Page 1! of !

10

CSE 603 Project Report

Parallel BFS on Graphs using GPGPU
Name: Soumasish Goswami
Dec 18, 2015
Abstract
Graph algorithms are a fundamental paradigm of computer Science and is relevant to
many domains and application areas. Large graphs involving millions of vertices are
common in scientific and engineering applications. Reasonable time bound
implementations of graph algorithms using high-end computing resources have been
reported but are accessible to only a few. Graphics Processing Units (GPUs) are fast
emerging as inexpensive parallel processors due to their high computation power and
low price. The GeForce line of Nvidia GPUs provides the CUDA programming model
that treats the GPU as a SIMD processor array. Ive presented a fundamental graph
algorithm - the breadth first search, using this programming model on large graphs. The
results on a graph of 500, 000 vertices would suggest that the NVIDIA GPUs can be
used as a reasonable co-processor to accelerate parts of an application.

Soumasish Goswami

Parallel BFS on CUDA 7.5

Page !2 of !10

Introduction
Graph representations are common in many problems of scientific and engineering
applications. Some of these problems map to very large graphs, often involving millions
of vertices, viz, problems likeVLSI chip layout, phylogeny reconstruction, data mining,
and network analysis can require graphs with millions of vertices. Running serial
implementations of Breadth-First-Search traversal algorithms on such large datasets
can be a time-consuming task. This report summarizes observations on parallelizing this
task using CUDA. Owing to the limitations of hardware and environment as obtained in
Centre for Computational Research (CCR), the project has been tested in its entirety on
a personal computer(with above average specs). However the observations can be easily
generalized to a larger data set. The actual tests on the serial algorithm run significantly
longer than the timings plotted in the report, primarily because the sub routine of
traversing a list of edges and adding it to the graph is painfully slow. The relevant tables
in the observations section only summarizes the time taken to run the actual BreadthFirst-Search(hereby referred to as BFS) algorithm.
GPGPU
General purpose programming on Graphics processing Units (GPGPU) tries to solve a
problem by posing it as a graphics rendering problem, restricting the range of solutions
that can be ported to the GPU. A GPGPU solution is designed to follow the general flow
of the graphics pipeline (consisting of vertex, geometry and pixel processors), with each
iteration of the solution being one rendering pass. Since the GPU memory layout is also
optimized for graphics rendering the GPGPU solutions as an optimal data structure may
not be available. Creating efficient data structures using the GPU memory model is a
challenge in itself . Memory size on GPU is another restricting factor. A single data
structure on the GPU cannot be larger than the maximum texture size supported by it.
The CUDA Programming Model
For the programmer the CUDA model is a collection of threads running in parallel. A
warp is a collection of threads that can run simultaneously on a multiprocessor. The
warp size is fixed for a specific GPU. The programmer decides the number of threads to
be executed. If the number of threads is more than the warp size, they are time-shared
internally on the multiprocessor. A collection of threads (called a block) runs on a
multiprocessor at a given time. Multiple blocks can be assigned to a single
multiprocessor and their execution is time-shared. A single execution on a device
generates a number of blocks. A collection of all blocks in a single execution is called a
grid. All threads of all blocks executing on a single multiprocessor divide its resources
equally amongst themselves. Each thread and block is given a unique ID that can be
accessed within the thread during its execution. Each thread executes a single
instruction set called the kernel.
The kernel is the core code to be executed on each thread. Using the thread and block

Soumasish Goswami

Parallel BFS on CUDA 7.5

Page !3 of !10

IDs each thread can perform the kernel task on different set of data. Since the device
memory is available to all the threads, it can access any memory location. The CUDA
programming interface provides a Parallel Random Access Machine (PRAM)
architecture, if one uses the device memory alone. However, the multiprocessors follow
a SIMD model, the performance improves with the use of shared memory which can be
accessed faster than the device memory. The hardware architecture allows multiple
instruction sets to be executed on different multiprocessors. The current CUDA
programming model, however, cannot assign different kernels to different
multiprocessors.
With CUDA, the GPU ca be viewed as a massive parallel SIMD processor, limited only
by the amount of memory available on the graphics hardware. The GT 750M graphics
card has 2048 MB memory. Large graphs can reside in this memory, given a suitable
representation.
Device Specs
GeForce GT 750M
CUDA Driver Version/ Runtime Version

7.5/7.5

CUDA Capability Major/Minor version

3.0

Total amount of global memory

2048 MBytes

2 Multiprocessors, 192 CUDA cores/MP

384 CUDA Cores

GPU Max Clock rate

926 Mhz

Memory Clock Rate

2508 Mhz

Memory Bus Width

128- bit

L2 Cache Size

262144 bytes

Maximum Texture Dimension Size(x,y,z)

1D=(65536), 2D = (65536, 65536), 3D= (4096,

4096, 4096)

Maximum layered 1D Texture size

1D=(16384), 2048 layers

Maximum layered 2D Texture size

2D= (16384, 16384), 2048 layers

Total amount of constant memory

65536 bytes

Warp Size

32

Maximum number of threads per multiprocessor

2048
1024

Maximum number of threads per block

Maximum dimension of a thread block(x, y, z)

(1024, 1024, 64)

Maximum dimension of a grid size(x, y, z)

(2147483647, 65535, 65535)

Soumasish Goswami

Parallel BFS on CUDA 7.5

Page 4
! of !10

Thus the device limitations being set is that a total of 786,432 threads could be run at
the same time owing to the device capability of 384 cores of 2048 threads each.
Data Set
The data set taken in as an argument by both the implementations is a vector/array of
edge lists. To this extent this is a rough parallel to he Erdos Renyi model of graph
generation. A brief reference can be found here:
The random seed used to generate the graph in this case is, we use a C++11 seed
generator:
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<> dis(0, (NUM_VERTICES -1));
Key assumptions
For the sake of simplicity we assume each vertex to be an integer. An edge is thus a pair
of integers representing from and to. An edge list is an array of such pairs. All edges are
bi-directional and have equal weights.
Core Serial Algorithm
The serial implementation follows a simple C++ STL Queue/Vector based graph. The
Graph class maintains a vector of vectors for each vertex listing the connected edges for
the respective vertex. The BFS algorithm uses a queue to traverse across each frontier /
level and touches all connected components from a start node.
Parallelization of BFS
Since BFS lends itself well to parallelization there are two common yet distinct strategies
that are followed in the parallel execution of BFS:
1.
2.

The level-synchronous algorithm.

The fixed-point algorithm

The level synchronous algorithm uses the following approach; it manages three sets of
nodes - the visited set V , the current-level set C, and the next-level set N. Iteratively, the
algorithm visits (in parallel) all the nodes in set C and transfers them to set V (in
parallel). C is then populated with the nodes from set N, and N is cleared for the new
iteration. This iterative process continues until naturally there is no node in the next
level. The level synchronous algorithm effectively visits in parallel all nodes in each BFS
level, with the parallel execution synchronizing at the end of each level iteration.

Soumasish Goswami

Parallel BFS on CUDA 7.5

Page 5
! of 10
!

The fixed-point algorithm,on the other hand, continuously updates the BFS level of
every node, based on BFS levels of all neighboring nodes until no more updates are
made. This method is sub-optimal at times because of the lack of communication
between neighboring nodes in parallel environments. For the purpose of this project Ive
based my implementation on this approach.
Observations & Speed Up
Starting from the base data set of 1024 vertices and an equal number on nodes the data
set has been scaled up progressively in multiples of 2. The details of the scale up in the
vertex to edge ratio and the time taken to run BFS on them are enumerated in the tables
below.
Base testing data set -1024 vertices
Vertices

Edges

Serial Runtime

Parallel
Runtime

Speed Up

1024

1024

0.000012

0.001332

-11000%

1024

2048

0.000011

0.000944

-8481.8%

1024

4096

0.000554

0.000877

-58.3%

1024

8192

0.001467

0.001084

26.1%

1024

16,384

0.005263

0.001106

99.9%

1024

32,768

0.018046

0.001338

92.5%

1024

65,536

0.064879

0.001344

97.9%

Serial

Parallel

70000

52500

35000

17500

0
1X

Soumasish Goswami

2x

4x

8x

16x

32x

64x

Parallel BFS on CUDA 7.5

Page 6
! of !10

Data on a more sparse graph

Vertices

Edges

Serial Runtime

Parallel
Runtime

Speed Up

65536

32,768

0.000034

0.000974

-2764.7%

65536

65,536

0.000031

0.003547

-11341.9%

Serial

Parallel

4000

3000

2000

1000

0
1x

Soumasish Goswami

1/2 x

Parallel BFS on CUDA 7.5

Page 7
! of !10

Scale up - 8X
Vertices

Edges

Serial Runtime

Parallel
Runtime

Speed Up

8192

8192

0.000026

0.000235

-803.8%

8192

16384

0.001461

0.001979

-35.4%

8192

32762

0.004600

0.006388

-38.8%

8192

65536

0.013723

0.003123

77.2%

8192

131072

0.042567

0.002640

93.7%

8192

262096

0.155348

0.002776

98.2%

8192

524288

0.567887

0.001600

99.7%

Serial

Parallel

600000

450000

300000

150000

0
1x

Soumasish Goswami

2x

4x

8x

16x

32x

64x

Parallel BFS on CUDA 7.5

Page 8
! of 10
!

Scale up - 16X
Vertices

Edges

Serial Runtime

Parallel
Runtime

Speed Up

65536

65536

0.000181

0.014693

-8017.6%

65536

131072

0.015386

0.014237

7.4%

65536

262096

0.051042

0.016766

67.1%

65536

524288

0.134016

0.023420

82.5%

Serial

Parallel

140000

105000

70000

35000

0
1x

Soumasish Goswami

2x

4x

8x

Parallel BFS on CUDA 7.5

Page 9
! of !10

Analysis
The following trends emanate from the plotting of the two algorithms run on the same
data set.
1. The serial algorithm tends to work better on a extremely sparse graphs with a 1:1
vertex to edge ratio.
2. The parallel algorithm works better on moderately sparse graphs and shows a
significant runtime improvement as the vertex to edge ratio goes up.
3. As the graph gets denser below 1x to 1/2x or 1/4x, the parallel algorithm continues to
perform poorer and poorer, primarily owing to the multiple serial calls being made
to launch the kernel. This is the serial bottleneck of this algorithm.
4. As the data set grows there seems to be a gradual fall in the performance difference
of the two algorithms on graphs with a 1:1/1:2 vertex edge ratio. Thus from 1024 to
65536 the speed up values are -8481% and 7%(1:2 vertex to edge). It can be safely
assumed that as the data set grows the relative runtime of the serial algorithm on the
a sparse graph will continue to decrease. Thus the parallel algorithm will
proportionately keep performing better over larger and larger data sets.
5. The catch up is also quicker as the data set grows. For the base data set of 1024
nodes the edge density had to increase to 8x over vertex density for a yield of 27%.
However in case of a graph with 65536 vertices the edge increase of only 4x yielded
67% speed up. Clearly the more the data the better the speedup - in both relative and
absolute terms.
6. Aside from the timings noted here each algorithm takes a non-trivial amount of time
doing initializations. The serial algorithm has to read an array of edge list and
compose the graph before it can start BFS. That in effect doubles the execution time
of the program. In case of the parallel algorithm the frontier array has to be
initialized to -1 for all levels except the start vertex. For a large data set this might
have to be done in batches which could add non-trivial overhead on the execution of
the program.
7. There was some difference observed in running the same data on CUDA multiple
times. Though not significant for a data set less than 500, 000 nodes this is
something which needs to be observed on a bigger data set of billion nodes. In all
likelihood the difference in runtime is attributable to other processes using the GPU
on the machine.
8. One key observation while running the tests was that the level of connectivity of a
node played a key role in determining the runtime. For the same data set, different
start points had significantly different running times based on its level of
connectivity. Thus it was of paramount importance to use the same start vertex for
both the algorithms.
Limitations of the Algorithm
Its not a surprise that this algorithm works well with a multi-connected graph where
one vertex is connected to many edges but limited to a certain number of levels. This is
because for every level theres a fresh kernel call that is made and the kernel though
Soumasish Goswami

Parallel BFS on CUDA 7.5

Page !10 of !10

launched with threads equal to the number of vertices only writes on those indices if the
array which are connected at that level. This is wasteful in nature and continuous
copying of data from host to device is a bottleneck in the algorithm. Also in case of a
graph which is linearly connected to all vertices, a linked list, the serial and the parallel
execution time would be the same.
However the beauty of the algorithm lies in the fact that it easily lends itself to complete
parallelization. On the current data set the algorithm performs as desirably as one could
expect.
Thus the algorithm is ideally recommended for a moderately sparse graphs. Also the
algorithm scales proportionately with data.
Recommendations
The algorithms speedup begins to improve as the number of edges goes up relative to
the number of vertices, this is a minor concern because most large-scale real-world
networks are sparse where the number of edges is much smaller than the maximum
number of possible edges. In that sense this implementation will find more usage in
streaming data where the graph is more connected.
Code Repositories
Data Set Generation: https://github.com/soumasish/GenerateDataSet
Serial BFS: https://github.com/soumasish/SerialBFS
Parallel BFS: https://github.com/soumasish/ParallelBreadthFirstSearch
Conclusion
The size of the device memory limits the size of the graphs handled on a single GPU. The
CUDA programming model provides an interface to use multiple GPUs in parallel using
multi-GPU bridges. Up to 2 synchronized GPUs can be combined using the SLI
interface. NVIDIA QuadroPlex is a CUDA enabled graphics solution with two Quadro
5600 cards each. Two such systems can be supported by a single CPU to give even better
performance than the GT 750M . NVIDIA has announced its Tesla S870 range of GPUs,
with up to four cores and 6GB system memory capacity, targeted at high performance
computing. Further research is required on partitioning the problem and streaming the
data from the CPU to GPU to handle even larger problems. External memory
approaches can also be adapted to the GPUs for this purpose.
References
1.
2.
3.
4.
5.

http://impact.crhc.illinois.edu/shared/papers/effective2010.pdf
https://en.wikipedia.org/wiki/Erd%C5%91s%E2%80%93R%C3%A9nyi_model
https://docs.nvidia.com/cuda/cuda-c-programming-guide/
https://www.sci.utah.edu/publications/Fu2014a/UUSCI-2014-002.pdf
https://research.nvidia.com/publication/scalable-gpu-graph-traversal

Soumasish Goswami

Parallel BFS on CUDA 7.5