An optimized cost-based data allocation model for heterogeneous distributed ...
Hassan - Condor _48_x_36
1. • Green Infrastructures (GIs) are being used more today to
advance urban infrastructure systems performance, reduce
associated costs, and diminish greenhouse gasses emission
to achieve urban sustainability.
• GIs provide control at the area of generation, so there
won’t be capital and performance cost and emissions in
conveyance and storage facilities (examples in Fig 1).
Fig 1: Rainwater Harvesting Cistern as an example of GIs (left) and CSO storage tunnel as an
example of traditional urban water facilities (right)
• However, choosing an intellectually robust plan among the
millions of possible LID implementation plans according to
the variation in type, size and location through long-term
continuous simulation is a computational challenge.
Accelerating Stormwater Infrastructure Design
With High-throughput Computing Using HTCondor
Hassan Tavakol1, Steven Burian1, Maryam Imani2, Andrew Duncan2, Raziyeh Farmani2, David Butler2
1Urban Water Group, University of Utah; 2Centre for Water Systems, University of Exeter
RESULTSINTRODUCTION METHODS
is the combined sewer network of the
East Side Interceptor System at the City of Toledo
(4th most populous city in Ohio, US). This system
extends along the east side of the Maumee
River and serves a total of 25,000 acres, of
which 2,200 acres are combined. As a part
of LTCP (Long Term Control Plan), a model was
developed in the US-EPA SWMM (Storm
Water Management Model) 5.0 (Fig 3). 5years
continuous rainfall (1997-2001) was
Selected as the representative period.
was performed as the 1st scenario
to accelerate the optimization. In
this scenario, all the 4 physical
cores (Core i7 – 2600 CPU@3.44GHz)
were used to evaluate the objective
function through the parfor
function in MATLAB 2013b.
high-throughput computing framework - 8.0.5
CONCLUSION AND FUTURE WORKS
• Application of the presented approach in the studied area
could lead to the same CSO control with 2.2 Million $ less!!!
• More objective functions, more GIs, bigger systems?!
REFERENCE
Deb K, Pratap A, Agarwal S, (2002) A fast and elitist multiobjective genetic
algorithm NSGA-II[J]. Evolutionary Computation. 6(2): 182-197
Graham Saunders, Dr. Defne Apul, Andy Stepnick, Jay Devkota
ACKNOWLEDGEMENTS
STUDY AREA
• This research is guided by the question: “How Multi-
Objective Optimization of RainWater Harvesting (RWH)
storage to minimize the Combined Sewer Overflow (CSO)
volume and implementation cost can be accelerated in a
complex system and long-term continuous data?”
• In such system, evaluating the objective functions takes
more than 99% of the time of optimization. Therefore,
parallelization of CSO modeling for different individuals in
each generation of evolutionary algorithm (illustrated in Fig
2) was tested as an answer of the above question.
Fig 2: Parallelizing CSO modeling for different individuals during the evolutionary algorithm
generations
RESEARCH QUESTION
Fig 3: SWMM
model of the
studied
combined
sewer
network
Parallelization for
different individuals
in same generation
Fig 4: CPU cores usage in regular un-
parallelized optimization process (left)
and in parallelized case (right)
Fig 5: Schematic of the
HTCondor network
edition - was used as the 2nd scenario. It
provided 68 potential nodes through 4
researchers’ computers and 5 undergrad’s
computer site at university of Exeter (Fig 5)
when their CPU was idle for 15 min. The jobs
that were suspended (due to the owner use) were
resubmitted to different machines to avoid
holding the generation progress.
(Non-dominated Sorting Genetic
Objective
Functions
- Total CSO volume
- RWH implem. cost
Decision
Variables
41 (RWH storage in
each subcatchment)
Max Gen 500
Pop Size 40
framework was developed in
this research to connect NSGA-
II to SWMM, and automate the
job submission process (Fig 6).
was also
used based on one 200Lit rain
barrel per building for
comparison purposes.
for generation#i
GA operations %selection,…
for individual#j
create SWMM_input(i,j)
submit condor_job(i,j)
end
wait until all jobs are done
for individual#j
update cost objfun(i,j)
read SWMM_output(i,j)
update CSO objfun(i,j)
end
end
Table1: NSGA-II details
Fig 6: Developed framework
Algorithm) was used via NGPM-1.4
(NsGa-II Program in Matlab) developed
by Mathwork (details in Table1).
0
200
400
600
HTCondor CPU
Parallelization
Time(hr)
• HTCondor could accelerate the
optimization remarkably (Fig 7),
especially during weekend.
• Studying the convergence
process (Fig 8) showed that Pareto
front after 175th generation has
had minimal improvement.
0
1
2
3
4
5
6
7
8
9
10
1200 1250 1300 1350 1400
ImplementationCost(Million$)
CSO Volume (Million Galon)
Initial Generation
Generation#50
Generation#100
Generation#150
Generation#175
Generation#200
Fig 7: Comparing the elapsed time in
the two accelerating methods
Fig 8: Convergence process, Pareto front in
the last generation, and the selected
individual for comparison purposes (the one
with the grey filling colour)
Selecting the best individual among the Pareto front is not a
goal of this research. Stakeholders can choose it base on the
financial constraints, water quality requirements, and the
weights that they consider
for the two objectives. The
one with closest CSO to
Simple Design is selected
just for comparison (Fig 9).
0
1
2
3
4
0
500
1000
1500
2000
Million$
MillionGallon
CSO Volume
Cost
Fig 9: Comparing design approaches