This product is available via the open source license described at the end of this document.
EPICS Input/Output Controller
core (IOCcore) software has been ported to several
different operating systems (
Three real-time parameters
are measured: interrupt, context switch, and total response latency. On Linux,
more detailed interrupt latencies are measured: interrupt top half to bottom
half, and interrupt bottom half to user space interrupt service routine.
To implement the tests,
several software components were developed. In order to port to other operating
systems or hardware platforms, only one component has to be implemented.
Interrupt latency is the time from interrupt generation until the
interrupt service routine starts executing. Linux has an additional concept of
interrupt “top half” and “bottom half”. The top half
interrupt handler, which executes with interrupts disabled, should complete
quickly. If more work is required it can schedule a bottom half handler. On
both vxWorks and RTEMS, the bottom half is replaced
by a high priority thread. The VME support on Linux provides an additional
feature, a user-level interrupt handler. A user-level handler is registered to
receive a VME interrupt. POSIX signals are used to call the user-level handler.
On Linux all of the following latencies are measured: top half, top half to
bottom half, and bottom half to user-level Interrupt Service Routine (ISR).
Context switch latency is the time from one context switching to
another. The context switch time from a kernel space interrupt to user space-thread
context is measured in this paper.
Total response latency is the time from an interrupt to when a waiting
user thread begins execution.
The test IOC is heavily loaded, as described below, in all the tests.
Four different values of each parameter are collected: minimum, median,
maximum, and percentage of samples over some value. To make the results
comparable, most tests were conducted on a private network for one hour. To
look for network interference, some tests were run for 16 hours on a public
network. Another test was run to measure user-level interrupt latency. Tables 1–5
show the results.
All latencies are in microseconds. For vxWorks,
two results are shown. One has the task priority lower than the vxWorks network task. The other has the task priority higher
than the vxWorks network task.
The results show that vxWorks and RTEMS have excellent
real-time performance. The average performance of Linux 2.6.13 is similar to
Linux 2.4.30, but the maximum total response latency is greatly reduced.
The recent Linux 2.6 kernel looks promising for use in real-time
applications. It might have better performance if the real-time preemptive patch is applied. No big differences have been
seen between the low-latency preemptive kernel and the
non-preemptive kernel of Linux 2.6 at this time.
Switching context involved saving the CPU’s registers and loading
a new state, flushing the caches, and changing the virtual memory mapping. Context
switch latency is highly architecture dependent and different hardware may get
different results.
If these tests were repeated or if the tests lasted for a longer time,
the maximum value of each parameter could be higher.
On Linux all
of the latency measurements are of interest. The application real-time
requirements determine how interrupts should be handled. At one extreme, user-level
interrupt handling will be sufficient. At the other extreme, the top-level
interrupt handler will have to process most interrupts and only involve the
bottom half and user threads at appropriate times.
Reconfigure the envPaths or cdCommands
and st.cmd file in iocBoot/iocsampler/
and start the IOC. Run EPICS client medm GUI: runmedm
Set "samples per set".
Click the "startOnce"
button to collect one set of samples
Click the menu botton
to select the period time to collect samples.
To look at the details of each real time
parameter configure the corresponding "X cutoff" or "Y
cutoff".
Run performCaget and performCaput,
on the host, issue the command:
<top>/bin/<host>/performCaget 5000 1
<top>/bin/<host>/performCaput 5000 5
Porting
to Other Platform
There are three platform dependent files rtemsSampler.c,
linuxSampler.c and vxworksSampler.c
in sampler/samplerApp/src/, which are for RTEMS, Linux
and vxWorks respectively.
To measure the performance on other platform, a platform dependent file must be developed. It must be able to generate an interrupt and provide a microsecond resolution timestamp.
Reference
"REAL-TIME PERFORMANCE MEASUREMENTS
OF EPICS IOCCORE", ICALEPCS 2005,