Advances in clustering technology have redefined the price-to-performance curve for many High Performance Computing (HPC) application areas. The use of specialized high-speed interconnects and fast commodity processors have pushed the envelope to where it is today.

Not all applications need this level of hardware (and cost) to achieve leading-edge price to performance. Indeed, there have been several technological advances that may invite a step back from the traditional edge-of-technology approach to HPC clustering. These advances include the following developments:

• Introduction of low-cost high-performance multi-core processors
• Introduction of high-density motherboards and packaging solutions
• Introduction of optimized Linux Gigabit Ethernet performance

Of particular importance is the fact that these advances are in the commodity sector where high demand and economies of scale have created reasonable price points. Furthermore, alternative high-cost approaches that employ enhanced interconnects and multi-socket motherboards may not be required for certain application classes. Users in this category can expect the commodity approach to deliver new levels of industry-leading price to performance.

In this article, we will discuss how these advances can be used to optimize cluster performance. In addition, we will highlight application areas where these types of clusters are expected to provide optimum performance.

Gaining the Multi-Core Advantage
The multi-core revolution is here. All major processor families have begun using multiple CPU cores to enhance performance. Currently, dual core processors are available at various performance and price levels. Remarkably, most HPC users can immediately benefit from these advances, as most HPC cluster software is designed to use multiple processors.

Specifically, current dual core designs allow HPC users to effectively double the number of processing units while still enjoying traditional commodity price points. In the HPC market, more CPUs are always welcome, but the right design (choice of processor, motherboard, and packaging) is critical to achieving the desired performance.

The recently introduced Pentium D (Presler) processor and Xeon 3000 processers from Intel are examples of commodity high-performance processors. The Presler series is a dual core processor manufactured using the latest 65nm process and is currently available at speeds up to 3.40 GHz. More important for HPC users is that each Presler has a total of 4MB of on-chip cache that it divides evenly between the two cores (2MB each). These caches are fed using an 800MHz FSB and DDR2 memory.

In the HPC cluster sector, the processor battle has typically been between the high-end Intel Xeon or the AMD Opteron. Little consideration has been given to "lesser processors" in the HPC space. As this report will show, this assumption may not hold when actual price and performance numbers are determined.

Check the Numbers - Presler Is on Top
The SPEC benchmarks are usually a good rating of overall processors performance. Table 1 shows the SPEC benchmarks for an Intel Pentium D (model 940) and an AMD Opteron (model 270). Pentium D 940 performs at a level 10% greater than the Opteron 270, yet at this point in time, it's priced at half the cost of an Opteron 270.

While the SPEC benchmarks are an important yardstick, real application benchmarks often provide a second data point with which to compare processors. The GROMACS molecular dynamics package is known to push processors very hard and is therefore a good test of overall number-crunching capability. The results shown in Table 2 are for the Gromacs Benchmark Suite (Linux Version 3.3). See the references at the end of this article for more information on GROMACS. All results are normalized to the Pentium D (lower means slower) and were run using one processor.

The results show a substantial performance advantage over the Opteron 270 processor. The Opteron 270 numbers were taken from the Gromacs Web site (www.gromacs.org).

Breakthrough Design - The Caretta Motherboard
When designing clusters, the "more is better" model often works. However, the number of processors (and hence cores) that can be placed on a motherboard needs to be considered carefully. Modern cluster designs currently take advantage of dual socket motherboards and single core processors. While this approach has helped improve processor density, extending this design with dual cores may have some unexpected results. Using dual core processors on dual socket motherboards requires that the memory subsystems and interconnect now service four cores (instead of two) at the same time.

This situation can, in certain cases, seriously degrade the maximum achievable performance of each core. Optimizing onboard memory subsystems is one way to mitigate memory contention, but this approach also introduces a "nonlocal" or NUMA (Non-uniform Memory Access) type of memory structure. In the end, the application determines the best approach, but rethinking the dense core motherboard approach may have some advantages.

A potentially more optimal solution for many applications would be a small single socket motherboard on which a dual core processor can reside. Such a system would resemble current dual socket motherboards/single core clusters designs in use today, on which memory and interconnect contention is well understood.

The recent introduction of the Intel Caretta motherboard (S3000PT) has been designed to fill this need. The Caretta motherboard supports the Intel Xeon 3000, Pentium D, and Pentium 4 processors, four DIMM slots (DDR2 533/667 with ECC, two-way interleaved, unbuffered), Integrated two port SATA 3.0Gb/s with RAID 0 &1, an ATI ES1000 (16MB), Dual Gigabit Ethernet LAN, and a 5.95 inch x13 inch Form Factor. Interestingly, the Form Factor is one half the size of an Extended ATX motherboard (12"x13"). These dimensions allow a standard Rack Mount ATX enclosure to hold two Caretta motherboards.

The Caretta allows the density found on dualcore/dualsocket motherboards, but provides each processor with its own local memory environment.

This approach has further advantages as well. As more cores/processors are placed on the motherboard, a node failure (motherboard/power supply/hard drive) removes all the cores/processors from the cluster. By using a separate motherboard for each processor, failures are limited to two cores (one processor).

The HyperBlade Advantage
When deploying a high density production HPC cluster, correct system packaging will ensure continuous operation. While many users find utility in deploying 1U server packaging solutions, blade systems are designed with a higher level of custom integration. Blades are typically easier to manage, but more expensive that 1U servers. A hybrid solution where commodity components can be packaged in a blade-like fashion has been developed by Appro International. The advantages of this solution include the use of commodity components inside the "blade" and the integration and manageability of bladed systems.

Like blades, the Appro HyperBlades are modular servers plugged into a common backplane that eliminates cable clutter. By using a vertical mount approach, the Appro HyperBlade offers an enhanced density, providing up to 50 servers in a standard 42U rack cabinet. Large and smaller rack systems are available as well. Because the HyperBlade is designed to include the flexibility of a typical 1U server, high speed interconnects options - including Myrinet, Dolphin, Quadrics and InfiniBand™ - are easily deployed.

In addition, HyperBlades offer the power advantage of 1U servers while, at the same time, they provide an integrated power control and serial management capability found in more expensive blade systems. Finally, each Appro HyperBlade can hold two Caretta motherboards, thus providing excellent processor density (four cores per HyperBlade) and easy management.

The Gigabit Ethernet Advantage
While there are many choices for cluster interconnects, the preferred and lowest cost option is Gigabit Ethernet. While often dismissed as underpowered for today's clusters, actual tests show the exact opposite is true for some application classes. Shown in Figure 1 is a NetPIPE TCP throughput graph for a Gigabit Ethernet link between two Pentium D 940 processors.

The connection used an onboard Intel 82573 chipset, an e1000 driver, and a 1500 byte MTU. It should be noted that the single byte latency was 36 microseconds. Not all applications can scale well with this level of performance. However, there are many that will find commodity Gigabit Ethernet more than adequate for their computing needs.