135 lines
6.5 KiB
Plaintext
135 lines
6.5 KiB
Plaintext
/*
|
|
ChibiOS/RT - Copyright (C) 2006-2007 Giovanni Di Sirio.
|
|
|
|
This file is part of ChibiOS/RT.
|
|
|
|
ChibiOS/RT is free software; you can redistribute it and/or modify
|
|
it under the terms of the GNU General Public License as published by
|
|
the Free Software Foundation; either version 3 of the License, or
|
|
(at your option) any later version.
|
|
|
|
ChibiOS/RT is distributed in the hope that it will be useful,
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
GNU General Public License for more details.
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
along with this program. If not, see <http://www.gnu.org/licenses/>.
|
|
*/
|
|
|
|
/**
|
|
* @page article_jitter Response Time and Jitter
|
|
* @{
|
|
* Response time jitter is one of the most sneaky source of problems when
|
|
* designing a real time system. When using a RTOS like ChibiOS/RT one must
|
|
* be aware of what the jitter is and how it can affect the performance of the
|
|
* system. A good place to start is this
|
|
* <a href="http://en.wikipedia.org/wiki/Jitter">Wikipedia article</a>.
|
|
*
|
|
* <h2>Interrupt handlers execution time</h2>
|
|
* The total execution time of an interrupt handler includes:
|
|
* - Hardware interrupts latency, this parameter is pretty much fixed and
|
|
* characteristic of the system.
|
|
* - Fixed handler overhead, as example registers stacking/unstacking.
|
|
* - Interrupt specific handler code execution time, as example, in a serial
|
|
* driver, this is the time used by the handler to transfer data from/to
|
|
* the UART.
|
|
* - OS overhead. Any operating system requires to run some extra code
|
|
* in interrupt handlers in order to handle correct preemption and Context
|
|
* Switching.
|
|
* .
|
|
* <h2>Interrupt Response Time</h2>
|
|
* The Interrupt Response Time is the time from an interrupt event and the
|
|
* execution of the handler code. Unfortunately this time is not constant
|
|
* in most cases, see the following graph:
|
|
*
|
|
* @dot
|
|
digraph example {
|
|
rankdir="LR";
|
|
node [shape=circle, fontname=Helvetica, fontsize=8, fixedsize="true", width="0.75", height="0.75"];
|
|
edge [fontname=Helvetica, fontsize=8];
|
|
int [label="Interrupt"];
|
|
busy [label="Busy"];
|
|
served [label="Interrupt\nServed"];
|
|
int -> served [label="Not Busy (minimum latency)"];
|
|
int -> busy [label="Not Ready"];
|
|
busy -> busy [label="Still Busy\n(added latency)"];
|
|
busy -> served [label="Finally Ready"];
|
|
* @enddot
|
|
*
|
|
* In this scenario the jitter (busy state) is represented by the sum of:
|
|
* - Higher or equal priority interrupt handlers execution time combined.
|
|
* This time can go from zero to the maximum randomly. This value can be
|
|
* guaranteed to be zero only if the interrupt has the highest priority in
|
|
* the system.
|
|
* - Highest execution time among lower priority handlers. This value is zero
|
|
* on those architectures (Cortex-M3 as example) where interrupt handlers
|
|
* can be preempted by higher priority sources.
|
|
* - Longest time in a kernel lock zone that can delay interrupt servicing.
|
|
* This value is zero for fast interrupt sources, see @ref system_states.
|
|
* .
|
|
* <h2>Threads Flyback Time</h2>
|
|
* This is the time between an event, as example an interrupt, and the
|
|
* execution of the thread that will process it. Imagine the following
|
|
* graph as the continuation of the previous one.
|
|
*
|
|
* @dot
|
|
digraph example {
|
|
rankdir="LR";
|
|
node [shape=circle, fontname=Helvetica, fontsize=8, fixedsize="true", width="0.75", height="0.75"];
|
|
edge [fontname=Helvetica, fontsize=8];
|
|
served [label="Interrupt\nServed"];
|
|
busy [label="Busy"];
|
|
thread [label="Thread\nAwakened"];
|
|
served -> busy [label="Not highest Priority"];
|
|
busy -> busy [label="Higher priority Threads\n(added latency)"];
|
|
busy -> thread [label="Highest Priority"];
|
|
served -> thread [label="Highest Priority (minimum latency)"];
|
|
* @enddot
|
|
*
|
|
* In this scenario all the jitter sources previously discussed are also
|
|
* present and there is the added jitter caused by the activity of the
|
|
* higher priority threads.
|
|
*
|
|
* <h2>Jitter Mitigation</h2>
|
|
* For each of the previously described jitter sources there are possible
|
|
* mitigation actions.
|
|
*
|
|
* <h3>Interrupt handlers optimization</h3>
|
|
* An obvious mitigation action is to optimize the interrupt handler code as
|
|
* much as possible for speed.<br>
|
|
* Complex actions should never be performed in interrupt handlers.
|
|
* An handler should serve the interrupt and wakeup a dedicated thread in order
|
|
* to handle the bulk of the work.<br>
|
|
* Another possible mitigation action is to evaluate if a specific interrupt
|
|
* handler really needs to interact with the OS, if the handler uses full
|
|
* stand-alone code then it is possible to remove the OS related overhead.<br>
|
|
*
|
|
* <h3>Kernel lock zones</h3>
|
|
* The OS kernel protects some critical internal data structure by disabling
|
|
* (fully in simple architecture, to some extent in more advanced
|
|
* microcontrollers) the interrupt sources. Because of this the kernel itself
|
|
* is a jitter cause, a good OS design minimizes the jitter generated by the
|
|
* kernel by using adequate data structures, algorithms and coding
|
|
* practices.<br>
|
|
* A good OS design is not the whole story, some OS primitives may generate
|
|
* more or less jitter depending on the system state, as example the maximum
|
|
* number of threads on a certain queue, the maximum number of nested mutexes
|
|
* and so on. Some algorithms employed internally can have constant execution
|
|
* time but others may have linear execution time or be even more complex.
|
|
*
|
|
* <h3>Higher priority threads activity</h3>
|
|
* At thread level, the response time is affected by the interrupt-related
|
|
* jitter but mainly by the activity of the higher priority threads and
|
|
* contention on protected resources.<br>
|
|
* It is possible to improve the system overall response time and reduce jitter
|
|
* by carefully assigning priorities to the various threads and carefully
|
|
* designing mutual exclusion zones.<br>
|
|
* The use of the proper synchronization mechanism (semaphores, mutexes, events,
|
|
* messages and so on) also helps to improve the overall system performance.
|
|
* The use of the Priority Inheritance algorithm implemented in the mutexes
|
|
* subsystem can improve the overall response time and reduce jitter but it is
|
|
* not a magic wand, a proper system design comes first.
|
|
*/
|
|
/** @} */
|