According to Chris Fournier, who is teaching a class titled "Analyzing Embedded Multicore Processor and System Capabilities (ESC-207), " at the Fall 2008 Embedded Systems Conference in Boston, when it comes to such things as the definition of "fast," it is also important for embedded systems developers to pay attention to context when assessing performance and benchmarking multi-core systems.

Fournier, a senior product marketing manager in AMD's embedded computing solutions division, said that switching to multi-core technology presents as many challenges as it does benefits.

"One of those challenges is simply analyzing the potential performance benefit of a multicore processor, system, or system-on-a-chip (SoC)," he said. " How much faster will a multicore approach be? Guessing and rules of thumb don't work here.

"Putting multiple processor cores into a single chip doesn't automatically guarantee big multiples of processing power. Two cores aren't always twice as fast; eight cores definitely won't be 8x faster. Furthermore, there's no guarantee that a multicore processor will deliver any dramatic increases at all in your system's capabilities, its computing resources, or its throughput."

Just as the definition of the word "fast" depends on the way it is being used and where it is being used, assessing the performance of a multicore system depends on many factors. In the context of assessing the performance of a multi-core system, he said, the benchmarks that a developer chooses depend on the context in which they are used and must test for and isolate each of the factors that may impact on performance.

"So multiple different benchmarks are essential to get even a rough approximation of a multicore system's behavior," said Fournier. "For this reason, the much-abused Dhrystone MIPS numbers - which aren't even useful for single-core performance - are particularly suspect when it comes to describing multicore systems.

"As always, the very best benchmark is your own application, but short of porting your code to every potential platform, using industry-standard benchmarks can offer a close approximation."

To be accurate and useful, a multicore benchmark must measure three fundamental areas: data throughput, computational throughput, and data decomposition. Data throughput shows how well an application can scale with more data inputs.

This can be accomplished by duplicating the same computation and applying it to multiple different datasets. Real world examples of this include the decoding of multiple different JPEG images (as may occur when viewing a web page), decoding multi-channel audio, or running a VoIP application with multiple channels whereby each channel receives different input data.

Computational throughput, in contrast, involves tests that can initiate more than one task at a time, implementing concurrency over both the data and the code. This demonstrates the scalability of a solution for general-purpose processing.

As an example, consider the execution of MPEG decode(x) followed by MPEG encode(x), which is similar to what you might find in a set-top box where the satellite signal is received, decoded, and encoded into a different quality signal for storing on the hard disk.

Data decomposition is where an algorithm is divided into multiple threads working on a common data set, demonstrating support for fine-grained parallelism. In this situation, the algorithm could be working on a single combined audio/video data stream, but the code can be split in such a way so as to distribute the workload among different threads, each of which could be handled by a different processor core.

In his class, Fournier focuses much of his attention on how to use new industry-standard benchmark tests from the Embedded Microprocessor Benchmark Consortium (EEMBC) to demonstrate and assess a multicore system's behavior across a variety of scenarios.

"These can help you predict how your own multicore system will function," he said. "Initial results from these benchmarks reveal some interesting secrets that, we hope, will encourage you to pay close attention to the combined effects of the multicore processor, the memory subsystem, the operating system, and other system-level characteristics.

To register to attend this and other classes and events at the conference, go to the Fall ESC Registration Page and sign up.