How to Choose Between Dedicated Versus Host-Based Processing -- Part II

Several years ago, the question of whether to use dedicated or host-based processing was a simpler decision than it is today. In years past, you could make most measurements on dedicated instruments that performed all analysis on raw data with dedicated processing within the instrument. You only performed host-based processing on a computer if you could analyze the data long after the acquisition. Recent advances in computers have seen a dramatic increase in processor speed and bus interface speeds, making the differences between dedicated processing and host-based processing harder to distinguish. Read on to discover what you need to consider when choosing between these two approaches. In this part I of this article, we consider a typical measurement system that consists of an instrument and a computer performing signal analysis on an acquired time-domain waveform.

You need to consider the question of dedicated versus host-based processing when deciding between different dynamic signal acquisition (DSA) products from National Instruments. The NI 455x Series dynamic signal analyzers have an onboard DSP, which is dedicated to performing signal processing on data from the analog to digital converters (ADCs). This Series uses the PCI bus interface. The NI 445x Series uses the PCI bus interfaces to the computer. These products rely on the host processor to perform all signal analysis.

Introduction

In Part I, we discussed a typical measurement system that consists of an instrument and a computer performing signal analysis on an acquired time-domain waveform. In Part II, we consider the effects of other operations on you processing routine. In addition, we offer suggestions to help you determine whether dedicated or host-based processing is the best choice for your application.

Determine Analysis Routines

The choice of analysis routines and how you use them can have a big impact on processing speed and performance. For example, you can calculate the FFT to a varying degree of resolution. Most dynamic signal analyzers with a dedicated processor can perform an FFT with 1,600 lines with real-time or gap-free performance. With current PC processors, you can also achieve this with host-based processing. If you want a larger FFT, a 12,800 line FFT for example, dedicated processing is not an option with currently available technology. You can perform an FFT of this size or larger with post-processing on the host computer. This example also illustrates an advantage of host-based processing over dedicated processing. It is possible to perform very large user-defined analysis functions. This flexibility of analysis tools is not present on most instruments with dedicated processing.

Consider Streaming to Disk

A common feature on many measurement systems is the capability of streaming data directly to disk. This feature is not available on measurement systems based on the GPIB interface bus or other slower interface buses. Data that is sent to disk can be direct time-domain data or data resulting from computations in a processor. Disk drives have limitations on how fast they can write data to disk and can become a bottleneck in the measurement system. The danger is losing data because the disk cannot write the data as fast as it is being acquired or processed. Thus, you would compromise real-time and gap-free processing and lose data. All considerations mentioned in this paper, channel count, frequency range, analysis routines, data display, and other operations can affect real-time and gap-free performance when writing to disk. Alternatively, you could purchase a faster disk drive, reduce the channel count, reduce the sampling rate, and reduce or remove the analysis during the operation in order to perform post-processing after the acquisition has completed.

Judge Data Display

The amount of data that is displayed during the acquisition on a computer can affect the acquisition and real-time or gap-free performance. Each data plot and graph consumes processor time, thus more graphs and plots mean less time to compute analysis on the host processor. Having a good video card on a computer can help relieve the burden of displaying data from the host-processor. This consideration also demonstrates why a fast host-processor is valuable for use even with instruments using dedicated processing. If a fast host-processor is used with an instrument using dedicated processing, the measurements are much less sensitive to additional results displayed at the same time.

Look at Other Operations

One of the key advantages to using dedicated processing is its relative immunity to other operations that are simultaneously performed on the host processor. Host-based processing on the other hand is greatly affected by other operations occurring simultaneously. The host processor must divide time to all tasks that are under its control. If there are many other tasks performed simultaneously, then time can be taken away from the processing, thus slowing it down, limiting gap-free performance.

There are a number of operations that can occur on a computer simultaneous with the signal processing, some of which are beyond the user’s control. Some operating systems dynamically allocate memory and perform a number of memory management operations behind the scenes without user intervention. While this is useful in many situations, the user does not have much control over when this occurs. Device drivers are often interrupt driven, meaning that when they make a request for time from the operating system they can interrupt the current process. For example, every time the mouse is moved on screen or a key is entered from the keyboard, time is taken from the computer processor to carry out these operations. The user interface or computer display is another consumer of host processor time. Any time data is displayed or user input is entered, time is taken from the host processor and thus potentially from the signal processing. Of course, any time another program is running at the same the same host processor is controlling time or another instrument; operations are going to occur simultaneously with the signal processing.

Resulting Benchmarks

We performed benchmarks using an NI 4551, which has a dedicated processor, and using an NI 4451, which uses the host processor. We calculated an auto power spectrum on signals acquired simultaneously from two channels of the DSA devices. The X-axis shows the number of FFT lines of resolution. The NI 4551 uses the NI-DSA instrument driver and can perform FFT analysis with up to 1,600 lines of resolution. The host processor has an almost unlimited amount of resources to perform high resolution FFT analysis. We used two desktop computers to perform the host-based processing, one with a 400 MHz processor and another with a 700 MHz processor. The maximum frequency bandwidth of the DSA devices is 95 kHz. We show this parameter on the Y-axis. Because the sampling rate determines the frequency range of the instrument and how fast data is streamed to the processor, the frequency range is affected by the performance of the processor. See results below.


The tests show progressing analysis complexity as the X-axis moves from left to right by increasing the lines of frequency resolution of the FFT. The height on the Y-axis shows the maximum frequency that you can analyze gap-free at the given frequency resolution. Note that with NI-DSA (using the NI 455x instruments), the processing power of the dedicated DSP is limited to 1,600 lines of resolution when running two channels. At lower frequency resolutions it is as fast or faster than processing on the host. Note that the NI 445x device with the 700 MHz computer processor can keep up with the NI 455x device with DSP in almost all cases, illustrating that it is possible to perform many measurements gap-free or real-time with a very fast host processor.

Conclusions and Suggestions

The line that was once drawn differentiating dedicated and host-based processing has become blurred with new advances in computer processor technology. Processing speeds of computers have increased dramatically so that you can achieve real-time and gap-free performance with host-based processing, not just with dedicated processing. There are still some limitations depending on the application, which can compromise real-time and gap-free performance with host-based processing. We have discussed several alternatives and trade-offs to help you select the best measurement system for the application.

Suggestions
Below is a quick list of items that you should consider when determining if host-based processing is suitable for an application. Knowing these items can help you calculate whether the measurement system can fulfill your application requirements.
· Speed of the processor
· Speed and size of the disk is, if writing directly to disk
· Amount of RAM memory in the computer
· Amount of data being transferred
· Sampling rate (speed at which data are being sent to the processor)
· Amount of data to be displayed during acquisition
· Other tasks being executed simultaneously

Quick Answer

What if there is not enough time to make all the necessary calculations to determine if the application requires dedicated processing or host-based processing? Here are some quick questions to ask that can give you initial guidance in making this decision.
· Are you replacing an existing stand-alone instrument? If yes, select dedicated processing.
· Does the application require absolute real-time or gap-free performance? If yes, dedicated processing is the safest solution, however host-based processing can often fulfill these requirements if the appropriate conditions are met.
· Is cost a large issue? Host-based processing is typically less expensive per channel.
· Do you need custom signal analysis? Host-based processing is typically the only configuration for custom signal analysis. Note that the NI 455x can also stream unprocessed data to disk for custom analysis on the host processor.
· Do you need multiple device synchronization? Using the National Instruments solutions, host-based configurations are the best solution.
· What measurements do you need? Currently, with NI products, swept sine and zoom FFT analysis are only possible with the NI 455x Series dynamic signal analyzers. Host-based processing with National Instruments products provides a wide range of analysis. Consult the catalog for a listing of the measurements provided in LabVIEW, Measurement Studio, the Sound and Vibration Toolset, and the Signal Processing Suite.

Reader Comments | Submit a comment »