Declining prices of computer hardware coupled with ever-increasing performance will continue to directly benefit the CFD practitioner. Nowhere has this been more apparent than in the PC space, where competition for the personal and business computer market has led to a rapid evolution of fast and affordable computing technologies. Similar trends exist for networking hardware where, for example, fast ethernet switches and network interface cards are now considered commodity products. Related to this has been a proliferation of miniature computing devices, including powerful laptops and handheld devices such as personal digital assistants (PDAs). Wireless technologies are also emerging, allowing such devices to communicate without a loss of portability.

Fluent's computer services and product development staff recently had the opportunity to put some of these technologies to task using their CFD software products. In one example, Fluent enlisted the service of several HP Omnibook 6000 laptops, prior to their being deployed to sales and training engineers. Each laptop contained an Intel 850 MHz Mobile Pentium III processor and 512MB of RAM, and came equipped with an integrated network interface, so creating a cluster simply entailed plugging each machine into a fast ethernet switch. Both Microsoft Windows 2000 and Red Hat Linux 7.1 operating systems were loaded onto each machine.

Figure 1: Line contours of vorticity magnitude on the combustor center-plane and filled contours of temperature on the flame surface

The laptop cluster, running Linux, was used to solve a combustion problem in parallel. The example was from the FLUENT 6.0 test matrix in the area of gas turbine combustion. In the simulation, the partially premixed combustion model in conjunction with the large eddy simulation (LES) turbulence model was used to solve for the mean and variance of the mixture fraction and the reaction progress variables. The model contained 250,000 cells. Figure 1 shows contour lines of vorticity magnitude on the combustor center-plane and contours of temperature on the flame surface. Parallel performance is presented in Figure 2. The superlinear performance is due to FLUENT's parallel multigrid solver requiring fewer cycles to achieve the same level of convergence as the serial run for this particular simulation ¹.

Figure 2: Parallel performance for the gas turbine com-bustor simulation solved with the LES turbulence model

The Linux OS is making rapid inroads, not only on traditional computers, but also on handheld devices. At the 2001 US UGM, Fluent demonstrated the laptop cluster with results displayed on a Compaq iPAQ H3670 handheld device over wire-less ethernet (Figure 3). The iPAQ, with 64MB RAM, was running Linux and X11 windows. Although the bandwidth provided by wireless ethernet (11Mbs peak) is not sufficient to provide message-passing communication for the parallel solver, it is more than enough to handle the X11 traffic needed to interact with FLUENT's user interface ².

Figure 3: Results from a CFD simu-lation run on a laptop cluster are displayed on a Compaq iPAQ H3670 handheld device, using wireless ether-net

Recently, Intel released its Pentium 4 processor. The Pentium 4 is significant not only for its fast clock rates (now up to 2GHz), but also for its fast 400 MHz system bus which delivers a three-fold improvement in bandwidth over the Intel Pentium III. FLUENT's performance is heavily dependent on memory bandwidth, and the improvements in system bandwidth account for the impressive benchmark results reported on our web site (see www.fluent.com/software/fluent/fl5bench). The FLUENT 5 benchmarks posted by HP and Dell on Pentium 4 systems are among the fastest we have seen. Figure 4 shows the excellent performance and scaling of the FL5M2 benchmark on a cluster of Dell Precision Workstation 330 machines that use 1.4GHz Pentium 4 processors. Fluent will continue to advance its capability for cluster computing, and is collaborating with several partners on this front.

Figure 4: Performance and scaling of the FL5M2 benchmark on a cluster of Dell Precision Workstation 330 machines

Notes

1. Parallel scaling is dependent on hardware, parallel implementation, and model properties. Most simulations will not exhibit superlinear behavior.
2. FLUENT's user interface, Cortex, was running on one of the laptops, but its display was sent to the iPAQ using X11's standard client/server capabilities.

FluentNEWS