editorial

The Hardware Under the OS Wars The critical distinctions between workstations and "ordinary" desktop machines become obvious when the platforms are running the larger EDA jobs. by Tets Maniwa

In the process of evaluating the tools and technologies for alternatives to the Unix workstations, some people may be measuring the wrong parameters, or trying to compare apples and oranges. A critical task in evaluating alternative hardware and software for important tasks like design sign-off is to separate the hardware performance and capabilities from that of the software. The wrong set of hardware can undermine the performance of any tool--or the underlying operating system itself.

I've heard of one ASIC vendor who was starting to qualify a set of NT tools for sign-off. Using a method similar to our benchmark evaluations, the vendor compared the performance of the basic tool set on a range of designs--small, medium, and large--across several platforms. Not surprisingly, the workstations and the PCs exhibited similar performance on the small and medium designs. When the vendor tested the large design, however, the PCs took substantially longer to complete their tasks than the workstations did.

Because this behavior didn't make immediate sense, the engineers had to investigate and explain the discrepancy. They noted that the processor speeds and much of the balance of the hardware were very similar. The main difference, then? The PCs were configured as "typical" desktop machines with only 64 Mbytes of RAM. The limited memory forced the machine to thrash data almost continuously to and from the hard drive in moving the large files necessary for EDA tasks through to the processor. This small complement of memory can't hope to provide comparable performance to workstations with a Gbyte of RAM. Obviously, the standard desktop PC isn't in the same performance category as a machine configured as a workstation, even when the processor speeds are similar.

Workstations have always outperformed standard desktop computers because manufacturers optimize the entire hardware package for much higher performance than a typical PC can deliver. Since the workstation must address very large memory spaces, it contains a high throughput cache--not just a write-through cache but more likely an n-way set associative cache--to keep the data and instruction pipelines full. The high-performance microprocessors run with clocks over 500 MHz and process multiple instructions per clock. Therefore, the memory and cache systems need to supply up to 6 instructions and data sets every 2 nanoseconds.

As the processor speeds continue to increase, we are starting to see the memory and I/O subsystems become the limiting factors in overall system performance. The design of these elements of the workstation will thus become even more important to overall performance. To support the data demands of the largest EDA jobs, the workstations must contain very large, high-speed hard drives (and now a RAID structure to add speed and redundancy), as well as the capability for fast networking and I/O. A team of design engineers could easily challenge or exceed the capacity of even a gigabit Ethernet network when sharing files for a system-on-a-chip design. In addition to the networking and I/O, the typical workstation supports large monitors driven by high-performance video subsystems.

To address some of these issues of relative system performance and provide legitimate comparisons, Integrated System Design is continuing to work with the workstation vendors and Seva Technologies in Fremont, Calif. to gauge the performance of the Windows NT and Unix platforms for EDA applications. We plan to evaluate the latest hardware and software to highlight the abilities of different platforms to perform the common EDA tasks. We will try to use roughly comparable hardware configurations for the evaluations, to prevent mismatches in subsystem performance from skewing the overall evaluations. We are shooting for apples and apples, oranges and oranges.

Because fierce competition drives workstation vendors to meet ever-increasing performance demands, we've found that the differences between the fully loaded workstations have become negligible. In the near future, we will look at performing benchmark evaluations with different sets of hardware--a typical workstation with 256 Mbytes of RAM, a large machine with 1 Gbyte of RAM, and possibly a cluster of workstations or servers. These different configurations will allow us to measure how much the subsystems affect the overall throughput in the machines, compared to the performance of the underlying microprocessor.

To voice an opinion on this or any Integrated System Design article, please email your message to jeff@isdmag.com.

integrated system design  May 1999