1 |
1026 |
ivang |
@c
|
2 |
|
|
@c COPYRIGHT (c) 1988-2002.
|
3 |
|
|
@c On-Line Applications Research Corporation (OAR).
|
4 |
|
|
@c All rights reserved.
|
5 |
|
|
@c
|
6 |
|
|
@c timeFORCE386.t,v 1.10 2002/01/17 21:47:46 joel Exp
|
7 |
|
|
@c
|
8 |
|
|
|
9 |
|
|
@include common/timemac.texi
|
10 |
|
|
@tex
|
11 |
|
|
\global\advance \smallskipamount by -4pt
|
12 |
|
|
@end tex
|
13 |
|
|
|
14 |
|
|
@chapter CPU386 Timing Data
|
15 |
|
|
|
16 |
|
|
@section Introduction
|
17 |
|
|
|
18 |
|
|
The timing data for the i386 version of RTEMS is
|
19 |
|
|
provided along with the target dependent aspects concerning the
|
20 |
|
|
gathering of the timing data. The hardware platform used to
|
21 |
|
|
gather the times is described to give the reader a better
|
22 |
|
|
understanding of each directive time provided. Also, provided
|
23 |
|
|
is a description of the interrupt latency and the context
|
24 |
|
|
switch times as they pertain to the i386 version of RTEMS.
|
25 |
|
|
|
26 |
|
|
@section Hardware Platform
|
27 |
|
|
|
28 |
|
|
All times reported except for the maximum period
|
29 |
|
|
interrupts are disabled by RTEMS were measured using a Force
|
30 |
|
|
Computers CPU386 board. The CPU386 is a 16 Mhz board with zero
|
31 |
|
|
wait state dynamic memory and an i80387 numeric coprocessor.
|
32 |
|
|
One of the count-down timers provided by a Motorola MC68901 was
|
33 |
|
|
used to measure elapsed time with one microsecond resolution.
|
34 |
|
|
All sources of hardware interrupts are disabled, although the
|
35 |
|
|
interrupt level of the i386 allows all interrupts.
|
36 |
|
|
|
37 |
|
|
The maximum period interrupts are disabled was
|
38 |
|
|
measured by summing the number of CPU cycles required by each
|
39 |
|
|
assembly language instruction executed while interrupts were
|
40 |
|
|
disabled. Zero wait state memory was assumed. The total CPU
|
41 |
|
|
cycles executed with interrupts disabled, including the
|
42 |
|
|
instructions to disable and enable interrupts, was divided by 16
|
43 |
|
|
to simulate a i386 executing at 16 Mhz.
|
44 |
|
|
|
45 |
|
|
@section Interrupt Latency
|
46 |
|
|
|
47 |
|
|
The maximum period with interrupts disabled within
|
48 |
|
|
RTEMS is less than RTEMS_MAXIMUM_DISABLE_PERIOD microseconds
|
49 |
|
|
including the instructions
|
50 |
|
|
which disable and re-enable interrupts. The time required for
|
51 |
|
|
the i386 to generate an interrupt using the int instruction,
|
52 |
|
|
vectoring to an interrupt handler, and for the RTEMS entry
|
53 |
|
|
overhead before invoking the user's interrupt handler are a
|
54 |
|
|
total of 12 microseconds. These combine to yield a worst case
|
55 |
|
|
interrupt latency of less
|
56 |
|
|
RTEMS_MAXIMUM_DISABLE_PERIOD + RTEMS_INTR_ENTRY_RETURNS_TO_PREEMPTING_TASK
|
57 |
|
|
microseconds. [NOTE: The
|
58 |
|
|
maximum period with interrupts disabled within RTEMS was last
|
59 |
|
|
calculated for Release RTEMS_RELEASE_FOR_MAXIMUM_DISABLE_PERIOD.]
|
60 |
|
|
|
61 |
|
|
It should be noted again that the maximum period with
|
62 |
|
|
interrupts disabled within RTEMS is hand-timed. The interrupt
|
63 |
|
|
vector and entry overhead time was generated on the Force
|
64 |
|
|
Computers CPU386 benchmark platform using the int instruction as
|
65 |
|
|
the interrupt source.
|
66 |
|
|
|
67 |
|
|
@section Context Switch
|
68 |
|
|
|
69 |
|
|
The RTEMS processor context switch time is RTEMS_NO_FP_CONTEXTS
|
70 |
|
|
microseconds on the Force Computers CPU386 benchmark platform.
|
71 |
|
|
This time represents the raw context switch time with no user
|
72 |
|
|
extensions configured. Additional execution time is required
|
73 |
|
|
when a TASK_SWITCH user extension is configured. The use of the
|
74 |
|
|
TASK_SWITCH extension is application dependent. Thus, its
|
75 |
|
|
execution time is not considered part of the base context switch
|
76 |
|
|
time.
|
77 |
|
|
|
78 |
|
|
Since RTEMS was designed specifically for embedded
|
79 |
|
|
missile applications which are floating point intensive, the
|
80 |
|
|
executive is optimized to avoid unnecessarily saving and
|
81 |
|
|
restoring the state of the numeric coprocessor. The state of
|
82 |
|
|
the numeric coprocessor is only saved when a FLOATING_POINT task
|
83 |
|
|
is dispatched and that task was not the last task to utilize the
|
84 |
|
|
coprocessor. In a system with only one FLOATING_POINT task, the
|
85 |
|
|
state of the numeric coprocessor will never be saved or
|
86 |
|
|
restored. When the first FLOATING_POINT task is dispatched,
|
87 |
|
|
RTEMS does not need to save the current state of the numeric
|
88 |
|
|
coprocessor.
|
89 |
|
|
|
90 |
|
|
The exact amount of time required to save and restore
|
91 |
|
|
floating point context is dependent on the state of the numeric
|
92 |
|
|
coprocessor. RTEMS places the coprocessor in the initialized
|
93 |
|
|
state when a task is started or restarted. Once the task has
|
94 |
|
|
utilized the coprocessor, it is in the idle state when floating
|
95 |
|
|
point instructions are not executing and the busy state when
|
96 |
|
|
floating point instructions are executing. The state of the
|
97 |
|
|
coprocessor is task specific.
|
98 |
|
|
|
99 |
|
|
The following table summarizes the context switch
|
100 |
|
|
times for the Force Computers CPU386 benchmark platform:
|
101 |
|
|
|