Reputation: 713
I am using Boost-Log and the global severity logger for my application's logging platform.
Profiling shows that the boost::log::v2s_mt_posix::core::open_record
can take up to
25% of the total execution time.
I do have many log messages BUT I would not expect them to be so costly since they are lower severity and they are filtered.
Is there a way to make those messages not so costly in run-time? (again: I expect to have a small overhead even when filtered and of course a larger when not-filrered).
Compile-time is relatively easy to "fix" this by creating some wrapper macros.
UPDATED with sample working code:
#include <cmath>
#include <iostream>
#include <string>
#include <boost/lexical_cast.hpp>
#include <boost/log/sources/global_logger_storage.hpp>
#include <boost/log/sources/severity_logger.hpp>
#include <boost/log/trivial.hpp>
#include <boost/log/sources/severity_channel_logger.hpp>
#include <boost/log/common.hpp>
#include <boost/log/expressions.hpp>
#include <boost/log/sinks.hpp>
#include <boost/log/utility/setup/common_attributes.hpp>
#include <boost/log/utility/setup/console.hpp>
#include <boost/log/support/date_time.hpp>
#include <boost/utility/empty_deleter.hpp>
#include <boost/shared_ptr.hpp>
#include <boost/make_shared.hpp>
// severity levels
enum severity_level {INFO};
// Logging macro
#define LOG(level) BOOST_LOG_SEV(global_logger::get(), level)
// Initializing global boost::log logger
typedef boost::log::sources::severity_channel_logger_mt<
severity_level, std::string> global_logger_type;
BOOST_LOG_INLINE_GLOBAL_LOGGER_INIT(global_logger, global_logger_type)
{
return global_logger_type(boost::log::keywords::channel = "global_logger");
}
// Initialize terminal logger
void init_logger(int verbosity)
{
namespace bl = boost::log;
typedef bl::sinks::synchronous_sink<bl::sinks::text_ostream_backend>
text_sink;
boost::shared_ptr<text_sink> sink = boost::make_shared<text_sink>();
sink->locked_backend()->add_stream(
boost::shared_ptr<std::ostream>(&std::cout, boost::empty_deleter()));
sink->locked_backend()->auto_flush(true);
sink->set_formatter(bl::expressions::stream << bl::expressions::message);
sink->set_filter(bl::expressions::attr<severity_level>("Severity")
< (severity_level) verbosity);
bl::core::get()->add_sink(sink);
bl::add_common_attributes();
}
int main(int argc, char* argv[])
{
init_logger(boost::lexical_cast<int>(argv[1]));
for(int i = 0; i < 200; ++i)
{
std::sin(std::sin(17.2)); // dummy processing
LOG(INFO) << "A message!";
}
return 0;
}
Running with argument 0
does not print any log messages but it takes twice (!) the
time than commenting out the LOG message.
Upvotes: 10
Views: 2735
Reputation: 33655
One workaround (this is a marmite solution btw.) is to use a macro to wrap the logging, for example:
#define LOG_INFO(x) \
if (<test if log level is INFO>) { \
LOG(INFO) << x; \
}
Then:
LOG_INFO("Foo" << "Bar" << someExpensiveFunction());
You can be sure that the someExpensiveFunction()
will only be executed at a level that supports INFO
.
Heck, you could even go as far as compiling it out completely,
#ifndef _NINFO_
#define LOG_INFO(x) \
if (<test if log level is INFO>) { \
LOG(INFO) << x; \
}
#else
#define LOG_INFO(x)
#endif
Another advantage of this would be that you could change the logging implementation by modifying the macro rather than than changing all your code every where..
Upvotes: 0
Reputation: 1
Often, when logging, what is being logged has to be evaluated, and that evaluation happens before the actual decision whether to log or not is made. For example:
log.debug("The current state of the system is: " + system.expensiveFunctionToGatherState());
We had a home grown logging system that added the current time to the log, which was a call to gettimeofday()
. Turned out 20% of the CPU consumption in the program was calls into gettimeofday()
for log functions that never went anywhere.
Upvotes: 0