HGraph hgraph;
11
hgraph.addPointProperty(dist, int);
12
hgraph.addProperty(changed, int);
13
hgraph.read(argv[1]);
14
hgraph.makePartition(1,1,ORDERED);
15
hgraph.updateFunction(fun1);
16
foreach(t In hgraph.points) t.dist=1234567890;
17
hgraph.points[0].dist=0;
18
while( 1 ){
19
hgraph.changed[0]=0;
20
foreach(t In hgraph.points)relaxgraph(t,hgraph);
21
if(hgraph.changed[0]==0)break;
22
hgraph.updatePartition();
23
}
24
for(int i = 0;i <hgraph.npoints; i++)
25
printf(“%d”, hgraph.points[i].dist);
26
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 Part1(PRG, CDG) 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100 , 75 , INF 0 , 100 , 75 , 500 0 , 100 , 75 , 500 Part2(FRA,DEL) 0 , INF, INF , INF 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100, 75 , 675 0 , 100 , 75 , 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 11 / 12
Introduction Experimental Results Conclusion backup slides GPUs and MultiGPUs CPU Dynamic(Morph) Algorithms Heterogeneous execution
Partitioned Execution of one Algorithm on Multiple Devices
1
fun1(Point ori, Point incom){
2
if(orig.dist >incom.dist)
3
4
}
5
relaxgraph(Point p, HGraph hgraph){
6
foreach(t in p.outnbrs)
7
MIN(t.dist,p.dist+hgraph.getWeight(p,t), hgraph.changed[0]);
8
}
9
main(int argc, char *argv[]) {
10
HGraph hgraph;
11
hgraph.addPointProperty(dist, int);
12
hgraph.addProperty(changed, int);
13
hgraph.read(argv[1]);
14
hgraph.makePartition(1,1,ORDERED);
15
hgraph.updateFunction(fun1);
16
foreach(t In hgraph.points) t.dist=1234567890;
17
hgraph.points[0].dist=0;
18
while( 1 ){
19
hgraph.changed[0]=0;
20
foreach(t In hgraph.points)relaxgraph(t,hgraph);
21
if(hgraph.changed[0]==0)break;
22
hgraph.updatePartition();
23
}
24
for(int i = 0;i <hgraph.npoints; i++)
25
printf(“%d”, hgraph.points[i].dist);
26
} A SIMPLE EXAMPLE PRG CDG DEL FRA 100 75 400 6 Part1(PRG, CDG) 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100 , 75 , INF 0 , 100 , 75 , 500 0 , 100 , 75 , 500 Part2(FRA,DEL) 0 , INF, INF , INF 0 , INF , INF , INF 0 , 100 , 75 , INF 0 , 100, 75 , 675 0 , 100 , 75 , 500 Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 11 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
2 Programmer does not have to worry on target architecture ,
thread & memory management.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
2 Programmer does not have to worry on target architecture ,
thread & memory management.
3 Future Works in mind
i) to extend it for CPU clusters. ii) Making DSL more simple(say removing <GPU>tag). iii) Support for non-Nvidia GPUs by providing OpenCL backend. iv) adding optimizations on Compiler.
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
2 Programmer does not have to worry on target architecture ,
thread & memory management.
3 Future Works in mind
i) to extend it for CPU clusters. ii) Making DSL more simple(say removing <GPU>tag). iii) Support for non-Nvidia GPUs by providing OpenCL backend. iv) adding optimizations on Compiler.
4 for queries email me on unni c@csa.iisc.ernet.in. Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
2 Programmer does not have to worry on target architecture ,
thread & memory management.
3 Future Works in mind
i) to extend it for CPU clusters. ii) Making DSL more simple(say removing <GPU>tag). iii) Support for non-Nvidia GPUs by providing OpenCL backend. iv) adding optimizations on Compiler.
4 for queries email me on unni c@csa.iisc.ernet.in. 5 Some sample autogenereated code can be downloaded from
my home page(http://clweb.csa.iisc.ernet.in/unni c).
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides 1 We have a introduced a new DSL for Graph algorithms which
targets heterogeneous architectures.
2 Programmer does not have to worry on target architecture ,
thread & memory management.
3 Future Works in mind
i) to extend it for CPU clusters. ii) Making DSL more simple(say removing <GPU>tag). iii) Support for non-Nvidia GPUs by providing OpenCL backend. iv) adding optimizations on Compiler.
4 for queries email me on unni c@csa.iisc.ernet.in. 5 Some sample autogenereated code can be downloaded from
my home page(http://clweb.csa.iisc.ernet.in/unni c).
6 The Falcon compiler will made available online. Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 12 / 12
Introduction Experimental Results Conclusion backup slides
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 13 / 12
Introduction Experimental Results Conclusion backup slides
Unnikrishnan C, Rupesh Nasre, Y N Srikant Falcon: A Graph Manipulation Language for Heterogeneous Systems 14 / 12
Introduction Experimental Results Conclusion backup slides
1
relaxgraph(Point <GPU>p, Graph <GPU>graph) {
2
foreach (t In p.outnbrs)
3
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
4
}
1
#define t (((struct hgraph *)(graph.extra)))
2
global void relaxgraph(GGraph graph, int x) {
3
id = blockIdx.x * blockDim.x + threadIdx.x + x;
4
if (id <graph.npoints){
5
int falcft0 = graph.index[id];
6
int falcft1 = graph.index[id+1]-graph.index[id];
8
for (falcft2 = 0; falcft2 <falcft1; falcft2++) {
9
int ut0 = 2 * (falcft0 + falcft2); //edge index
10
int ut1 = graph.edges[ut0].ipe; //dest point
11
int ut2 = graph.edges[ut0 + 1].ipe;
12
GMIN(&t->dist[ut1], t->dist[id] + ut2, changed);
13
}
14
}
Code generated for GPU
1
relaxgraph(Point p, Graph graph) {
2
foreach (t In p.outnbrs)
3
MIN(t.dist, p.dist + graph.getweight(p, t), changed);
4
}
1
#define t (((struct hgraph *)(graph.extra)))
2
relaxgraph(int &p ,HGraph &graph) {
3
int falcft0 = graph.index[p];
4
int falcft1 = graph.index[p+1]-graph.index[p];
7
for (int falcft2 = 0; falcft2 <falcft1; falcft2++) {
8
int ut0 = 2 * (falcft0 + falcft2);
9
int ut1 = graph.edges[ut0].ipe;
10
int ut2 = graph.edges[ut0 + 1].ipe;
11
HMIN(&t->dist[ut1], t->dist[p] + ut2, ut1, changed);
12
}
14
}
Code generated for CPU Code generated for relaxgraph function of SSSP example
Introduction Experimental Results Conclusion backup slides
1
int <GPU>changed=0,lev=0;
2
BFS(Point <GPU>p,Graph <GPU>graph,int lev) {
3
foreach( t In p.outnbrs ){
4
if(t.dist>(lev+1)) {
5
t.dist=lev+1;changed=1;
6
}
7
}
8
}
9
main(int argc, char *name[]) {
10
Graph hgraph;
11
hgraph.addPointProperty(dist,int);
12
hgraph.read(name[1]);
13
hgraph.getType() <GPU>graph;
14
graph=hgraph;
15
foreach(t In graph.points)t.dist=1234567890;
16
graph.points[0].dist=0;
17
while( 1 ){
18
changed=0;
19
foreach(t In graph.points)(t.dist==lev) BFS(t,graph);
20
if(changed==0)break;
21
lev++;
22
}
23
for(int i=0;i<graph.npoints;i++)
24
printf(“%d\n”,graph.points[i].dist);
25
}
Falcon code for BFS on GPU
1
int changed=0,lev=0;
2
BFS(Point p,Graph graph,int lev) {
3
foreach( t In p.outnbrs ){
4
if(t.dist>(lev+1)) {
5
t.dist=lev+1;changed=1;
6
}
7
}
8
}
9
main(int argc, char *name[]) {
10
Graph hgraph;
11
hgraph.addPointProperty(dist,int);
12
hgraph.read(name[1]);
13
foreach(t In hgraph.points) t.dist=1234567890;
14
hgraph.points[0].dist=0;
15
while( 1 ){
16
changed=0;
17
foreach(t In hgraph.points) (t.dist==lev)BFS(t,hgraph,lev);
18
if(changed==0)break;
19
lev++;
20
}
21
for(int i=0;i<hgraph.npoints;i++)
22
printf(“%d\n”,hgraph.points[i].dist);
23
}
Falcon code for BFS on CPU
Introduction Experimental Results Conclusion backup slides
Dynamic Graph algorithms
1 Falcon provides addPoint(),addEdge() function on Graph
class.
Introduction Experimental Results Conclusion backup slides
Dynamic Graph algorithms
1 Falcon provides addPoint(),addEdge() function on Graph
class.
2 Falcon compiler checks for such function during code
generation.
Introduction Experimental Results Conclusion backup slides
Dynamic Graph algorithms
1 Falcon provides addPoint(),addEdge() function on Graph
class.
2 Falcon compiler checks for such function during code
generation.
3 If such a function is found in DSL, Graph is pre-allocated with
more space(say 3 times).
Introduction Experimental Results Conclusion backup slides
Dynamic Graph algorithms
1 Falcon provides addPoint(),addEdge() function on Graph
class.
2 Falcon compiler checks for such function during code
generation.
3 If such a function is found in DSL, Graph is pre-allocated with
more space(say 3 times).
4 Deletion of Points/Edges are managed by marking. No
Garbage Collection.
Introduction Experimental Results Conclusion backup slides
Graph member function addProperty() function
1 graph.addProperty(struct node,triangle);
Introduction Experimental Results Conclusion backup slides
Graph member function addProperty() function
1 graph.addProperty(struct node,triangle); 2 After this you can use triangle in same way as Edge and Point
in DSL code.
Introduction Experimental Results Conclusion backup slides
Graph member function addProperty() function
1 graph.addProperty(struct node,triangle); 2 After this you can use triangle in same way as Edge and Point
in DSL code.
3 It provides to programmer
i)an iterator triangle ii) variable ntriangle, which store number of triangles.
Introduction Experimental Results Conclusion backup slides
Graph member function addProperty() function
1 graph.addProperty(struct node,triangle); 2 After this you can use triangle in same way as Edge and Point
in DSL code.
3 It provides to programmer
i)an iterator triangle ii) variable ntriangle, which store number of triangles.
4 This function can be used to view Graph as Collection of
triangles(DMR) or as Collection of clauses(Survey Propagation).
Introduction Experimental Results Conclusion backup slides
Code for Parallel Execution of SSSP and BFS on two GPUs in Falcon
1
int <GPU>changed;
2
SSSPBFS(char *name) { //begin SSSPBFS
3
Graph hgraph;//Graph object on CPU
4
hgraph.addPointProperty(dist,int);
5
hgraph.addProperty(changed,int);
6
hgraph.getType() <GPU>graph0;//Graph on GPU0
7
hgraph.getType() <GPU>graph1;//Graph on GPU1
8
hgraph.addPointProperty(dist1,int);
9
hgraph.read(name);//read Graph from file to CPU
10
graph0=hgraph;//copy entire Graph to GPU0
11
graph1=hgraph;//copy entire Graph to GPU1
12
foreach(t In graph0.points)t.dist=1234567890;
13
foreach(t In graph1.points)t.dist=1234567890;
14
graph0.points[0].dist=0;
16
parallel sections { //do in parallel
17
section {//compute BFS on GPU1
18
while(1){
19
graph1.changed[0]=0;
20
foreach(t In graph1.points)BFS(t,graph1);
21
if(graph1.changed[0]==0) break;
22
}
23
}
24
section {//compute SSSP on GPU0
25
while(1){
26
graph0.changed[0]=0;
27
foreach(t In graph0.points)SSSP(t,graph0);
28
if(graph0.changed[0]==0) break;
29
}
30
}
31
}//end SSSPBFS
32
}