|
| | ~ScorePLogger () |
| |
| bool | getLogMachineName () const |
| | Is public as some analysis frameworks check explicitly whether these features are switched on.
|
| |
| bool | getLogThreadName () const |
| | Is public as some analysis frameworks check explicitly whether these features are switched on.
|
| |
| bool | getLogTrace () const |
| | Is public as some analysis frameworks check explicitly whether these features are switched on.
|
| |
| bool | getLogTimeStamp () const |
| | Is public as some analysis frameworks check explicitly whether these features are switched on.
|
| |
| void | debug (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| |
| void | info (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| |
| void | warning (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| | Write Warning.
|
| |
| void | error (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| | Write Error.
|
| |
| void | traceIn (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| |
| void | traceOut (long int timestampNanoseconds, int rank, int threadId, const std::string &trace, const std::string &message) |
| |
| void | indent (bool indent, const std::string &trace, const std::string &message) |
| | Tells the logger to increment/decrement the indent.
|
| |
| void | close () |
| |
Command Line Logger.
Standard log output device. Implements the LogHandle. Usually there's only one instance of this class for only the Log type aggregates it in a static way. For a detailed description of this compiler flags please browse through the static consts of the class or the howto page.
An interesting issue is the behaviour of the command line logger in the parallel mode if debug is switched on. In this case the number of messages soon becomes unreadable for any human. Thus, the command line logger internally creates one output file rank.log for each rank and writes all the information into this file. info, warnings and errors are written (in short form) to std::cout / std::cerr, too.
If the debug mode is not set, all information is written to std::cout.
Note, that if multithreading is switched on (TBB-flag), the access to std::cout or the logfile respectively is synchronized. The other methods are not thread-safe and should be called only in safe context for defined behaviour.
- Version
- Revision
- 1.19
- Author
- Tobias Weinzierl, Wolfgang Eckhardt
Definition at line 54 of file ScorePLogger.h.
| std::string tarch::logging::ScorePLogger::constructMessageString |
( |
std::string | messageType, |
|
|
long int | timestampNanoseconds, |
|
|
int | rank, |
|
|
int | threadId, |
|
|
const std::string & | trace, |
|
|
const std::string & | message ) |
|
private |
Construct message string.
!!! Thread Safety
The message string relies on the global field _indent. This one might change throughout the execution of this method. However, I accept such a behavior: Changing _indent throughout the message execution makes the method add the wrong number of whitespaces in front of the message. That is a 'bug' we can accept.
To satisfy Intel Inspector et al at least slightly, I copy over _indent before I actually construct the message string. So the indent can't change while we add the spaces/tabs to the output.
| void tarch::logging::ScorePLogger::indent |
( |
bool | indent, |
|
|
const std::string & | trace, |
|
|
const std::string & | message ) |
Tells the logger to increment/decrement the indent.
!!! Thread Safety
_indent is a global static field shared by all threads. If we increment or decrement it, this is first of all a read followed by a write. Consequently data races could occur and the counter could become smaller than zero. This ain't possible in the sequential code as each increment is accompanied by a decrement. The following table illustrates the race:
|| value of _indent || Thread 1 || Thread 2 | 2 | initial condition | initial condition | | enter indent(false) | enter indent(true) | | fetch indent into register | fetch indent into register | | register value -= 2 | register value += 2 | 4 | is a little bit slower | write back new value of indent | 0 | write back new value of indent |
To avoid this data race, I introduced a semaphore. This one could also be implemented with TBB's atomic construct, e.g., but I prefer the semaphor / critical section technique.
- Parameters
-
| trace | Needed in debug mode to be able to find out who called indent(false) without an indent(true) |
| message | Needed in debug mode to be able to find out who called indent(false) without an indent(true) |