A HighPerformance Hybrid Computing Approach to Massive Contingency Analysis in the Power Grid.
ABSTRACT Operating the electrical power grid to prevent power blackouts is a complex task. An important aspect of this is contingency analysis, which involves understanding and mitigating potential failures in power grid elements such as transmission lines. When taking into account the potential for multiple simultaneous failures (known as the Nx contingency problem), contingency analysis becomes a massively computational task. In this paper we describe a novel hybrid computational approach to contingency analysis. This approach exploits the unique graph processing performance of the Cray XMT in conjunction with a conventional massively parallel compute cluster to identify likely simultaneous failures that could cause widespread cascading power failures that have massive economic and social impact on society. The approach has the potential to provide the first practical and scalable solution to the Nx contingency problem. When deployed in power grid operations, it will increase the grid operator's ability to deal effectively with outages and failures with power grid components while preserving stable and safe operation of the grid. The paper describes the architecture of our solution and presents preliminary performance results that validate the efficacy of our approach.

Conference Paper: EmPower: An Efficient Load Balancing Approach for Massive Dynamic Contingency Analysis in Power Systems
[Show abstract] [Hide abstract]
ABSTRACT: Power system simulations involving solution of thousands of stiff differential and algebraic equations (DAE) are computationally intensive and yet crucial for grid security and reliability. Online simulations of a large number of contingencies require very high computational efficiency. Furthermore, since the simulation times across the contingencies vary considerably, dynamic load balancing of parallel contingency analysis (CA) is required to ensure maximum resource utilization. However, the stateoftheart contingency analysis techniques fail to fulfill this requirement. In this paper, we present EmPower, an Efficient load balancing approach for massive dynamic contingency analysis in Power systems. For single contingency analysis, EmPower uses time domain simulations and incorporates efficient numerical algorithms for solving the DAE. Further, the contingency analysis approach is scaled for large scale contingency analysis using MPI based parallelization. For enabling an efficient, nonblocking implementation of workstealing, multithreading is employed within each processor. Simulations of thousands of contingencies on a supercomputer have been performed and the results show the effectiveness of EmPower in providing good scalability and huge computational savings.High Performance Computing, Networking, Storage and Analysis (SCC), 2012 SC Companion:; 01/2012  [Show abstract] [Hide abstract]
ABSTRACT: Access to a computational tractable method for representing the system more realistically has always been an important issue in power system reliability assessment. The Post Optimal Analysis (POA), as a well recognized technique to attack a set of similar optimization problems, has been successfully used to assess the reliability of composite systems. This method exploits the similarity of the system states to speed up the contingency evaluation procedure without sacrificing the accuracy of the results. In this paper, the performance and practical feasibility of the POA technique for power system reliability evaluation is tested using the Iran power grid. The POA based approach is applicable in both sorts of state sampling methods namely analytical enumeration (AE) and Monte Carlo simulation (MCS). Considering the dimension of the Iran power grid, in this paper system states are sampled using the MCS approach. In the simulations, a diverse range of reliability indices at both overall system and individual load point levels are computed with/without POA accommodation. The accuracy and execution time associated with the POA based method are compared with those obtained using the conventional method. Although considerable efforts have been devoted to collect and compile equipments performance data in the Iran grid data collection system, the uncertainty in the availability data of components is inevitable. Accordingly, a sensitivity analysis is carried out to investigate the effects of possible errors in equipments forced outage rates (FORs). The impacts on system risk of different peak load levels are also investigated.01/2012; 
Conference Paper: Modified centrality measures of power grid to identify critical components: Method, impact, and rank similarity
[Show abstract] [Hide abstract]
ABSTRACT: This paper justifies the criticality of nodes in a power system found from previous analysis. Critical nodes are identified from complex network theory based modified centrality approaches whose failure would stress the system badly in terms of line overloading. Impact of removal of critical nodes on two different standard test systems are analyzed. Simulation results suggest that these indices do not change with continually varying power system load and generation.Power and Energy Society General Meeting, 2012 IEEE; 01/2012
Page 1
1
A HighPerformance Hybrid Computing Approach to Massive Contingency
Analysis in the Power Grid
Ian Gorton, Zhenyu Huang, Yousu Chen, Benson Kalahar, Shuangshuang Jin, Daniel ChavarríaMiranda,
Doug Baxter, John Feo
Pacific Northwest National Laboratory,
Richland WA 99352, USA
*ian.gorton@pnl.gov
Abstract
Operating the electrical power grid to prevent
power blackouts is a complex task. An important
aspect of this is contingency analysis, which involves
understanding and mitigating potential failures in
power grid elements such as transmission lines. When
taking into account the potential for multiple
simultaneous failures (known as the Nx contingency
problem), contingency analysis becomes a massively
computational task. In this paper we describe a novel
hybrid computational approach to contingency
analysis. This approach exploits the unique graph
processing performance of the Cray XMT in
conjunction with a conventional massively parallel
compute cluster to identify likely simultaneous failures
that could cause widespread cascading power failures
that have massive economic and social impact on
society. The approach has the potential to provide the
first practical and scalable solution to the Nx
contingency problem. When deployed in power grid
operations, it will increase the grid operator’s ability
to deal effectively with outages and failures with power
grid components while preserving stable and safe
operation of the grid. The paper describes the
architecture of our solution and presents preliminary
performance results that validate the efficacy of our
approach.
Introduction
Electric power transmission involves the delivery of
electricity from power generators to consumers. A
power transmission network connects power plants to
multiple substations in a populated area, and the wiring
from substations to customers provides the final step of
electricity distribution. Electric power transmission
allows widely geographically dispersed energy
generators (such as hydroelectric, coal and nuclear
power plants) to be connected to consumers.
A power transmission network is called a grid.
Multiple redundant lines between nodes in the network
are provided so that power can be routed using a
variety of paths from any power plant to any load
center. The exact route chosen can be based on the
economics of the transmission path and the cost of
power.
If a problem occurs on the electric power grid, such
as a transmission line going out of service due to
contact with vegetation, there are various issues that
may arise. For example, if large amounts of electric
power are being transferred from one geographic area
to another, the loss the connecting line will have
impacts on loads and voltages across the grid. Loads
may be lost and voltages may drop or even collapse in
various areas as a result, leading to power outages.
Power grid operators in North America refer to such
unexpected outages as “contingencies” and manage the
system in a way that ensures any single contingency
will not propagate into a cascading blackout. This is
known as the N1 contingency standard issued by the
North American Electric Reliability Corporation
(NERC).
Past power grid blackouts like the Northeast
Blackout in 2003[1] have involved 1020 simultaneous
contingencies happening across the grid. Therefore, if
multiple contingencies occur simultaneously, the
urgency to restore the grid to a normal condition is far
greater. This clearly indicates the need for Nx
contingency analysis, i.e. analysis of the potential
simultaneous occurrence of multiple contingencies. N
x contingency analysis can prepare grid operators with
mitigation procedures so as to avoid cascading failures.
Nx contingency analysis is also, as we explain later,
an extremely complex computational task.
In this paper we describe our computational
approach for Nx contingency analysis. The approach
exploits the multithreaded Cray XMT architecture to
select likely contingencies from the power grid
network. It transfers the identified contingencies to a
conventional compute cluster to perform contingency
analysis on selected cases based AC power flow. This
analysis generates potentially terabytes of data which
must be transferred back to the XMT to identify likely
power grid vulnerabilities and utilize advanced visual
analytical methods to present the results to operators
for remedial actions.
Page 2
2
We have previously demonstrated the potential of
the Cray MTA2 (the XMT’s predecessor) for solving
the static state estimation problem for the electric
power grid [2]. Based on this experience, this project
exploits the unique heterogeneous architecture of the
Cray XMT to tackle the enormous computational
challenge of advanced contingency analysis for the
electric power grid [3,4,5].
A novel aspect of our approach is the hybrid
architecture we employ. We draw on the strengths of
the Cray XMT to perform the necessary graph analyses
for contingency selection [7]. The subsequent
contingency analysis involves the execution of
thousands of independent tasks, and is hence most
appropriately executed on a conventional cluster
architecture. We connect the two computational
platforms using our MeDICi integration framework
[6], which provides highperformance transport of
large amounts of data.
The paper briefly explains the complexity of Nx
contingency analysis, and describes our hybrid
architecture and the major aspects of the algorithms
used on each platform. It then presents some initial
endtoend performance figures which demonstrate the
viability of this approach.
Background: Contingency Planning for the
Power Grid
In North America there are three interconnected
power grids – the eastern interconnection, western
interconnection, and the Texas grid. Although each
grid is a single interconnected machine, it is divided
into different administrative entities called Balancing
Areas (or BAs) which own, operate, and/or manage
their own pieces of the grid. For example, the western
interconnection has 35 BAs, as shown in Figure 1
Figure 1 Major Balancing Areas in the Western
power grid
When performing contingency analysis, each BA
looks no further than its own boundaries. For areas
within the interconnection where several BAs reside
next to each other, the potential problem caused by
simultaneous failures across multiple BAs can be
missed. Individually, model results from each BA may
show that each contingency does not cause a problem.
However, if these contingencies occur simultaneously,
there will likely be a very large systemwide impact.
Unfortunately, the urgency to restore the system is not
fully recognized with today’s N1 contingency
analysis.
In the western power grid, there are approximately
20,000 elements that could fail. Checking each
element takes about ½ cpu second, therefore, the entire
N1 contingency case set would take about 104 cpu
seconds. In order to check all the combinations of x
contingencies (x = 2,3,4,…), it would take 104x cpu
seconds. Given that there are 35 different BAs in the
western grid, one could consider there are often several
simultaneous contingencies occurring – but in different
BAs. Assuming x=10, checking all these N10
contingency cases rigorously
computational time approximately in the order of 1040
cpu seconds and is obviously infeasible.
Therefore, an important element in a practical
solution is how to identify the Nx contingencies from
a systemwide perspective. This will allow high
performance computing to check a limited set of
contingencies and identify detrimental consequences.
For each contingency that exhibits these problematic
results, automated algorithms must be created that will
identify the remedial operator actions necessary to
maintain grid stability should that contingency actually
occur. The automation of this algorithm is necessary
since this process will run continuously in an
operational setting. The effectiveness of the evaluation
will be limited by the efficiency and efficacy of the
algorithms and the processing capacity of the
computers used. As we explain in the next section, our
solution uses a hybrid solution, exploiting the key
strengths of the Cray XMT multithreaded system and a
conventional supercomputing cluster.
Hybrid Computing
Analysis
Given the sheer number of contingency cases in the
problem space and the realtime requirements for
power grid operations, today’s industry algorithms and
tools are not able to handle comprehensive contingency
analysis as described in the previous section. A
comprehensive contingency analysis process includes
the following three elements:
1) contingency selection,
2) parallel contingency analysis and
would require
for Contingency
Page 3
3
3) postprocessing of contingency analysis
For contingency selection, we use a new method
based on the concept of graph edge betweenness
centrality. The power grid can be treated as a weighted
undirected graph, and the edge betweenness centrality
identifies the most "traveled" edges in the graph. The
most traveled edges are considered the most important
branches in a power grid, and their failures must be
analyzed, while the least traveled edges are identified
as lowimpact branches, and their failures do not need
to be analyzed because they are of little importance to
power grid stability.
Detailed contingency analysis is performed for the
selected contingency cases based on the ranking results
from the contingency selection step. Each contingency
case is a nonlinear algebraic equation, called AC
power flow. The NewtonRaphson method is used to
solve the equation iteratively. As each case is
independent, it can be run separately on multiple nodes
in a cluster. Contingency analysis identifies violations,
which are defined as electric quantities (bus voltages
and power flows) that exceed their prespecified limits.
Violations are an indication of power grid vulnerability
and should be identified and presented to power grid
operators so the situation can be understood and
remedial actions can be taken.
Postprocessing is necessary to reduce the cognitive
load on power grid operators. Today’s industry tools
simply presents the violation data in a tabular form
without little postprocessing. When the power grid is
under stress, the number of violations can be very
large, and operators have no way to understand the
tabular presentation of so many violations in a real
time manner. In order to effectively interpret
contingency analysis results and support situational
awareness, postprocessing of violation data is
necessary. Violation data can be interpreted as impact
of contingencies on various parts of the power grid.
The more violations, the more severe the impact is.
The impact can be visualized on a geographic map as
colored contours indicating vulnerability levels with
color intensity throughout the grid. Figure 2 shows an
example graph of four contingency cases.
Our implementation of this 3 step process involves a
Cray XMT and a conventional
supercomputer connected via a 1 GBit Ethernet
network. To simplify the connectivity complexity, we
utilize a MeDICi [6] pipeline.
The Cray XMT system is a scalable multithreaded
high performance computing platform. The XMT
supports a global shared memory accessible by all
processors on the system and can scale to a total of
8,192 processors. With a hardware multithreading
mechanism to tolerate memory access latencies
clusterbased
frequently encountered in irregular, graphprocessing
algorithms, the XMT is a suitable platform for parallel
processing of largescale power grid graphs.
Figure 2 Power grid vulnerability assessment
through visual analytics
Our implementation of this 3 step process involves a
Cray XMT and a conventional
supercomputer connected via a 1 GBit Ethernet
network. To simplify the connectivity complexity, we
utilize a MeDICi [6] pipeline.
The Cray XMT system is a scalable multithreaded
high performance computing platform. The XMT
supports a global shared memory accessible by all
processors on the system and can scale to a total of
8,192 processors. With a hardware multithreading
mechanism to tolerate memory access latencies
frequently encountered in irregular, graphprocessing
algorithms, the XMT is a suitable platform for parallel
processing of largescale power grid graphs.
Our current contingency analysis platform is a Linux
cluster with 192 nodes with 8 cores each. The nodes
are connected using a highperformance Quad Data
Rate (QDR) InfiniBand network.
The MeDICi Integration Framework (MIF) is a
middleware platform for integrating heterogeneous
software codes (known as components) into a
processing pipeline. MIF separates the integration logic
from the processing logic used in components and the
transports used for connectivity. Components are
oblivious to the transports used to connect them
together, and the topology that they are connected in.
Transports can then be configured independently of the
component logic, allowing a declarative configuration
of a range of transports to meet the performance and
quality of service requirements for an application.
Graph Analysis for Contingency Selection
Vertex betweenness is a centrality measure of a
vertex within a graph [8, 9]. Edge betweenness is a
centrality measure of an edge within a graph. Edges
clusterbased
Page 4
that occur on many shortest paths betw
pair have higher betweenness than thos
Ulrik Brandes’ algorithm [10] is fre
calculate vertex betweenness for unwe
graphs. It requires O(n+m) space and
unweighted graphs, and O(nm + n2
weighted graphs, where n and m are
vertices and edges in the graph, respec
to apply edge betweenness centrality
graphs, we adapted Brandes’ algorith
edge betweenness for weighted power
power grids it is possible that ther
shortest paths between two buses, ther
algorithm [11] was also adapted to
shortest paths. The pseudo code for
algorithm is given in [7].
We have implemented this algorith
XMT using the parallelization capa
XMT’s multithreading C/C++ compil
betweenness centrality computation,
Brandes’ algorithm is called for eac
power grid graph. In each iteration,
paths are identified between this given
other vertices in the graph, and then
accumulates the calculated edge betw
for particular edges. After all the vertic
edge betweenness of all edges is
iteration is independent and can be exe
since it analyzes independentlysourced
We have used XMTspecific pragma
parallel) to guide the compiler
other sections of the code. We also
update pragma (#pragma mta up
updates to the shared edge betweenne
the compiler knows that this statem
performed atomically.
We have also implemented a po
version of the same code for comparis
mainstream CPUs. We have used a 12
Itanium 2 1.5 GHz processors, 2
memory) HP Superdome system t
performance against the Cray XMT.
Figure 3 compares the perform
implementations of power grid betwee
on both the Cray XMT and the HP Sup
problems sizes: 46,000 buses (com
Eastern US power grid) and
(comparable to a more detailed view of
grid). The data presented on both figu
time normalized to the time on one p
HP Superdome (time on one HP Super
is 1.0). Both figures use a logarit
improve the readability of the graph.
from both figures the perprocessor per
XMT is lower than the Superdome, but
ween any vertex
e that do not.
equently used to
eighted/weighted
O(nm) time for
2logn) time for
e the number of
ctively. In order
y to power grid
hm to calculate
r grid graphs. In
re are multiple
refore Dijkstra’s
o store multiple
r this modified
hm on the Cray
abilities of the
ler. In the edge
, the modified
h vertex in the
all the shortest
n vertex and any
n the algorithm
weenness values
ces are scanned,
available. Each
ecuted in parallel
d shortest paths.
(#pragma mta
in parallelizing
use the atomic
pdate) before
ess variables so
ment should be
ortable OpenMP
son purposes on
28core (64 dual
56 GB shared
to compare its
mance of our
enness centrality
perdome for two
mparable to the
170,000 buses
f the Eastern US
ures is execution
processor on the
rdome processor
thmic yaxis to
As can be seen
rformance of the
t its scalability is
better leading to overall reduced
scale.
Figure 3 Comparison of XMT
performance for Contingency
different problem sizes
Parallel Contingency Analy
Our framework for parallel con
shown in Figure 4. Each contingenc
a power flow run. To manage th
allocation and load balancing,
designated as the master process (
in addition to running contingency
Figure 4 Framework for pa
analysis
0.01
0.10
1.00
10.00
124816
332
Relative Execution Time (170,000 bu
Proc 0
Proc 0
Proc 1
Proc 1
Proc 2
Proc 2
Proc 0:
(1) Distribute base ca
(2) Perform load bala
(3) Distribute case in
(4) Perform continge
Other Proc’s:
(1) Update Y matrix based on ca
(2) Perform contingency analysis
4
d execution time at
T and Superdome
Selection for 2
ysis
ntingency analysis is
cy case is essentially
he contingency case
one processor is
(Proc 0 in Figure 4),
cases.
arallel contingency
64
us system)
XMT time
Superdome Time
Proc N
Proc N
…
ase Y0 matrix
ancing (static/dynamic)
nformation to other processors
ency analysis
se information: Y = Y0 + ΔY
s
Page 5
5
A straightforward mechanism for load balancing is
to statically preallocate an equal number of cases to
each processor, i.e. static load balancing. With static
load balancing, the master processor only needs to
allocate the cases once at the beginning of the analysis.
Since the convergence performance is different for
different cases, each contingency analysis run will
require a different number of iterations and thus take a
different time to finish. The extreme case would be
nonconverged cases which iterate until the maximum
number of iterations is reached. These variations in
execution time would result in unevenness, with the
overall computational efficiency determined by the
longest execution time of individual processors.
Therefore, computational power is not fully utilized as
many processors are idle while waiting for the last one
to finish.
Therefore a dynamic load balancing scheme was
used to allocate tasks to processors based on the
availability of a processor. Contingency cases are
dynamically allocated to free individual processors to
evenly distribute load and minimize processor idle
time. The scheme is based on a shared task counter
updated by atomic fetchandadd operations. The
master processor (Proc 0) does not distribute all the
cases at the beginning. Instead, it maintains a task
counter. Whenever a processor finishes its assigned
case, the processor requests more tasks from the master
processor and the task counter is updated. This process
is illustrated in Figure 5. With this approach, the
number of cases executed by each processor will not be
equal, but the computation time on each processor is
optimally utilized.
Figure
computational load balancing scheme
5 Taskcounterbased dynamic
This approach has been compared with a static
allocation scheme for 512 “N1” contingency cases on
a 32 node cluster. The results are shown in Figure 6.
The performance of dynamic load balancing, in
comparison with its static counterpart, shows much
greater linear scalability. The dynamic scheme
achieves eight times more speedup with 32 processors,
and the difference is expected to be greater with as the
problem is executed on a much large cluster.
Figure 6 Performance comparison of static and
dynamic computation load balancing schemes with
512 WECC contingency cases
As further validation, Figure 7 compares the
processor execution time for the case with 32
processors. With dynamic load balancing, the
execution time for all the processors is within a small
variation of the average 23.4 seconds, while static load
balancing has variations as large as 20 seconds or 86%.
The dynamic load balancing scheme successfully
improves speedups. It is also worth pointing out that
the contingency analysis process of 512 WECC cases
with full NewtonRaphson power flow solutions can be
finished within about 25 seconds. It is a significant
improvement compared to several minutes in current
industry practice. More case studies with up to 300,000
contingency cases can be found in [12]. All the results
show the advantages of dynamic load balancing
scheme.
Figure 7 Evenness of execution time with different
computational load balancing schemes
0
5
10
15
20
25
30
0510
Number of processors
15202530 35
Speedup
Dynamic load balancing
Static load balancing
0
5
10
15
20
25
30
35
40
45
1357911 13 15 17 19 21 23 25 27 29 31
Processor #
Time (sec)
Static Load Balancing
Dynamic Load Balancing
23.4 sec
Proc 0
Proc 0
Proc 1
Proc 1
Proc 2
Proc 2
Proc N
Proc 3
Computation
Time tc
I/O
Time tio
Counter Update
Time tcnt
Waiting
Time tw
Page 6
6
Integration with MeDICi
The MeDICi Integration Framework (MIF) is used
to facilitate seamless, flexible communication between
the XMT and the traditional compute cluster. The
framework abstracts away the complexity of high
performance networking code and allows the
programmer to concentrate on higher level integration
logic. General MIF applications take the form of a
processing pipeline, comprising software components
connected asynchronously or synchronously. Data
flows between components which process the data in
some manner and pass the results to the next
component(s) in the pipeline. Individual connections
between components can be declaratively configured
to specify a range of protocols supported by MIF,
allowing for simple tuning to match quality of service
requirements without making
integration code [6].
In our contingency analysis architecture, a MIF
pipeline has been constructed to:
1) Accept the selected contingencies from the
XMT and transfer the data to the compute
cluster
2) Invoke the contingency simulation code and
gather the results
3) Transfer the results back to the XMT and make
available for the postprocessing phase
For step (1), the data sizes are the order of 10
150KB, and hence we configure MIF to use the TCP
transport to facilitate a low overhead connection with
guaranteed delivery of data. A MIF component on an
XMT Linux node accepts a list of selected
contingencies from the analysis code on the XMT and
sends it via TCP to the compute cluster. A MeDICi
component running on the cluster communicates with
the job scheduler to launch the contingency simulation.
The next component in the pipeline then monitors for
output and forwards completed contingency data to the
XMT.
When running simulations for large contingency
lists, LZO compression is used to decrease both disk
and network overhead. LZO is a compression
algorithm and associated library that achieves up to
75% compression on our data. It is exceedingly fast,
so it does not introduce unacceptable levels of
computational overhead.
As the compressed data is currently in the 100’s of
MB to GB range depending on the number of
contingency cases, the MIF pipeline uses the high
performance bbcp protocol,
multithreaded chunked transfer, with no encryption
overhead. It provides a multithreaded chunking
changes to the
which provides
transfer, with no encryption overhead to transfer the
data files. It is invoked by MeDICi’s external
command module, which makes it entirely transparent
to the rest of the system, and simple to replace with a
faster protocol. In our tests, the MIF pipeline driving
this protocol delivers on average 30MB/sec over the
shared laboratory gigabit network between the XMT
and the compute cluster.
Performance Analysis
In order to validate
performance, we have integrated all of the software
codes that are critical to contingency analysis to create
an endtoend contingency test platform. The test case
includes all elements except the postprocessing and
visualization, which is not performance critical.
Our current hardware testbed is:
•
64 node Cray XMT with 512Mb of shared
memory
•
8 dual quadcore Intel Xeon “Clovertown”
nodes running at 2.33 GHz (64 cores total), with
nodes connected by a highperformance Quad
Data Rate (QDR) InfiniBand network.
•
1 gigabit shared interconnect
Figure 8 shows the total computation time in seconds
of test cases ranging from 200 to 17000 selected
contingencies. Contingency analysis, including the
transfer of data to/from the cluster using MeDICi,
dominates these times, ranging from approximately
85% of the total time for 200 contingencies, to 98% of
the total time for 17000 contingencies. However, our
tests only utilized 64 cores, and hence the contingency
analysis time for the larger contingency cases will be
greatly reduced on larger clusters, as we have
demonstrated in [12].
Data sizes transferred to the cluster after contingency
selection are in the 10’s of Kbytes ranges, and hence
their transfer time is negligible. Compressed data sizes
generated by the contingency analysis grow non
linearly and range from
contingencies to 2.5GBytes for 17000 contingencies.
The transfer time becomes a significant fraction of the
clusterbased contingency analysis time for larger
number of cases. For example, for 200 cases, data
transfer is less than 2 seconds, or approximately 3% of
the contingency analysis. This rises to 5% for 500
cases, 7% for 1000 cases, and eventually 17% for
17000 cases. Again, the performance of the data
transfer would be significantly improved on a
dedicated high speed network between the two
compute platforms. When fully optimized, bbcp is
capable of transferring data at close to line speeds,
our hybrid system’s
44Mbytes for 200
Page 7
7
which is 4 times faster than the performance we are
experiencing on the shared laboratory network used in
these tests. This would significantly reduce the
overheads of data transfer as data sizes grow.
Contingency selection time on the XMT is almost
constant in all these test cases (12.83 seconds for 200
cases and 13.16 seconds for 17000). Other variables in
performance are introduced by sorting and transferring
larger data sizes as the number of cases is increased.
Figure 8 Total computation time for N1
Contingency Analysis cases in seconds
Overall these initial results show that our hybrid
architecture has the potential to perform contingency
analysis in a time frame that is appropriate for power
grid operations. These results provide a strong
validation of the algorithms we are using and the
hybrid computational approach.
Conclusions and Further Work
Providing a highperformance solution to the Nx
contingency analysis problem is a key challenge for the
future of power grid management. In this project, we
are investigating how a hybrid solution can be built
and scale to realistic problem sizes.
This paper describes our approach and describes
performance results that validate the components of the
hybrid architecture. We also present initial endtoend
performance results which clearly show the potential
for our algorithms to scale and evolve into an
operational scenario on nextgeneration hybrid
architectures.
We are highly encouraged by these results and are
currently working to scale up tests to a 128 node XMT
and PNNL’s 161 TFLOP supercomputer so that we can
execute significantly larger Nx contingency analyses.
We are already seeing the need for using significantly
more cores for contingency analysis for N2 cases, and
increases in compressed data output sizes of up to five
times. Hence, handling these larger computations will
provide even deeper insights into the novel hybrid
architecture we are using for this application.
8. References
1. U.S.Canada Power System Outage Task Force,
Final Report on the August 14, 2003 Blackout in
the United State and Canada: Causes and
Recommendations,
https://reports.energy.gov/.
2. J. Nieplocha, D. ChavarríaMiranda, V. Tipparaju,
Zhenyu Huang, A. Marquez. Parallel WLS State
Estimator on Shared Memory Computers, in:
Proceedings of IPEC2007 – the 8th International
Power Engineering Conference, Singapore, 36
December 2007.
3. J. Deuse, K. Karoui, A. Bihain, J. Dubois,
Comprehensive approach of power system
contingency analysis, 2003 IEEE Power Tech
Conference Proceedings, Bologna, Volume 3, 23
26 June, 2003.
4. Q. Morante, N. Ranaldo, A. Vaccaro, E. Zimeo,
Pervasive Grid for LargeScale Power Systems
Contingency Analysis, IEEE Transactions on
Industrial Informatics, vol. 2, no. 3, August 2006
5. R.H. Chen, G. Jingde, O.P. Malik, W. ShiYing,
N. Xiang, Automatic contingency analysis and
classification, The
Conference on Power System Control and
Management, 1618 April, 1996.
6. I. Gorton, A. Wynne, J. Almquist, J. Chatterton:
The MeDICi Integration Framework: A Platform
for High Performance
Applications, In Procs Working IFIP/IEEE
Conference on Software Architecture, Feb 2008:
95104, IEEE
7. Y. Chen, S Jin, D ChavarríaMiranda, Z Huang,
Application of Cray XMT for Power Grid
Contingency Selection, Proceedings of Cray User
Group 2009, Atlanta, GA, May 47, 2009.
8. L.C. Freeman, A set of measures of centrality
based on betweenness, Sociometry, 1977, 40:35
41.
9. J.M. Anthonisse, The rush in a directed graph,
Technical Report
Mathematisch Centrum, Amsterdam, 1971.
10. U. Brandes, A Faster Algorithm for Betweenness
Centrality, Journal of Mathematical Sociology,
Vol. 25, 2001, pp. 163177.
11. E. W. Dijkstra, A note on Two Problems
Inconnexion with
Mathematik, 1 (1959), S. 269–271.
12. Z. Huang, Y. Chen, J. Nieplocha, Massive
Contingency Analysis with High Performance
Computing, Proc. IEEE PES General Meeting,
Calgary, Canada, July 2009
April 2004. at
Fourth International
Data Streaming
BN 9/71, Stichting
Graphs, Numerische
0
100
200
300
400
500
600
20050010002000500017000
Total Time
View other sources
Hide other sources
 Available from pnl.gov
 Available from Zhenyu Huang · Jun 4, 2014