Large-scale simulation and visualization are essential topics in areas as different as sociology, physics, urbanism, training, entertainment among others. This kind of systems requires a vast computational power and memory resources commonly available in High Performance Computing HPC platforms. Currently, the most potent clusters have heterogeneous architectures with hundreds of thousands and even millions of cores. The industry trends inferred that exascale clusters would have thousands of millions. The technical challenges for simulation and visualization process in the exascale era are intertwined with difficulties in other areas of research, including storage, communication, programming models and hardware. For this reason, it is necessary prototyping, testing, and deployment a variety of approaches to address the technical challenges identified and evaluate the advantages and disadvantages of each proposed solution. The focus of this research is interactive large-scale crowd simulation and visualization. To exploit to the maximum the capacity of the current HPC infrastructure and be prepared to take advantage of the next generation. The project develops a new approach to scale crowd simulation and visualization on heterogeneous computing cluster using a task-based technique. Its main characteristic is hardware agnostic. It abstracts the difficulties that imply the use of heterogeneous architectures like memory management, scheduling, communications, and synchronization — facilitating development, maintenance, and scalability. With the goal of flexibility and take advantage of computing resources as best as possible, the project explores different configurations to connect the simulation with the visualization engine. This kind of system has an essential use in emergencies. Therefore, urban scenes were implemented as realistic as possible; in this way, users will be ready to face real events. Path planning for large-scale crowds is a challenge to solve, due to the inherent dynamism in the scenes and vast search space. A new path-finding algorithm was developed. It has a hierarchical approach which offers different advantages: it divides the search space reducing the problem complexity, it can obtain a partial path instead of wait for the complete one, which allows a character to start moving and compute the rest asynchronously. It can reprocess only a part if necessary with different levels of abstraction. A case study is presented for a crowd simulation in urban scenarios. Geolocated data are used, they were produced by mobile devices to predict individual and crowd behavior and detect abnormal situations in the presence of specific events. It was also address the challenge of combining all these individual’s location with a 3D rendering of the urban environment. The data processing and simulation approach are computationally expensive and time-critical, it relies thus on a hybrid Cloud-HPC architecture to produce an efficient solution. Within the project, new models of behavior based on data analytics were developed. It was developed the infrastructure to be able to consult various data sources such as social networks, government agencies or transport companies such as Uber. Every time there is more geolocation data available and better computation resources which allow performing analysis of greater depth, this lays the foundations to improve the simulation models of current crowds. The use of simulations and their visualization allows to observe and organize the crowds in real time. The analysis before, during and after daily mass events can reduce the risks and associated logistics costs.