Appcademy presentation at Mobile App Accelerator Camp Athens event
cloud_futures_2.0_Papazachos
1. Producing realistic earthquake shake maps using cloud computing: A cloud
implementation of a stochastic simulation approach to compute expected ground
motions for strong earthquakes
Papazachos C.B., G. Spyrou and A.A. Skarlatoudis
Geophysical Laboratory, Aristotle University of Thessaloniki, Greece
Abstract
With the proposed technique we aim to exploit the Microsoft Azure cloud
infrastructure for performing scientific computations for the near-real time assessment of
seismic motions after the occurrence of a strong earthquake. The main problem that
motivates this effort is the fact that after significant earthquakes limited information is
almost immediately available regarding the experienced seismic motions and especially
the potential damage on the infrastructures and lifelines of the urban areas that have
been affected by the earthquake. Even worse, information for the areas of higher
damage is not only unavailable but often false, as in most cases information (e.g. phone
calls to the police or fire brigade) is available from areas of less impact. Therefore, the
implementation of an automated application is proposed, in order to produce realistic
maps of the expected (estimated) spatial distribution of appropriate ground motion
measures and possibly induced damage, in a very short time (near-real time) after the
occurrence of a large earthquake.
The proposed application uses the automatic information for an earthquake,
distributed by the Seismological Station of the Aristotle University
(http://geophysics.geo.auth.gr/ss/station_index_en.html), and produces synthetic ground
motion parameters and seismograms, in order to derive the necessary ground motion
measures. The simulations for producing the synthetic seismograms are performed with
the stochastic method (using both point and finite source models) for a dense grid of
receivers that covers the study area. Stochastic simulations of strong ground motion
(either point source or more advanced finite-fault simulations) have been used during the
last 15 years mainly at a research level for simulations of individual earthquakes. The
original method, as well as its latest modifications are described in detail in [1-4]. In the
proposed application we mainly rely on the use of the modifications made to EXSIM [5]
by the implementation by Boore (2009) [6].
The stochastic method has been selected mainly because of its proven accuracy in
seismic impact assessment and wide applicability despite its simplicity compared to
more elaborate and/or deterministic methods. In addition, the execution time of the
various implementations of the method for a single receiver is of the order of few
minutes, which ranks the method among the first ones in terms of computational cost
and accuracy of the expected results. Nevertheless, in cases where a significant number
of sequential simulations has to be computed (e.g. the dense grid of receivers used in
our application) the execution time can be significantly larger, inadequate for practical
2. applications. Therefore, the successful implementation of the proposed application in
Microsoft Azure cloud infrastructure that will be based on the large number of available
computing nodes, allows the realization of the involved sequential simulation jobs within
a few minutes. Moreover, the application demands can be efficiently scaled up when
necessary (e.g. more scenarios, additional simulation sites, etc.), taking advantage of
the elasticity properties of the cloud-computing infrastructure.
From the technical point of view the proposed method utilizes the Microsoft Azure
infrastructure as a calculation backend by implementing the map-reduce execution
model. The first step is to populate a table in Azure Table Storage with the parameters
for the grid of virtual receivers for which the stochastic simulations will be performed
(receiver’s table). The process that manages the parameters stored receiver’s table for
each one them can be executed manually, giving the user the ability to add, delete and
modify the contents of the table as well as automatically based on preselected
parameters such area coverage, epicenter location (inland or sea), etc.
The map reduce model is being implemented in Windows Azure with the following
three simple worker roles, similar to [7]: A partitioner role , a simulation role and a
reducer role . Inter-process communication within the Azure datacenter, as well as with
external systems, such the servers in Seismological Data Centers, is achieved by
utilizing the Azure Queues system. Whenever an earthquake is automatically located
within the Seismological Station servers (running Linux operating system) a message
with the specific event information is being sent through Azure Queues to the Windows
Azure data centers . This message is being handled by the partionioner role, which
performs a query to the Azure Table Storage and based on the results it sends a number
of messages via Azure Queue that determine the map reduce pattern to the simulation
role. Each simulation role performs the following procedure in order to produce results:
1) Creates the input file based on the station parameters and earthquake data.
2) Executes the legacy simulation process (implemented in FORTRAN) and stores
the output files to local storage
3) Uploads the output files to Azure Blob storage for persistence
4) De-queues the specific message from the Azure Queue, so it can continue
working on another node of the simulation
5) Finally when all the simulation nodes are completed, the reducer role gathers all
the necessary data from Azure Blob Storage and creates several output products, such
a color shake map file that represents a specific ground motion measure, e.g. PGA
(Peak Ground Acceleration, see example of figure 1).
The fact that earthquakes phenomena by nature appear in a random (spike-type)
manner makes them a perfect candidate for utilizing the elastic capabilities of a cloud
infrastructure. When an earthquake occurs based on the its parameters (such as
preliminary location, magnitude, depth etc ) the number of worker nodes can be
allocated dynamically in order to achieve the execution of the simulation in the minimum
amount of time. After the simulation is completed these computational resources may
3. be released in order to minimize the cost. By using the elasticity of the Windows Azure
the Seismological Data Center can achieve both the execution of the simulation in a
minimum time, as well as keeping the cost of these calculation at the optimal level.
Figure 1. Typical shake map for Peak
Ground Acceleration (PGA in
cm/sec2
), generated for a Santorini
volcano intra-caldera earthquake with
magnitude M=5.5
Based on the previous description the proposed working flow after the occurrence of
the automatic location of an earthquake comprises of the following steps:
a) Execution of stochastic simulations assuming a point-source model, for a grid of
several hundred (e.g. 400-1000) of virtual receivers (earthquake recording stations) in
order to produce preliminary strong ground motion spatial distribution maps, using the
automatic location of the earthquake.
b) Using existing typical focal mechanisms (estimated from current knowledge for the
seismotectonic properties of the area), stochastic simulations will be performed for the
same grid of receivers in order to produce more accurate results, assuming a finite-
source model. The strong ground motion spatial distribution maps will be updated with
the more accurate results from this modeling.
c) Critical information such as potential high-damage areas and maps with strong ground
motion spatial distribution will be disseminated to Aristotle University Seismological
Station scientists for scientific assessment.
d) The previous steps will be re-executed immediately after a manual solution for the
location of the earthquake or its focal mechanism are available. All maps will be
updated after each step’s re-execution, in order to provide more accurate spatial shake
maps.
Critical information such as potential high-damage areas and maps with strong ground
motion spatial distribution will be disseminated to Aristotle University Seismological
Station scientists for scientific assessment, as well as to local and state authorities in
order to proceed in all the necessary risk mitigation actions.
4. The benefits from the proposed application have multiple levels of impact. The use of
the stochastic method for near-real time purposes is novel computationally, as this will
be the first time that this method will be realized on a cloud infrastructure (to the best of
our knowledge). A typical application for a single earthquake involves simulations for 600
receivers, for which 30 typical rupture scenarios would be examined. The sequential
execution for this number of receivers corresponds to an execution time of 1000-3000
minutes (depending on the configuration), with minimal memory requirements (a few Mb)
per simulation. Such computations can be easily realized on personal computers,
however, results would need several hours (or even days), making their use for near-real
time assessment of seismic motions impossible. This limitation can be efficiently handled
by cloud-computing infrastructures, since:
1) The large number of available computing nodes allows the realization of the
involved sequential simulation jobs within a few minutes. For example, execution of the
previously described application (single event-30 scenarios-300 simulation sites) on e.g.
200 cloud computing CPU cores would require a few minutes, making the near real-time
use possible.
2) The application demands can be efficiently scaled up when necessary (e.g. more
scenarios, additional simulation sites, etc.), taking advantage of the cloud elasticity,
without the need to invest the large cost involved with permanent computing
infrastructures.
3) Even when the involved application is used for high seismicity areas (e.g. broader
Aegean area), relatively strong earthquakes occur rather sparsely (e.g. once every
month), corresponding to ~10-20 annual “production” runs. Therefore, the maintenance
of a large dedicated computational infrastructure for the specific application (even at a
national level) corresponds to a very high-cost investment that will be relatively sparsely
used and practically never depreciated. Hence, cloud computing is a very efficient cost-
wise approach, since computational resources are allocated and used only whenever
the corresponding application need occurs.
The method is currently being tested for a preselected area (namely the broader city
of Thessaloniki and the Santorini island, Greece) in order to check the reliability of the
technique and calibrating (using cloud computing) the procedure against pre-existing
advance modeling results [8-10] but also due to the availability of instrumental/modeling
data for this test area. The produced results can have a significant impact on scientists,
as well as the wider public. Scientists like seismologists, soil- and civil-engineers, etc.
that are either interested on computing or using information related to the estimation of
the direct impact of strong earthquakes (e.g. expected or observed seismic motions and
damage distribution after a major seismic event) or even more diverse disciplines
(sociologists, psychologists, etc.) that are interested in the indirect, earthquake-induced
effects such as mass behavior, etc could benefit from our results. Moreover, the
produced results could be used by agencies that participate in the seismic risk mitigation
and especially those that are responsible to organize post-earthquake relief measures,
such as general or specialized governmental agencies and non-governmental
organizations involved in natural disaster support actions. Finally, parts of the produced
5. results could be used for public awareness and/or to inform people that are either
interested to share experience about earthquakes and disseminate this information
through the web or simply want to have reliable information about the possibly expected
or even observed consequences of strong seismic events.
Acknowledgements
This work has been partly funded by the CLOUD-QUAKE pilot project of the VENUS-
C research initiative (http://www.venus-c.eu), funded by the European Union through the
7th Framework Programme (Grant Agreement number 261565). All cloud computing
was performed on the Windows Azure infrastructure (https://windows.azure.com) for
which Microsoft Corporation & Microsoft Research has provided free access (~70,000
CPU hours), as part of the VENUS-C project. The help of both Microsoft
Corporation/Microsoft Research and VENUS-C project is gratefully acknowledged.
References
[1] Boore, D. M. (1983). Stochastic simulation of high-frequency ground motions based
on seismological models of the radiated spectra, Bull. Seism. Soc. Am., 73, 1865-1894.
[2] Boore, D. M. (1996). SMSIM--Fortran programs for simulating ground motions from
earthquakes: version 1.0, U.S. Geol. Surv. Open-File Rept. 96-80-A and 96-80-B, 73 pp.
[3] Boore, D. M. (2003). Simulation of ground motion using the stochastic method, Pure
and Applied Geophysics 160, 635-675.
[4] Joyner, W.B. and D.M. Boore (1988). Measurement, characterization, and prediction
of strong ground motion, in Earthquake Engineering and Soil Dynamics II, Proc. Am.
Soc. Civil Eng. Geotech. Eng. Div. Specialty Conf., June 27-30, 1988, Park City,
Utah, 43-102.
[5] Motazedian, D., and G. Atkinson (2005). Stochastic finite-fault model based on
dynamic corner frequency, Bull. Seism. Soc. Am., 95, 995–1010.
[6] Boore, D. M. (2009). Comparing stochastic point-source and finite-source ground-
motion simulations: SMSIM and EXSIM, Bull. Seism. Soc. Am. 99, 3202-3216.
[7] Carrión, A., I. Blanquer, V. Hernández (2012), A Service-based BLAST command
tool supported by Cloud Infrastructures, Proceedings of the HealthGrid 2012,
Amsterdam, 21-23 May 2012, (accepted and publication pending).
[8] Skarlatoudis A.A., C.B. Papazachos, N. Theodoulidis, J. Kristek and P. Moczo,
(2010). Local site-effects for the city of Thessaloniki (N. Greece) using a 3D Finite-
Difference method: A case of complex dependence on source and model parameters,
Geoph. J. Int.,182, 279-298.
[9] Skarlatoudis A.A., C.B. Papazachos and N. Theodoulidis, (2011). Spatial distribution
of site-effects and wave propagation properties in Thessaloniki (N. Greece) using a 3D
Finite Difference method, Geoph. J. Int., 185, 485-513.
[10] Skarlatoudis A.A., C.B. Papazachos and N. Theodoulidis, (2012). Site response
study of the city of Thessaloniki (N. Greece), for the 04/07/1978 (M5.1) aftershock, using
a 3D Finite-Difference wave propagation method, Bull. Seism. Soc. Am., (accept. publ.).