Timing analysis
Deadlines identified per function, with latency budgets allocated across interrupt, scheduling and processing time.
T · [CLIENT TO PROVIDE] E · [CLIENT TO PROVIDE]
Overview
A system that responds in 200 microseconds almost always, and 40 milliseconds occasionally, is not a fast system — it is an unpredictable one. Real-time engineering is about bounding that worst case and proving the bound.
We start from the timing requirements: what must happen, within what deadline, triggered by what event, and what happens if the deadline is missed. Those become task boundaries, priority assignments and interrupt design rather than implementation details discovered later.
Shared resources, blocking calls, priority inversion, interrupt duration and stack usage are all examined deliberately, because each is a route to timing behaviour that only shows up under load.
Engineering principles
Being correct late is the same as being wrong.
The real-time constraint
A system that answers in 200 microseconds almost always, and 40 milliseconds occasionally, is not fast — it is unpredictable.
Real-time scope
Each of these decisions changes the worst-case behaviour of the system.
Deadlines identified per function, with latency budgets allocated across interrupt, scheduling and processing time.
Task decomposition and priority assignment derived from the deadlines rather than from how the code happens to be organised.
Short, bounded interrupt handlers with deferred processing, so a long handler cannot delay a critical response.
Shared resource access designed to avoid priority inversion, deadlock and unbounded blocking.
Static allocation where possible, stack sizing analysed rather than guessed, and overflow detection built in.
Measurement of real response times under representative and worst-case load, with results recorded as evidence.
Real-time process
Timing treated as a specified, verified property of the system.
Events, deadlines, tolerances and the consequence of a missed deadline captured per function.
Kernel chosen for the platform, the timing requirements, the toolchain and the long-term support picture.
Decomposition into tasks with priorities derived from deadlines, and interaction between them made explicit.
Handler duration bounded, work deferred appropriately and interrupt priorities aligned with the timing model.
Firmware built against the architecture, with blocking behaviour and resource sharing controlled.
Response times, jitter, CPU load and stack usage measured on hardware under load.
Architecture adjusted where measurement disagrees with the model, then verified again.
Evidence
Analysis guides the architecture; measurement decides whether it worked. Response time, jitter, interrupt duration, CPU load and stack usage are all recorded on real hardware under representative and worst-case load.
Where measurement disagrees with the model, the architecture changes.
Applications
Some products tolerate variable timing. These generally do not.
FAQ
Where real-time projects usually succeed or come unstuck.
It depends on how many independent timed activities exist and how they interact. A single control loop with a couple of interrupts often does not need a kernel. Several concurrent activities with different deadlines usually do.
By measuring on hardware: response times, jitter, interrupt duration, CPU load and stack usage under representative and worst-case load. Analysis guides the design; measurement provides the evidence.
Often. It starts with measuring what the system actually does, which frequently contradicts what it was assumed to do. From there the fix may be interrupt redesign, priority changes or restructuring a blocking path.
The kernel is selected per project against platform support, timing needs, toolchain and lifecycle. Specific vendor partnerships or certified kernel usage: [CLIENT TO PROVIDE].
Send us the deadlines and the events that trigger them. We will tell you what the architecture needs to look like.