When building a real-time system, care must be taken to ensure that the system will be able to perform properly within the constraints of that system. One of these constraints may be how fast certain operations can be performed. Another might be how deterministic the overall behavior of the system is. Lastly the memory footprint (size) and unit cost may be important.
One of the major problems encountered while evaluating a system will be how to compare it with possible alternatives. Most manufacturers of real-time systems publish performance numbers, ostensibly so that users can compare the different offerings. However, what these numbers mean and how they were gathered is often not clear. The values are typically measured on a particular piece of hardware, so in order to truly compare, one must obtain measurements for exactly the same set of hardware that were gathered in a similar fashion.
Two major items need to be present in any given set of measurements. First, the raw values for the various operations; these are typically quite easy to measure and will be available for most systems. Second, the determinacy of the numbers; in other words how much the value might change depending on other factors within the system. This value is affected by a number of factors: how long interrupts might be masked, whether or not the function can be interrupted, even very hardware-specific effects such as cache locality and pipeline usage. It is very difficult to measure the determinacy of any given operation, but that determinacy is fundamentally important to proper overall characterization of a system.
In the discussion and numbers that follow, three key measurements are provided. The first measurement is an estimate of the interrupt latency: this is the length of time from when a hardware interrupt occurs until its Interrupt Service Routine (ISR) is called. The second measurement is an estimate of overall interrupt overhead: this is the length of time average interrupt processing takes, as measured by the real-time clock interrupt (other interrupt sources will certainly take a different amount of time, but this data cannot be easily gathered). The third measurement consists of the timings for the various kernel primitives.
Key operations in the kernel were measured by using a simple test program which exercises the various kernel primitive operations. A hardware timer, normally the one used to drive the real-time clock, was used for these measurements. In most cases this timer can be read with quite high resolution, typically in the range of a few microseconds. For each measurement, the operation was repeated a number of times. Time stamps were obtained directly before and after the operation was performed. The data gathered for the entire set of operations was then analyzed, generating average (mean), maximum and minimum values. The sample variance (a measure of how close most samples are to the mean) was also calculated. The cost of obtaining the real-time clock timer values was also measured, and was subtracted from all other times.
Most kernel functions can be measured separately. In each case, a reasonable number of iterations are performed. Where the test case involves a kernel object, for example creating a task, each iteration is performed on a different object. There is also a set of tests which measures the interactions between multiple tasks and certain kernel primitives. Most functions are tested in such a way as to determine the variations introduced by varying numbers of objects in the system. For example, the mailbox tests measure the cost of a "peek" operation when the mailbox is empty, has a single item, and has multiple items present. In this way, any effects of the state of the object or how many items it contains can be determined.
There are a few things to consider about these measurements. Firstly, they are quite micro in scale and only measure the operation in question. These measurements do not adequately describe how the timings would be perturbed in a real system with multiple interrupting sources. Secondly, the possible aberration incurred by the real-time clock (system heartbeat tick) is explicitly avoided. Virtually all kernel functions have been designed to be interruptible. Thus the times presented are typical, but best case, since any particular function may be interrupted by the clock tick processing. This number is explicitly calculated so that the value may be included in any deadline calculations required by the end user. Lastly, the reported measurements were obtained from a system built with all options at their default values. Kernel instrumentation and asserts are also disabled for these measurements. Any number of configuration options can change the measured results, sometimes quite dramatically. For example, mutexes are using priority inheritance in these measurements. The numbers will change if the system is built with priority inheritance on mutex variables turned off.
The final value that is measured is an estimate of interrupt latency. This particular value is not explicitly calculated in the test program used, but rather by instrumenting the kernel itself. The raw number of timer ticks that elapse between the time the timer generates an interrupt and the start of the timer ISR is kept in the kernel. These values are printed by the test program after all other operations have been tested. Thus this should be a reasonable estimate of the interrupt latency over time.