This article advocates the use of a task-graph-based model for specifying large scientific applications to be executed on cloud environments. The approach is particularly suited for large applications consisting of independent computational parts that can be mapped onto different parts of a distributed execution platform. The task-based approach facilitates the distribution of computations to suitable resources of the cloud environment, since the distribution can be dynamically adapted to the current load of the resources. The software-defined environment of the cloud infrastructure comprises a dispatch unit which serves as interface for clients submitting applications and a server control unit which is responsible for the execution of task graphs. Task-graph-based applications are submitted to the cloud environment via the dispatch unit which also provides a performance and cost estimation. A dynamic scheduling is used to adapt the distribution of the tasks to the resources of the dynamically changing execution environment.