Break Through Application Performance
No mater how good low level benchmarks perform, the combined performance of a cluster is what ultimately determines its viability as a high performance computing system. In order to test the scalability and overall effectiveness of a true commodity cluster, eight Pentium D nodes were configured into a test cluster. Each node had one Pentium D 940 running at 3.2 GHz, 8 GBytes of RAM, and was connected to an 8 port Gigabit Ethernet switch (SMC 8505T).

The NAS Parallel Benchmarks (NPB Version 2.3) are good overall tests of cluster scalability. These tests are a small set of programs designed to help evaluate the performance of parallel supercomputers. The benchmarks, which are derived from computational fluid dynamics (CFD) applications, consist of five kernels and three pseudoapplications. See the references at the end of this article for more information on the specific programs. Table 3 illustrates the Linux cluster scaling capability of the Class B test suite. Scaling is defined as speed of Ncores/speed of one core.

The results using four and eight processors are quite good considering the cluster is using a low cost Gigabit Ethernet interconnect. (Test IS is known not to scale well with Ethernet.) Given the speed of the Pentium D, it is quite remarkable that almost all the tests show acceptable scaling at four and/or eight processors. Perhaps the most interesting result is the extra performance gained by using all 16 cores for some of the tests. From the table above, EP, FT, and LU were able to gain additional performance by utilizing the extra core on each processor.

In the case of the LU benchmark, 8 processors using 16 cores delivered 6.3 GFLOPS of combined performance which was an increase of 2.5 GFLOPS over the 8 processor single core result.

Those tests that did not scale well using the extra cores (e.g., CG, BT) may require additional network throughput to accommodate the extra eight parallel tasks.

Scaling GROMACS
In addition to single processor performance benchmarks, GROMACS also supports Linux parallel execution using MPI (Message Passing Interface). Table Four provides the performance and scaling results for the DPPC Phospholipid membrane benchmark.

These results are important for two reasons. First, the use of commodity hardware, including low cost Gigabit Ethernet, is still a viable method to achieve supercomputer level performance at low cost. Second, the presence of an additional core in the Pentium D, at essentially no extra cost, pushed the performance level to that of much larger and more expensive solutions.

For the GROMACS benchmark, the 8-way scaling is excellent; and gaining (for free) an additional 3.2 GFLOPS from the second cores using Gigabit Ethernet is a huge win for this class of cluster.

Price-to-Performance Data
When evaluating high performance computing systems, it's important to keep in mind the price-to-performance of the components. If a Pentium D 940 cluster and an Opteron 270 cluster were purchased, many of the components would be identical (i.e., HyperBlade, hard disk, etc.). The price differences would exist in the processor, motherboard, and memory. If these prices are taken into account using current online market pricing, the Opteron node would cost approximately 1.4 times more than the Pentium D node.

If the price ratio data are combined with the GROMACS performance data, then the price-to-performance of the Opteron 270 solution is almost double that of the Pentium D 940 solution.

The price and performance differences are shown in Table Five. The GROMACS performance is an average of the data in Table Two. It should be noted that the scaling behavior of a similar Opteron cluster node was not available for this benchmark. Assumptions about scalability should be always be tested.

Recommendations
Based on the above analysis, the following recommendations will help maximize price-to-performance of an HPC cluster:

1)  The Pentium D platform offers substantial price-to-performance advantages over more traditional server based systems and should be considered for HPC clusters. In addition, the use of Gigabit Ethernet has also been shown to produce exceptional results for this range of testing.
2)  Combining optimized commodity hardware with leading edge HyperNode packaging from Appro International will allow complete integration of the high density Caretta motherboards into the cluster environment.
3)  The need for benchmarking should be emphasized when evaluating a cluster architecture. As shown in this article, proper benchmarking can often test assumptions about hardware and provide an optimum price performance for your HPC needs. Any choice of cluster hardware should be made on the basis of a benchmark analysis.
4)  The use of dual core processors can be a huge win for certain codes. In this study we did not test the effect of running different MPI programs on separate cores on the same processor. We expect - due to different memory access patters and network usage patters - that this methodology should also see similar price-to-performance gains.

Testing Methodology
Tests were conducted using eight dual core Intel Pentium D (model 940) Presler servers operating at 3.2 GHz. Each server used a Nobhill motherboard (Intel Model SE7230NH1) which is functionally equivalent to the Caretta motherboard, but larger in size. Each node had 8GB of DDR2 RAM and two Gigabit Ethernet ports (only one of which was used for the testing). A SMC 8508T Ethernet switch was used to connect the servers. Ethernet drivers were also set to provide best performance for a given test. In addition, where appropriate, the MPI tests were run with "sysv" flag to cause the same processor cores to communicate through memory. Contact the author for details whose information can be found in the references of this article.

The software environment was as follows:

• Base OS: Linux Fedora Core 4 (kernel 2.6.11)
• Cluster Distribution: Basement Supercomputing Baseline Suite
• Compilers: gcc and gfortran version 4.0.2
• MPI: LAM/MPI version 7.1.1