The flexible design of the RegESM modeling system allows choosing a different atmospheric model component (ATM) in the configuration of the coupled model for a various type of application. Currently, two different atmospheric models are compatible with the RegESM modeling system: (1) RegCM4 , which is developed by the Abdus Salam International Centre for Theoretical Physics (ICTP), and (2) the Advanced Research WRF model ARW;, which is developed and sourced from National Center for Atmospheric Research (NCAR). In this study, RegCM 4.6 is selected as an atmospheric model component because the current implementation of WRF coupling interface is still experimental and does not support coupling with the co-processing component yet, but the next version of the modeling system (RegESM 1.2) will be able to couple the WRF atmospheric model with the co-processing component. The NUOPC cap of atmospheric model components defines state variables (i.e., sea surface temperature, surface wind components), rotates the winds relative to Earth, applies unit conversions and performs vertical interpolation to interact with the newly introduced co-processing component.

The current version of the coupled modeling system supports two different ocean model components (OCNs): (1) Regional Ocean Modeling System ROMS revision 809;, which is developed and distributed by Rutgers University, and (2) MIT general circulation model MITgcm version c63s;. In this case, the ROMS and MITgcm models are selected due to their large user communities and different vertical grid representations. Although the selection of ocean model components depends on user experience and application, often the choice of vertical grid system has a determining role in some specific applications. For example, the ROMS ocean model uses a terrain-following (namely s coordinates) vertical grid system that allows a better representation of the coastal processes, but MITgcm uses z levels generally used for applications that involve open oceans and seas. Similar to the atmospheric model component, both ocean models are slightly modified to allow data exchange with the other model components. In the current version of the coupled modeling system, there is no interaction between wave and ocean model components, which could be crucial for some applications (i.e., surface ocean circulation and wave interaction) that need to consider the two-way interaction between waves and ocean currents. The exchange fields defined in the coupled modeling system between ocean and atmosphere strictly depend on the application and the studied problem. In some studies, the ocean model requires heat, freshwater and momentum fluxes to be provided by the atmospheric component, while in others, the ocean component retrieves surface atmospheric conditions (i.e., surface temperature, humidity, surface pressure, wind components, precipitation) to calculate fluxes internally, by using bulk formulas . In the current design of the coupled modeling system, the driver allows selecting the desired exchange fields from the predefined list of the available fields. The exchange field list is a simple database with all fields that can be exported or imported by the component. In this way, the coupled modeling system can be adapted to different applications without any code customizations in both the driver and individual model components.

Surface waves play a crucial role in the dynamics of the PBL in the atmosphere and the currents in the ocean. Therefore, the wave component is included in the coupled modeling system to have a better representation of atmospheric PBL and surface conditions (i.e., surface roughness, friction velocity and wind speed). In this case, the wave component is based on WAM Cycle 4 (4.5.3-MPI). WAM is a third-generation model without any assumption on the spectral shape . It considers all the main processes that control the evolution of a wave field in deep water, namely the generation by wind, the non-linear wave–wave interactions and also white capping. The model was initially developed by the Helmholtz-Zentrum Geesthacht (GKSS, now HZG) in Germany. The original version of the WAM model was slightly modified to retrieve surface atmospheric conditions (i.e., wind speed components or friction velocity and wind direction) from the RegCM4 atmospheric model and to send back calculated surface roughness. In the current version of the modeling system, the wave component cannot be coupled with the WRF model due to the missing modifications on the WRF side. In RegCM4, the received surface roughness is used to calculate air–sea transfer coefficients and fluxes over sea using Zeng ocean air–sea parameterization . In this design, it is also possible to define a threshold for maximum roughness length (the default value is 0.02 m) and friction velocity (the default value is 0.02 m) in the configuration file of RegCM4 to ensure the stability of the overall modeling system. Initial investigation of the added value of atmosphere–wave coupling in the Mediterranean Sea can be found in .

To simulate the lateral freshwater fluxes (river discharge) at the land surface and to provide river discharge to ocean model component, the RegESM modeling system uses the Hydrological Discharge (HD, version 1.0.2) model developed by Max Planck Institute . The model is designed to run in a fixed global regular grid with 0.5∘ horizontal resolution using daily time series of surface runoff and drainage as input fields. In that case, the model uses the pre-computed river channel network to simulate the horizontal transport of the runoff within model watersheds using different flow processes such as overland flow, baseflow and river flow. The river-routing model (RTM) plays an essential role in the freshwater budget of the ocean model by closing the water cycle between the atmosphere and ocean model components. The original version of the model was slightly modified to support interaction with the coupled model components. To close the water cycle between land and ocean, the model retrieves surface and subsurface runoff from the atmospheric component (RegCM or WRF) and provides estimated river discharge to the selected ocean model component (ROMS or MITgcm). In the current design of the driver, rivers can be represented in two different ways: (1) individual point sources that are vertically distributed to model layers and (2) imposed as freshwater surface boundary condition like precipitation (P) or evaporation minus precipitation (E-P). In this case, the driver configuration file is used to select the river representation type (1 or 2) for each river individually. The first option is preferred if river plumes need to be defined correctly by distributing river discharge vertically among the ocean model vertical layers. The second option is used to distribute river discharge to the ocean surface when there is a need to apply river discharge to a large areal extent close to the river mouth. In this case, a special algorithm implemented in the NUOPC cap of ocean model components (ROMS and MITgcm) is used to find affected ocean model grids based on the effective radius (in kilometers) defined in the configuration file of the driver.

RegESM (version 1.1) is completely redesigned and improved version of the previously used and validated coupled atmosphere–ocean model (RegCM-ROMS) to study the regional climate of the Caspian Sea and its catchment area . To simplify the design and to create a more generic, extensible and flexible modeling system that aims to support easy integration of multiple model components and applications, RegESM uses a driver to implement the hub-and-spoke approach. In this case, all the model components are combined using ESMF (version 7.1.0) to structure the coupled modeling system. ESMF is selected because of its unique online regridding capability, which allows the driver to perform different interpolation types (i.e., bilinear, conservative) over the exchange fields (i.e., sea surface temperature, heat and momentum fluxes) and the NUOPC layer. The NUOPC layer is a software layer built on top of ESMF. It refines the capabilities of ESMF by providing a more precise definition of a component model and how components should interact and share data in a coupled system. ESMF also provides the capability of transferring computational grids in the model component memory, which has critical importance in the integration of the modeling system with a co-processing environment (see also Sect. ). The RegESM modeling system also uses ESMF and the NUOPC layer to support various configurations of component interactions such as defining multiple coupling time steps among the model components. An example configuration of the four-component (ATM, OCN, RTM and WAV) coupled modeling system can be seen in Fig. . In this case, the RTM component runs in a daily time step (slow) and interacts with ATM and OCN components, but ATM and OCN components can interact each other more frequently (fast) such as every 3 h.

The interaction (also called as run sequences) among the model components and driver is facilitated by the connector components provided by the NUOPC layer. Connector components are mainly used to create a link between individual model components and the driver. In this case, the number of active components and their interaction determines the number of connector components created in the modeling system. The interaction between model components can be either (1) bidirectional, such as atmosphere and ocean coupled modeling systems, or (2) unidirectional, such as atmosphere and co-processing modeling systems. In the unidirectional case, the co-processing component does not interact with the atmosphere model and only processes retrieved information; thus, there is one connector component.

The RegESM modeling system can use two different types of time-integration coupling schemes between the atmosphere and ocean components: (1) explicit and (2) semi-implicit (or leapfrog) (Fig. ). In the explicit type coupling, two connector components (ATM-OCN and OCN-ATM directions) are executed concurrently at every coupling time step, and model components start and stop at the same model time (Fig. a). In the semi-implicit coupling type (Fig. b), the ocean model receives surface boundary conditions from the atmospheric model at one coupling time step ahead of the current ocean model time. The semi-implicit coupling aimed at lowering the overall computational cost of a simulation by increasing stability for longer coupling time steps.

As described earlier, the execution of the model components is controlled by the driver. Both sequential and concurrent execution of the model components is allowed in the current version of the modeling system. If the model components and the driver are configured to run in sequence on the same set of persistent execution threads (PETs), then the modeling system executes in a sequential mode. This mode is a much more efficient way to run the modeling system in case of limited computing resources. In the concurrent type of execution, the model components run in mutually exclusive sets of PETs, but the NUOPC connector component uses a union of available computational resources (or PETs) of interacted model components. This way, the modeling system can support a variety of computing systems ranging from local servers to large computing systems that could include high-speed performance networks, accelerators (i.e., graphics processing unit or GPU) and parallel I/O capabilities. The main drawback of concurrent execution approach is assigning the correct amount of computing resource to individual model components, which is not an easy task and might require an extensive performance benchmark of a specific configuration of the model components, to achieve best available computational performance. In this case, a load-balancing analysis of individual components and driver play a critical role in the performance of the overall modeling system. For example, the LUCIA (Load-balancing Utility and Coupling Implementation Appraisal) tool can be used to collect all required information such as waiting time and calculation time of each system component for a load-balancing analysis in the OASIS3-MCT-based coupled system.

In general, the design and development of the coupled modeling systems involve a set of technical difficulties that arise due to the usage of the different computational grids in the model components. One of the most common examples is the mismatch between the land–sea masks of the model components (i.e., atmosphere and ocean models). In this case, the unaligned land–sea masks might produce artificial or unrealistic surface heat and momentum fluxes around the coastlines, narrow bays, straits and seas. The simplest solution is to modify the land–sea masks of the individual model components manually to align them; however, this requires time and is complex (especially when the horizontal grid resolution is high). Besides, the procedure needs to be repeated each time the model domain (i.e., shift or change in the model domain) or horizontal grid resolution is changed.

The RegESM modeling system uses a customized interpolation technique that also includes extrapolation to overcome the mismatched land–sea mask problem for the interaction between the atmosphere, ocean and wave components. This approach helps to create more generic and automatized solutions for the remapping of the exchange fields among the model components and enhance the flexibility of the modeling system to adapt to different regional modeling applications. There are three main stages in the customized interpolation technique: (1) finding destination grid points that the land–sea mask type does not match completely with the source grid (unmapped grid points; Fig. ), (2) performing bilinear interpolation to transfer the exchange field from source to destination grid and (3) performing extrapolation in destination grid to fill unmapped grid points that are found in first step.

To find the unmapped grid points, the algorithm first interpolates the field from source to destination grid (just over the sea) using a nearest-neighbor-type interpolation (from Field_A to Field_B). Similarly, the same operation is repeated by using a bilinear-type interpolation (from Field_A to Field_C). Then, the results of both interpolations (Field_B and Field_C) are compared to identify unmapped grid points for the bilinear interpolation (Fig. ).

The field can then be interpolated from the source to the destination grid using a two-step interpolation approach. In the first step, the field is interpolated from source to destination grid using a bilinear interpolation. Then, nearest-neighbor-type interpolation is used on the destination grid to fill unmapped grid points. One of the main drawbacks of this method is that the result field might include unrealistic values and sharp gradients in the areas of complex land–sea mask structure (i.e., channels, straits). The artifacts around the coastlines can be fixed by applying a light smoothing after interpolation or using more sophisticated extrapolation techniques such as the sea-over-land approach , which are not included in the current version of the modeling system. Also, the usage of the mosaic grid along with the second-order conservative interpolation method, which gives smoother results when the ratios between horizontal grid resolutions of the source and destination grids are high, can overcome unaligned land–sea mask problem. The next major release of ESMF library (8.0) will include the creep fill strategy to fill unmapped grid points.

The model simulations and performance benchmarks are done on a cluster (SARIYER) provided by the National Center for High-Performance Computing (UHeM) in Istanbul, Turkey. The CentOS 7.2 operating system installed in compute nodes is configured with a two Intel Xeon CPU E5-2680 v4 (2.40 GHz) processors (total 28 cores) and 128 GB RAM. In addition to the compute nodes, the cluster is connected to a high-performance parallel disk system (Lustre) with 349 TB storage capacity. The performance network, which is based on Infiniband FDR (56 Gbps), is designed to give the highest performance for the communication among the servers and the disk system. Due to the lack of GPU accelerators in the entire system, the in situ visualization integrated performance benchmarks are done with the support of software rendering provided by the Mesa library. Mesa is an open-source OpenGL implementation that supports a wide range of graphics hardware each with its back end called a renderer. Mesa also provides several software-based renderers for use on systems without graphics hardware. In this case, ParaView is installed with Mesa support to render information without using hardware-based accelerators.

RegESM 1.1 is configured to couple atmosphere (ATM; RegCM) and ocean (OCN; ROMS) models with the newly introduced co-processing component (ParaView/Catalyst version 5.4.1) to analyze the evolution of Hurricane Katrina and to assess the overall performance of the modeling system. In this case, two atmospheric model domains were designed for RegCM simulations using an offline nesting approach, as shown in Fig. . The outer atmospheric model domain (low resolution, LR) with a resolution of 27 km is centered at 25.0∘ N, 77.5∘ W, and covers almost the entire United States, the western part of the Atlantic Ocean and northeastern part of the Pacific Ocean for better representation of the large-scale atmospheric circulation systems. The outer domain is enlarged as much as possible to minimize the effect of the lateral boundaries of the atmospheric model in the simulation results of the inner model domain. The horizontal grid spacing of the inner domain (high resolution, HR) is 3 km and covers the entire GoM and the western Atlantic Ocean to provide high-resolution atmospheric forcing for coupled atmosphere–ocean model simulations and perform cloud-resolving simulations. Unlike the outer domain, the model for the inner domain is configured to use the non-hydrostatic dynamical core (available in RegCM 4.6) to allow better representation of local-scale vertical acceleration and essential pressure features.

The lateral boundary condition for the outer domain is obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) latest global atmospheric reanalysis ERA-Interim project;, which is available at 6 h intervals at a resolution of 0.75∘×0.75∘ in the horizontal and 37 pressure levels in the vertical. On the other hand, the lateral boundary condition of the inner domain is specified by the results of the outer model domain. The Massachusetts Institute of Technology-Emanuel convective parameterization scheme MIT-EMAN; for the cumulus representation along with the subgrid explicit moisture SUBEX; scheme for large-scale precipitation are used for the low-resolution outer domain.

As it can be seen in Fig. , the ROMS ocean model is configured to cover the entire GoM to allow better tracking of Hurricane Katrina. In this case, the used ocean model configuration is very similar to the configuration used by the Physical Oceanography Numerical Group (PONG), Texas A&M University (TAMU), in which the original model configuration can be accessed from their THREDDS (Thematic Real-time Environmental Distributed Data Services) data server (TDS). THREDDS is a service that aims to provide access to an extensive collection of real-time and archived datasets, and TDS is a web server that provides metadata and data access for scientific datasets, using a variety of remote data access protocols. The ocean model has a spatial resolution of 1/36∘, which corresponds to a non-uniform resolution of around 3 km (655×489 grid points) with highest grid resolution in the northern part of the domain. The model has 60 vertical sigma layers (θs=10.0, θb=2.0) to provide a detailed representation of the main circulation patterns of the region and vertical tracer gradients. The bottom topography data of the GoM are constructed using the ETOPO1 dataset , and minimum depth (hc) is set to 400 m. The bathymetry is also modified so that the ratio of depth of any two adjacent columns does not exceed 0.25 to enhance the stability of the model and ensure hydrostatic consistency that prevents pressure gradient error. The Mellor–Yamada level 2.5 turbulent closure MY; is used for vertical mixing, while rotated tensors of the harmonic formulation are used for horizontal mixing. The lateral boundary conditions for the ROMS ocean model are provided by Naval Oceanographic Office Global Navy Coastal Ocean Model (NCOM) during 27–30 August 2005.

The model coupling time step between the atmosphere and ocean model components is set to 1 h but a 6 min coupling time step is used to provide one-way interaction with the co-processing component to study Hurricane Katrina in a very high temporal resolution. In the coupled model simulations, the ocean model provides sea surface temperature (SST) data to the atmospheric model in the region where their numerical grids overlap. In the rest of the domain, the atmospheric model uses SST data provided by the ERA-Interim dataset (prescribed SST). The results of the performance benchmark also include additional tests with a smaller coupling time step such as 3 min for the interaction with the co-processing component. In this case, the model simulations for the analysis of Hurricane Katrina run over 3 days, but only 1 day of simulation length is chosen in the performance benchmarks to reduce the compute time.

A set of simulations is performed with different model configurations to assess the overall performance of the coupled modeling system by focusing on the overhead of the newly introduced co-processing component (Table ). The performance benchmarks include analysis of the extra overhead provided by the co-processing component, coupling interval between physical models and the co-processing component under different rendering loads such as various visualization pipelines (Table ). In this case, same model domains that are described in the previous section (Sect. ) are also used in the benchmark simulations. The LR atmospheric model domain includes around 900 000 grid points, while the HR domain contains 25 million grid points to test scaling up to a large number of processors. In both cases, the ocean model configuration is the same, and it has around 19 million grid points. Besides the use of a non-hydrostatic dynamical core in the atmospheric model component in the HR case, the rest of the model configuration is preserved. To isolate the overhead of the driver from the overhead of the co-processing component, first individual model components (ATM and OCN) are run in stand-alone mode, and then the best-scaled model configurations regarding two-dimensional decomposition configuration are used in the coupled model simulations; CPL (two-component case: atmosphere–ocean) and COP (three-component case: atmosphere, ocean and the co-processing component). Due to the current limitation in the integration of the co-processing component, the coupled model only supports sequential-type execution when the co-processing component is activated, but this limitation will be removed in the future version of the modeling system (RegESM 2.0). As mentioned in the previous section, the length of the simulations is kept relatively short (1 day) in the benchmark analysis to perform many simulations with different model configurations (i.e., coupling interval, visualization pipelines and domain decomposition parameters).

In the benchmark results, the slightly modified version of the speedup is used because the best possible sequential implementation of the utilized numerical model (stand-alone and coupled) does not exist for the used demonstration application and model configurations. In this case, the speedup is defined as the ratio of the parallel execution time for the minimum number of processors required to run the simulation (Tp(Nmin⁡); based on 140 cores in this study) to the parallel execution time (Tp(N); see Eq. ). S(N)=Tp(Nmin⁡)Tp(N) The measured wall-clock time and the calculated speedup of stand-alone model components (ATM and OCN) can be seen in Fig. . In this case, two different atmospheric model configurations are considered to see the effect of the domain size and non-hydrostatic dynamical core on the benchmark results (LR and HR; Fig. ). The results show that the model scales pretty well and it is clear that the HR case shows better scaling results than LR configuration of the atmospheric component (ATM) as expected. It is also shown that around 588 processors, which is the highest available compute resource, the communication among the processors dominates the benchmark results of LR case, but it is not evident in the HR case that scales very well without any performance problem (Fig. a). Similar to the atmospheric model component, the ocean model (OCN) is also tested to find the best two-dimensional domain decomposition configuration (tiles in x and y directions). As it can be seen from Fig. b, the selection of the tile configuration affects the overall performance of the ocean model. In general, the model scales better if the tile in the x direction is bigger than the tile in the y direction, but this is more evident in the small number of processors. The tile effect is mainly due to the memory management of the Fortran programming language (column-major order) as well as the total number of active grid points (not masked as land) placed in each tile. The tile options must be selected carefully while considering the dimension of the model domain in each direction. In some tile configuration, it is not possible to run the model due to the underlying numerical solver and the required minimum ghost points. To summarize, the ocean model scales well until 588 cores with the best tile configurations indicated in Fig. b.

The performance of the two-component modeling system (CPL) can be investigated using the benchmark results of the stand-alone atmosphere and ocean models. Similar to the benchmark results of the stand-alone model components, the measured wall-clock time and the calculated speedup of the coupled model simulations are also shown in Fig. . In this case, the best two-dimensional decomposition parameters of the stand-alone ocean model simulations are used in the coupled model simulations (Fig. b). The overhead is calculated by comparing the CPL wall-clock time to the sum of the stand-alone OCN and ATM wall-clock time as they run sequentially. The comparison of the stand-alone and coupled model simulations shows that the driver component introduces an additional 5 %–10 % (average is 5 % for LR and 6 % for HR cases) overhead in the total execution time, which slightly increases along with the used total number of processors. The extra overhead is mainly due to the interpolation (sparse matrix multiply performed by ESMF) and extrapolation along the coastlines to match land–sea masks of the atmosphere and ocean models and fill the unmapped grid points to exchange data (Fig. ) and slightly increases along with the increased number of cores as well as the number of MPI (Message Passing Interface) communications between the model components (Figs.  and a).

To further investigate the overhead introduced by the newly designed co-processing component, the three-component modeling system (COP) is tested with three different visualization pipelines (P1, P2 and P3; Table ) using two different atmospheric model configurations (LR and HR) and coupling intervals (3 and 6 min with co-processing). In this case, the measured total execution time during the COP benchmark results also includes vertical interpolation (performed in the ESMF cap of the model components) to map data from sigma coordinates to height (or depth) coordinates for both physical model components (ATM and OCN).

As shown in Fig. b–d, the co-processing components require 10 %–40 % extra execution time for both LR and HR cases depending on the used visualization pipeline when it is compared with CPL simulations. The results also reveal that the fastest visualization pipeline is P3 and the slowest one is P1 for the HR case (Fig. b and d). In this case, the components are all run sequentially, and the performance of the co-processing component becomes a bottleneck for the rest of the modeling system, especially for the computing environment without GPU support like the system used in the benchmark simulations. It is evident that if the co-processing were run concurrently in a dedicated computing resource, the overall performance of the modeling system would be improved because of the simultaneous execution of the physical models and co-processing components. Table  also includes the execution time of the single visualization pipeline (measured by using the MPI_Wtime call) isolated from the rest of the tasks. In this case, each rendering task gets 2–4 s for P1 and P2 cases and 7–15 s for the P3 case in LR atmospheric model configuration. For the HR case, P1 and P2 take around 17–80 s, and the P3 case is rendered in around 8–10 s. These results show that the time spent in the co-processing component (sending data to ParaView/Catalyst and rendering to create the output) fluctuates too much and that this component does not present a predictable and stable behavior. It might be due to the particular configuration of ParaView, which is configured to use software-based rendering to process data in CPUs and load in the used high-performance computing system (UHeM) even if the benchmark tests are repeated multiple times.

In addition to the testing modeling system with various data-processing loads, a benchmark with an increased coupling time step is also performed (see P23M in Fig. c). In this case, the coupling time step between physical model components and the co-processing component is decreased (from 6 to 3 min) to produce output in a doubled frame rate, but coupling intervals between physical model components (ATM and OCN) are kept same (1 h). The benchmark results show that increasing the coupling time step also raises overhead due to the co-processing from 45 % to 60 % for the HR case and pipeline P2 when it is compared with the results of two-component simulations (CPL; Fig. a). It is also shown that the execution time of the co-processing-enabled coupled simulations increases but the difference between the P2 and P23M cases is reduced from 66 % to 37 % when the number of processors is increased from 140 to 588.

In addition to the analysis of timing profiles of modeling system under different rendering loads, the amount of data exchanged and used in the in situ visualization case can be compared with the amount of data that would be required for offline visualization at the same temporal frequency to reveal the added value of the newly introduced co-processing component. For this purpose, the amount of data exchanged with the co-processing component is given in Table  for three different visualization pipelines (P1, P2 and P3). In co-processing mode, the data retrieved from model component memory (single time step) by the driver are passed to ParaView/Catalyst for rendering. In addition to processing data concurrently with the simulation on the co-processing component, the offline visualization (postprocessing) consists of computations done after the model is run and requires storing numerical results in a disk environment. For example, a 3-day long simulation with a 6 min coupling interval produces around 160 GB data (720 time steps) just for a single variable from high-resolution atmosphere component (P1 visualization pipeline) in the case using offline visualization. With co-processing, the same analysis can be done by applying the same visualization pipeline (P1), which requires processing only 224 MB data stored in the memory, in each coupling interval. Moreover, storing results of the 3-day long, high-resolution simulation of the RegCM atmosphere model (in NetCDF format) for offline visualization requires around 1.5 TB data in the case using a 6 min interval in the default configuration (7×3-D fields and 28×2-D fields). It is evident that the usage of the co-processing component reduces the amount of data stored in the disk and allows a more efficient data analysis pipeline.

Besides the minor fluctuations in the benchmark results, the modeling system with the co-processing component scales pretty well to the higher number of processors (or cores) without any significant performance pitfalls in the current configuration. On the other hand, the usage of accelerator-enabled ParaView configuration (i.e., using the NVIDIA EGL library) and ParaView plugins with accelerator support such as the NVIDIA IndeX volume rendering plugin and new VTK-m filters to process data on GPU will improve the benchmark result. The NVIDIA IndeX for the ParaView plugin enables large-scale and high-quality volume data visualization capabilities of the NVIDIA IndeX library inside ParaView and might help to reduce time to process high-resolution spatial data (HR case). In addition to NVIDIA IndeX plugin, VTK-m is a toolkit of scientific visualization algorithms for emerging processor architectures such as GPUs . The model configurations used in the simulations also write simulation results to the disk in NetCDF format. In the case of disabling of writing data to disk or configuring the models to write data with large time intervals (i.e., monthly), the simulations with an active co-processing component will run much faster and make the analysis of the model results in real time efficient especially in live mode (see Sect. ).

While the live visualization was designed to examine the simulation state at a specific point in time, the temporal filters such as ParticlePath, ParticleTracer and TemporalStatistics that are designed to process data using multiple time steps cannot be used in this mode. However, live visualization mode allows connecting to the running simulation anytime through the ParaView graphical user interface (GUI) in order to make a detailed analysis by modifying existing visualization pipelines defined by a Python script. In this case, the numerical simulation can be paused while the visualization pipeline is modified and will continue to run with the revised one. It is evident that the live visualization capability gives full control to the user to make further investigation about the simulation results and facilitate better insight into the underlying physical process and its evolution in time.

The current version of the co-processing-enabled modeling system can process data of multiple model components by using the multichannel input port feature of ParaView/Catalyst. In this case, each model has two input channels based on the rank of exchange fields. For example, the atmospheric model component has atm_input2d and atm_input3d input channels to make processing available for both two- and three-dimensional exchange fields. The underlying adaptor code resides between the NUOPC cap of the co-processing component and ParaView/Catalyst and provides two grid definitions (2-D and 3-D) for each model component for further analysis. In this design, the ParaView co-processing plugin is used to generate Python co-processing scripts, and the user needs to map data sources to input channels by using predefined names such as atm_input2d and ocn_input3d. Then, the adaptor provides the required data to the co-processing component through each channel to perform rendering and data analysis in real time. The fields that are used in the co-processing component are defined by generic ASCII-formatted driver configuration file (exfield.tbl), which is also used to define exchange fields among other model components such as atmosphere and ocean models. Figure  shows a screenshot of the live visualization of the three-dimensional relative humidity field provided by the low-resolution atmospheric model component, underlying topography information and vorticity of ocean surface that is provided by the model component.

In addition to the live visualization mode that is described briefly in the previous section, ParaView/Catalyst also allows rendering and storing data using a predefined co-processing pipeline (in Python) for further analysis. Co-processing mode can be used for three purposes: (1) the simulation output can be directed to the co-processing component to render data in batch mode and write image files to the disk, (2) added value information (i.e., vorticity from wind components, eddy kinetic energy from ocean current) can be calculated and stored in a disk for further analysis, and (3) simulation output can be stored in a higher temporal resolution to process it later (postprocessing) or create a representative dataset that can be used to design visualization pipeline for co-processing or live visualization modes. In this case, the newly designed modeling system can apply multiple visualization and data-processing pipelines to the simulation results at each coupling time step to make a different set of analyses over the results of same numerical simulation for more efficient data analysis. The modeling system also facilitates multiple input ports to process data flowing from multiple ESM components. In this design, input ports are defined automatically by the co-processing component based on activated model components (ATM, OCN, etc.) and each model component has two ports to handle two- and three-dimensional grids (and fields) separately such as atm_input2d, atm_input3d, ocn_input2d and ocn_input3d.

To test the capability of the co-processing component, the evolution of Hurricane Katrina is investigated by using two different configurations of the coupled model (COP_LR and COP_HR) that are also used to analyze the overall computational performance of the modeling system (see Sect. ). In this case, both model configurations use the same configuration of the OCN model component, but the different horizontal resolution of the ATM model is considered (27 km for LR and 3 km for HR cases).

Figure  shows 3-hourly snapshots of the model-simulated clouds that are generated by processing the three-dimensional relative humidity field calculated by the low-resolution version of the coupled model (COP_LR) using the NVIDIA IndeX volume rendering plugin as well as streamlines of Hurricane Katrina, which is calculated using a three-dimensional wind field. The visualization pipeline also includes sea surface height and surface current from the ocean model component to make an integrated analysis of the model results. Figure a–b show the streamlines that are produced by extracting the hurricane using the ParaView Threshold filter. In this case, the extracted region is used as a seed to calculate backward and forward streamlines. In Fig. c–e, sea surface height, sea surface current and surface wind vectors (10 m) are shown together to give insight about the interaction of ocean-related variables with the atmospheric wind. Lastly, the hurricane reaches land and starts to disappear due to increased surface roughness and lack of energy source (Fig. f). While the low-resolution atmosphere model configuration is used, the information produced by the new modeling system enabled investigating the evolution of the hurricane in a very high temporal resolution, which was impossible before. A day-long animation that is also used to create Fig.  can be found in the supplemental video .

In addition to low-resolution model results revealing the evolution of the hurricane in a very high temporal resolution, low- and high-resolution model results are also compared to see the added value of the increased horizontal resolution of the atmospheric model component regarding representation of the hurricane and its structure. To that end, a set of visualization pipelines is designed to investigate the vertical updraft in the hurricane, simulated track, precipitation pattern and ocean state. In this case, two time snapshots are considered: (1) 28 August 2005, 00:00 UTC, at the early stage of the hurricane in Category 5, and (2) 29 August 2005, 00:00 UTC, just before Katrina makes its third and final landfall near the Louisiana–Mississippi border, where the surface wind is powerful and surface currents had a strong onshore component . In the analysis of vertical structure, the hurricane is isolated based on the criteria of surface wind speed that exceeds 20 m s-1, and the seed (basically set of points defined as vtkPoints) for the ParaView StreamTracerWithCustomSource filter are defined dynamically using ProgrammableFilter as a circular plane with a radius of 1.2∘ and points distributed with 0.2∘ interval in both directions (x and y) around the center of mass of the isolated region. Then, forward and backward streamlines of vorticity are computed separately to see inflow at low and intermediate levels and outflow at upper levels for both low- (COP_LR; Fig. a, b, d and e) and high-resolution (COP_HR; Fig. a, b, d and e) cases. The analysis of simulations reveals that the vertical air movement shows higher spatial variability in high-resolution simulation (COP_HR) case even if the overall structure of the hurricane is similar in both cases. As expected, the strongest winds occur in a region forming a ring around the eyewall of the hurricane, which is where the lowest surface pressure occurs.

Also, the analysis of cloud liquid water content shows that low and intermediate levels of the hurricane have higher water content and spatial distribution of precipitation is better represented in the high-resolution case (Fig. a–b and d–e), which is consistent with the previous modeling study of .

It is also seen that the realistic principal and secondary precipitation bands around the eye of the hurricane are more apparent and well structured in the high-resolution simulation, while the low-resolution case does not show those small-scale features (Fig. a–b and d–e). On the ocean side, the Loop Current, which is a warm ocean current that flows northward between Cuba and the Yucatan Peninsula and moves north into the Gulf of Mexico, loops east and south before exiting to the east through the Florida Straits and joining the Gulf Stream and is well defined by the ocean model component in both cases (Figs. c and f; c and f). The track of the hurricane is also compared with the HURDAT2 second-generation North Atlantic (NATL) hurricane database, which is the longest and most complete record of tropical cyclone (TC) activity in any of the world's oceans . In this case, the eye of the hurricane is extracted as a region with surface pressure anomaly greater than 15 mbar (shown as a circular region near the best track). As can be seen from the figures, Katrina moves over in the central Gulf, which is mainly associated with the Loop Current and persistent warm and cold eddies, and intensifies as it passes over the region due to the high ocean heat content in both simulations (Figs. c and f and c and f). The comparison of the low- and high-resolution simulations also indicates that the diameter of hurricane-force winds at peak intensity is bigger in the high-resolution simulation case on 29 August 2005, 00:00 UTC (Figs. f and f). An animation that shows the comparison of low- and high-resolution model results can be found in the supplemental video .

While the main aim of this paper is to give design details of the new in situ visualization integrated modeling system and show its capability, the performance of the coupled modeling system to represent one of the most destructive hurricanes is very satisfactory especially for the high-resolution case (COP_HR). Nonetheless, the individual components (atmosphere and ocean) of the modeling system can be tuned to have better agreement with the available observations and previous studies. Specifically for the analysis of the hurricane, a better storm-tracking algorithm needs to be implemented using ParaView the ProgrammableFilter by porting existing legacy Fortran codes for more accurate storm tracking in both live and co-processing modes.