Real-time systems - timing
The term "real-time" is bandied about a great deal these days. You will hear about the "real-time web" and "real-time systems". But what exactly does "real-time system" mean though?
A simple definition serves as a starting point:
A real-time system is one in which time is a key consideration.
This serves us well as a simple definition, but as nearly all systems care about time to a lesser or greater extent, the definition can be refined depending on the context. For example, a chat program has a different set of timing characteristics compared to a car Anti-lock Braking System (ABS), and the ABS has different timing requirements compared to a video streaming application.
The actual scale of time of interest depends on the application. In some systems it may be in the milli- or micro-second timescale, as for example in a missile guidance system or car ABS. In other systems timescales may be measured in seconds, hours, or even days.
With time as a key aspect of a real-time system, there are at least three parameters of timing to consider:
- Response time (time to process an event)
- Determinism (predictability)
- Deadlines (hard, firm, soft)
Each of these will be explored in the following sections.
Response time can be defined differently depending on the context. In this series of articles it is taken to have the following definition unless otherwise stated:
Response time is the time it takes for the system to process an event to completion.
So, what is an event? An event is just something happening. It could be an event in the real world, such as a user hitting an emergency stop button, pressing a brake, or perhaps a child running into the road. It could be the arrival of a network packet. Usually that event will need to be processed in some way, at the very least it is necessary to record that the event happened.
What is completion? This depends on the event and how it is best processed, but it is possible to illustrate by example.
Consider the case of a hardware interrupt, such as a keyboard keypress. This generates a hardware interrupt which causes an interrupt service routine (ISR) to handle the interrupt. The ISR may read a hardware register to obtain the keyboard scan code and place the code in a buffer. In this example the event is considered complete to on the return of the ISR. At this point it is confirmed that the key event has been recorded at least at the lower level of the system - critically the key event has not been lost. If the keyboard ISR was itself interrupted by another ISR, and possibly other nested ISRs, the event could not be considered to have been recorded until the keyboard ISR itself returns. At this point the event processing has completed.
In the case of a web browser, clicking a link to retrieve some result may initiate an HTTP request which arrives at the server at some time later after taking into account the latency of the network. The server may perform some calculation or a database lookup and then send an HTTP response, which will again be subject to delays in the network. Finally, the response arrives back in the client, where the web browser displays the result. The event is now considered to have completed.
Consider one more example. An application performs an asynchronous (AIO) filesystem read. A callback handler is written in our application. The asynchronous read is called, registering the appropriate callback handler and perhaps specifying a buffer into which data can be placed. The read will not block the app (it's a non-blocking asynchronous call), so the read method returns immediately and the app then most likely sleeps. The operating system code runs, disks spin, read/write heads move ponderously about, and the requested bytes are read from the disk at some point. The OS signals completion by invoking the registered callback handler and the data is safely stored in the buffer. The operation has completed when the callback handler returns.
There are no hard and fast rules here, but to be able to measure response time it is necessary to be clear about what is being measured.
Now consider again a subsequent filesystem read where the bytes requested are already cached. The response time will be a lot less because the mechanical "gubbins" does not need to be spun. However, completion is still when the callback handler returns.
This last example also illustrates the variable nature of response times for the same type of event. It is not known in advance, at the application level, whether the data is cached or not.
Consider the response time of the ABS event. If in one test it was 10 ms and then the next time that event occurred the response time was 500ms, this may indicate a problem - even a potentially dangerous situation.
That brings us to the question of consistency or predictability in the timing constraints of the system.
In this series of articles the level of predictabiity is referred to as the determinism of the system.
Consider an ABS event where a response time of between 5 and 15 milliseconds is acceptable. On testing the response times obtained are 5ms, 5ms, 5ms, 5ms, ..., 5ms. The response time is within the desired range and is very predictable (deterministic).
With a different ABS design the following results set might be obtained: 5ms, 10ms, 15ms, 6ms, 14ms. These results are still within acceptable limits, but are much less predictable.
Now consider the following results (in ms): 16, 17, 16, 16, 17, 16, 16BACKSLASH. These results are fairly predictable, but outside the desired range (the deadline of 15ms has been exceeded).
Now consider the following results (in ms): 5, 500, 4, 2, 700, 300, 250, 15, 177BACKSLASH. These results not only fall outside the required range on occasion, but also appear quite unpredictable.
The degree of predictability is called the determinism of a system.
Determinism in a system is really a measure of how accurately response times can be predicted. Deterministic system are quite stable in their response times. Non-deterministic systems tend to have unpredictable response times. It is important to understand the levels of determinism in a system when designing applications with real-time constraints.
Looking at the ABS again, why might things be unpredictable? Well, it may have been something like a flurry of hardware events interrupts occurred at approximately the same time, delaying processing.
In the real world it is often not possible to predict if and when certain events will happen. Perhaps the engine management circuitry is already busy processing another interrupt such as a timer interrupt and cannot be interrupted, this represents an unpredictable delay even before the incoming interrupt is processed.
This problem of unpredictable real-time events might be accute in a system such as a self-driving car. Here you may have many real-world events happening all at once due to what is happening on the road. A car strays from its lane, a child steps into the road, the GPS updates, the LIDAR generates new data that needs processing, and so on.
The ABS ISR may itself be interrupted by another interrupt causing another Interrupt Service Routine (ISR) to run, which itself was interrupted causing another ISR to run and so on in a nested fashion. This results in unpredictability or lack of determinism.
The solution to the lack of determinism in the case of the ABS ISR may be to make it the highest priority interrupt, and so it cannot by interrupted until its code has completed. Note that if another ISR was running when the ABS interrupt occurred, the other ISR would be interrupted by the higher priority ABS ISR. This might bring the predictability back to within our timing constraints, as exceeding the 15ms deadline may be unacceptable.
This is now getting to the heart of designing effective real-time systems - the system designer cannot predict if and when real-world events will occur, but she can design the system in such a way that it processes these events, whenever they happen, with predictable response times and within appropriate time constraints.
A deadline is a timing constraint that must not be exceeded.
Real-time systems can be broadly categorized depending on the gravity of outcomes should the system fail to meet timing deadlines:
In hard real-time systems this could lead to a hazardous situation and potentially a loss of life. Examples of hard real-time systems include nuclear reactors, missile guidance systems, robots, ABS, self-driving cars, and aircraft flight control systems.
Firm real-time systems have less disastrous outcomes if deadlines are exceeded. An example might be a trade is not carried out in an algorithmic trading system, resulting in a loss of income.
In soft real-time systems a deadline exceeded might be reflected as a mild annoyance, such as a voice packet being missed, a game dropping a frame, or a annoyingly long database response time for a specific query.
Note that in some systems a response time might be micro or millseconds but these are soft deadlines. On the other hand a system such as a nuclear reactor might have response times of five seconds for some processes but these are deemed quite acceptable. On the other hand the same system might have a deadline of ten seconds that is a hard limit. It is the outcome of failing to meet a deadline that is the defining characteristic.
A real-time system is any system where timing is a key consideration.
Specifically three timing-related characteristics (at least) should be considered when designing a real-time system:
- Response time (the time to process an event to completion)
- Determinism (predictibility of response times)
- Deadline (timing constraints that must not be exceeded)
More broadly a real-time system can be considered to be a system in which real world events occur with a degree of unpredictability on the time axis, and where predictabililty of response times is desirable.
In the next article in this series I will take a quick look at networks and protocols in the context of real-time systems.