One reason for the success of in-memory (transactional) data grids lies on their ability to fit elasticity requirements imposed by the cloud oriented pay-as-you-go cost model. In fact, by relying on in-memory data maintenance, these platforms can be dynamically resized by simply setting up (or shutting down) instances of so called data cache servers. However, defining the well suited amount of cache servers to be deployed, and the degree of in-memory replication of slices of data, in order to optimize reliability/availability and performance tradeoffs, is far from being a trivial task. To cope with this issue, in this article we present a framework for high performance simulation of in-memory data grid systems, which can be employed as a support for timely what-if analysis and exploration of the effects of reconfiguration strategies. The framework consists of a discrete event simulation library modeling differentiated data grid components in a modular fashion, which allows easy (re)-modeling of different data grid architectures (e.g. characterized by different concurrency control schemes). Also, the library has been designed to be layered on top of the open source ROOT-Sim parallel simulation engine, natively offering facilities for optimized resource usage in the context of model execution on top of multi-core and cluster based architectures. Finally, instances of data-grid models supported by the framework have been validated against real measurements obtained by deploying the Infinispan data grid onto Amazon EC2 virtual clusters, and running the well known TPC-C benchmark. By the experiments we demonstrate closeness of simulation outputs and real measurements, while jointly showing extreme scalability of the framework, in terms of speedup and ability to manage extremely large data grid models.