NCAR's newest supercomputer will enable scientists to cut the time for complex scientific calculations in half. Called bluevista, the machine will work alongside the Mesa Lab's other two supercomputers: an IBM POWER4-based machine and an IBM e1350 Linux cluster. The combined computational power of the three supercomputers places NCAR among the top 25 institutions worldwide as measured by computational capacity, estimates Tom Bettge, deputy director of SCD.
The new machine is an IBM p575 based on IBM's POWER5 processor. It was delivered to NCAR in late August, and software installation began on 6 September. Testing will continue throughout the fall. During this period, all of bluevista's time that is not needed for the testing is allocated to an effort to nest a regional climate model into the Community Climate System Model (see sidebar).
SCD plans to make bluevista available to the community in January. "At that time, our supercomputing capacity will nearly double, instantaneously," Bettge says. "The number of sustained flops [floating-point operations per second] on our typical workload will go from 400 gigaflops to almost 800 gigaflops."
He adds that a typical application will run about twice as fast on bluevista as it does on bluesky, the POWER4-based machine. (NCAR computer names are spelled as a single lowercase word to meet Unix requirements.)
Bluevista is smaller, hotter for its size, and denser than bluesky. Although it takes up only a third as much floor space in the Mesa Lab as the older supercomputer, it requires two-thirds as much power and cooling. Bettge estimates that bluevista will need more than 250 kilowatts of power to operate. The average personal computer consumes 0.12 kilowatts.
Even as bluevista is being tested, SCD is working on plans for the next increase in NCAR's computing capacity. In 2007, it anticipates deploying a new supercomputer system, called the Integrated Computing Environment for Scientific Simulation. The system is expected to provide two to three times the computing capacity of bluevista.
A giant step down the road to multiscale modeling
by Carol Rasmussen
Modeling atmospheric phenomena such as hurricanes could make a scientist feel like Goldilocks. Global climate models are too big; these mesoscale phenomena fall through the cracks of the grid. On the other hand, regional or weather models are too small; they don't allow the researcher to observe inputs and feedbacks to global conditions. A model that is just right would allow the researcher to move effortlessly between scales, zooming in to watch the evolution of a smaller-scale feature and zooming out when the computer-intensive finer scale is no longer needed.
This kind of multiscale modeling is the long-term goal of a research project that has been going on at NCAR for several years. The project, called Predicting the Earth System Across Scales, is taking a large step forward this fall with the allocation of all of bluevista's free time during the machine's three-month testing period. Initiated by Greg Holland (MMM) and Jim Hurrell (CGD), the project now includes leadership from researchers in three NCAR divisions and at the Pacific Northwest National Laboratory (PNNL) who are reconfiguring the Weather Research and Forecasting model (WRF) to a climate configuration dubbed the Nested Regional Climate Model, or NRCM.
In the first phase of this work, researchers at NCAR and PNNL configured WRF so it could be coupled with the North American part of NCAR's global-scale Community Land Model in a downscaling mode (that is, the global model output could affect the course of the regional model but not vice versa). For the bluevista testing this fall, the researchers are trying a new configuration with the NRCM inserted in the Community Atmospheric Model to cover the entire tropics. Inside this tropical model, nestings that allow both downscaling and upscaling will zoom in to a scale where the local organization of convection and clouds can be simulated explicitly. The NCAR and PNNL colleagues will be testing how well the combination can model tropical moist convection and how that process interacts with tropical modes of variability, which have historically not been a strength of global climate models. The model will do a series of five-year runs to include an El Niño–Southern Oscillation cycle.
The work will produce a data set that can be used to test current ideas on the scale interactions involved in tropical
cyclogenesis. It may also throw some light on the relationship between global climate change and changes in tropical cyclone intensity. And finally, it will give the researchers a better understanding of the challenges that lie ahead on the road to seamless multiscale modeling.
A comparison of winter precipitation as observed (left) and simulated by the nested-grid (center) and global-scale (right) models during the first phase of the multiscale-modeling project shows the improved accuracy that the nested model provides. Click here or on the image to enlarge it.