Abstract:
In this report, we present a comparison of model performance indicators for several
operational coastal forecast systems and structural engineering predictions and evaluations
executed in local workstations, in HPC cluster nodes and in a pilot cloud, aiming at
contributing to the best choice for the National Infrastructure for Scientific Computing.
Results show that the scalability and flexibility of cloud computing resources makes them
an attractive alternative for the implementation of multiple forecast systems using serial,
non-MPI models, as well as for sensor data acquisition and processing applications.
For MPI-based models, tests using cloud virtual machines with resources equal to or lower
than the smaller physical bases performed well relative to the other resources. However, as
the cloud resources under testing did not reach the optimal number of processors for the
present use cases, the HPC cluster remained the best option, as it fits better the
requirements for the typical dimensions of computational grids for multi-scale (port to
ocean) analysis.
Federated cloud resources allowed a better performance for small pool sizes, allowing the
combination of the processing power of several hosts. However, the performance does
scale very badly if the choice relies in any combination that uses many processes (by using
many hosts or many processes within each host), even if we use resources with some
hardware assistance.
Further testing is still necessary to explore this possibility in detail, taking into account the
need to assure an adequate quality of service (QoS), especially to meet forecasting
deadlines and real-time streaming bandwidth.
We conclude that an evolution from the current cluster setup to a cloud-based architecture
will satisfy most of our simulation requirements while offering a more flexible and
affordable computing environment.
