gerhobbelt / nanolog2

Low Latency C++11 Logging Library

NanoLog

• Low Latency C++11 Logging Library.
• It's fast. Very fast. See Latency benchmark
• NanoLog only uses standard headers so it should work with any C++11 compliant compiler.
• Supports typical logger features namely multiple log levels, log file rolling and asynchronous writing to file.

Design highlights

• Zero copying of string literals.
• Lazy conversion of integers and doubles to ascii.
• No heap memory allocation for log lines representable in less than ~256 bytes.
• Minimalistic header includes. Avoids common pattern of header only library. Helps in compilation times of projects.

Guaranteed and Non Guaranteed logging

• Nanolog supports Guaranteed logging i.e. log messages are never dropped even at extreme logging rates.
• Nanolog also supports Non Guaranteed logging. Uses a ring buffer to hold log lines. In case of extreme logging rate when the ring gets full (i.e. the consumer thread cannot pop items fast enough), the previous log line in the slot will be dropped. Does not block producer even if the ring buffer is full.

Usage

#include "NanoLog.hpp"

int main()
{
  // Ensure initialize is called once prior to logging.
  // This will create log files like /tmp/nanolog1.txt, /tmp/nanolog2.txt etc.
  // Log will roll to the next file after every 1MB.
  // This will initialize the guaranteed logger.
  nanolog::initialize(nanolog::GuaranteedLogger(), "/tmp/", "nanolog", 1);
  
  // Or if you want to use the non guaranteed logger -
  // ring_buffer_size_mb - LogLines are pushed into a mpsc ring buffer whose size
  // is determined by this parameter. Since each LogLine is 256 bytes,
  // ring_buffer_size = ring_buffer_size_mb * 1024 * 1024 / 256
  // In this example ring_buffer_size_mb = 3.
  // nanolog::initialize(nanolog::NonGuaranteedLogger(3), "/tmp/", "nanolog", 1);
  
  for (int i = 0; i < 5; ++i)
  {
    LOG_INFO << "Sample NanoLog: " << i;
  }
  
  // Change log level at run-time.
  nanolog::set_log_level(nanolog::LogLevel::CRIT);
  LOG_WARN << "This log line will not be logged since we are at loglevel = CRIT";
  
  return 0;
}

Latency benchmark of Guaranteed logger

• A google search for fast logger C++ gives the first result spdlog
• There's an interesting article on worst case latency by the author of g3log
• So let's benchmark NanoLog vs spdlog vs g3log vs reckless.
• Take a look at nano_vs_spdlog_vs_g3log_vs_reckless.cpp
• Benchmark was compiled with g++ 4.8.4 running Linux Mint 17 on Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz

Thread Count 1 - percentile latency numbers in microseconds (lower number means better performance)

Logger	50th	75th	90th	99th	99.9th	Worst	Average
nanolog_guaranteed	0	1	1	4	8	68	0.347930
spdlog	3	3	3	5	11	129	2.588590
g3log	5	6	6	10	19	186	5.206230
reckless	0	0	1	1	175	1861	1.829760

Thread Count 2 - percentile latency numbers in microseconds (lower number means better performance)

Logger	50th	75th	90th	99th	99.9th	Worst	Average
nanolog_guaranteed	0	1	1	2	5	55	0.457240
nanolog_guaranteed	0	1	1	2	5	81	0.459090
spdlog	2	3	3	3	5	25	2.449580
spdlog	2	3	3	3	6	21	2.457150
g3log	4	5	6	12	18	64	4.574850
g3log	4	5	6	12	20	84	4.586590
reckless	0	1	1	11	417	1592	4.412750
reckless	0	1	1	12	417	2138	4.427810

Thread Count 3 - percentile latency numbers in microseconds (lower number means better performance)

Logger	50th	75th	90th	99th	99.9th	Worst	Average
nanolog_guaranteed	0	1	1	3	6	91	0.450700
nanolog_guaranteed	0	1	2	3	7	90	0.676050
nanolog_guaranteed	0	1	2	3	7	262	0.680430
spdlog	2	2	2	4	6	6729	1.803570
spdlog	3	3	3	5	8	25	2.679420
spdlog	3	3	3	5	10	50	2.685230
g3log	4	4	6	17	27	53	4.385530
g3log	4	4	6	16	26	55	4.435680
g3log	6	7	8	19	29	1031	5.896250
reckless	1	1	1	298	1643	3070	11.208420
reckless	1	1	1	382	2266	3006	12.310360
reckless	1	1	1	167	2839	3249	12.754520

Thread Count 4 - percentile latency numbers in microseconds (lower number means better performance)

Logger	50th	75th	90th	99th	99.9th	Worst	Average
nanolog_guaranteed	0	1	2	3	6	53	0.582140
nanolog_guaranteed	0	1	2	3	7	70	0.608980
nanolog_guaranteed	0	1	2	3	7	62	0.803630
nanolog_guaranteed	0	1	2	3	7	61	0.797270
spdlog	2	2	2	3	5	40	1.767930
spdlog	2	2	2	3	6	21	1.768640
spdlog	3	3	3	4	8	24	2.676170
spdlog	3	3	3	5	10	31	2.698580
g3log	4	4	5	17	30	7766	4.620760
g3log	6	7	9	21	35	8478	6.368940
g3log	6	7	8	22	32	1327	7.023880
g3log	7	8	9	23	36	8470	7.831750
reckless	1	1	1	506	3477	9224	18.959310
reckless	1	1	1	479	3636	8471	19.181160
reckless	1	1	1	530	2990	11658	19.245110
reckless	1	1	1	436	3641	8626	19.342780

Latency benchmark of Non guaranteed logger

• Take a look at non_guaranteed_nanolog_benchmark.cpp for the code used to generate the latency numbers.
• Benchmark was compiled with g++ 4.8.4 running Linux Mint 17 on Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz

Thread count: 1
	Average NanoLog Latency = 131 nanoseconds
Thread count: 2
	Average NanoLog Latency = 182 nanoseconds
	Average NanoLog Latency = 272 nanoseconds
Thread count: 3
	Average NanoLog Latency = 216 nanoseconds
	Average NanoLog Latency = 209 nanoseconds
	Average NanoLog Latency = 315 nanoseconds
Thread count: 4
	Average NanoLog Latency = 229 nanoseconds
	Average NanoLog Latency = 221 nanoseconds
	Average NanoLog Latency = 233 nanoseconds
	Average NanoLog Latency = 332 nanoseconds
Thread count: 5
	Average NanoLog Latency = 247 nanoseconds
	Average NanoLog Latency = 240 nanoseconds
	Average NanoLog Latency = 320 nanoseconds
	Average NanoLog Latency = 345 nanoseconds
	Average NanoLog Latency = 383 nanoseconds

Crash handling

• g3log has support for crash handling. I do not see the point in re-inventing the wheel. Have a look at that what's done there and if it works for you, give Kjell credit and use his crash handling code.

Tips to make it faster!

• NanoLog uses standard library chrono timestamps. Your platform / os may have non-standard but faster timestamps. Use them!