PlatformLab / NanoLog

Nanolog is an extremely performant nanosecond scale logging system for C++ that exposes a simple printf-like API.
Other
3k stars 343 forks source link
high-performance logging logging-library low-latency printf

NanoLog

Nanolog is an extremely performant nanosecond scale logging system for C++ that exposes a simple printf-like API and achieves over 80 million logs/second at a median latency of just over 7 nanoseconds.

How it achieves this insane performance is by extracting static log information at compile-time, only logging the dynamic components in runtime hotpath, and deferring formatting to an offline process. This basically shifts work out of the runtime and into the compilation and post-execution phases.

For more information about the techniques used in this logging system, please refer to either the NanoLog Paper published in the 2018 USENIX Annual Technical Conference or the original author's doctoral thesis.

Performance

This section shows the performance of NanoLog with existing logging systems such as spdlog v1.1.0, Log4j2 v2.8, Boost 1.55, glog v0.3.5, and Windows Event Tracing with Windows Software Trace Preprocessor on Windows 10 (WPP).

Throughput

Maximum throughput measured with 1 million messages logged back to back with no delay and 1-16 logging threads (NanoLog logged 100 million messages to generate a log file of comparable size). ETW is "Event Tracing for Windows." The log messages used can be found in the Log Message Map below. N|Solid

Runtime Latency

Measured in nanoseconds and each cell represents the 50th / 99.9th tail latencies. The log messages used can be found in the Log Message Map below.

Message NanoLog spdlog Log4j2 glog Boost ETW
staticString 7/ 37 214/ 2546 174 / 3364 1198/ 5968 1764/ 3772 161/ 2967
stringConcat 7/ 36 279/ 905 256 / 25087 1212/ 5881 1829/ 5548 191/ 3365
singleInteger 7/ 32 268/ 855 180 / 9305 1242/ 5482 1914/ 5759 167/ 3007
twoIntegers 8/ 62 437/ 1416 183 / 10896 1399/ 6100 2333/ 7235 177/ 3183
singleDouble 8/ 43 585/ 1562 175 / 4351 1983/ 6957 2610/ 7079 165/ 3182
complexFormat 8/ 40 1776/ 5267 202 / 18207 2569/ 8877 3334/ 11038 218/ 3426

Log Messages Map

Log messages used in the benchmarks above. Italics indicate dynamic log arguments.

Message ID Log Message Used
staticString Starting backup replica garbage collector thread
singleInteger Backup storage speeds (min): 181 MB/s read
twoIntegers buffer has consumed 1032024 bytes of extra storage, current allocation: 1016544 bytes
singleDouble Using tombstone ratio balancer with ratio = 0.4
complexFormat Initialized InfUdDriver buffers: 50000 receive buffers (97 MB), 50 transmit buffers (0 MB), took 26.2 ms
stringConcat Opened session with coordinator at basic+udp:host=192.168.1.140,port=12246

Using NanoLog

Prerequisites

Currently NanoLog only works for Linux-based systems and depends on the following:

NanoLog Pipeline

The NanoLog system enables low latency logging by deduplicating static log metadata and outputting the dynamic log data in a binary format. This means that log files produced by NanoLog are in binary and must be passed through a separate decompression program to produce the full, human readable ASCII log.

Compiling NanoLog

There are two versions of NanoLog (Preprocessor version and C++17 version) and you must chose one to use with your application as they’re not interoperable. The biggest difference between the two is that the Preprocessor version requires one to integrate a Python script in their build chain while the C++17 version is closer to a regular library (simply build and link against it). The benefit of using the Preprocessor version is that it performs more work at compile-time, resulting in a slightly more optimized runtime.

If you don’t know which one to use, go with C++17 NanoLog as it’s easier to use.

C++17 NanoLog

The C++17 version of NanoLog works like a traditional library; just #include "NanoLogCpp17.h" and link against the NanoLog library. A sample application can be found in the sample directory.

To build the C++17 NanoLog Runtime library, go in the runtime directory and invoke make. This will produce ./libNanoLog.a to against link your application and a ./decompressor application that can be used to re-inflate the binary logs.

When you compile your application, be sure to include the NanoLog header directory (-I ./runtime), link against NanoLog, pthreads, and POSIX AIO (-L ./runtime/ -lNanoLog -lrt -pthread), and enable format checking in the compiler (e.g. passing in -Werror=format as a compilation flag). The latter step is incredibly important as format errors may silently corrupt the log file at runtime. Sample g++ invocations can be found in the sample GNUmakefile.

After you compile and run the application, the log file generated can then be passed to the ./decompressor application to generate the full human-readable log file (instructions below).

Preprocessor NanoLog

The Preprocessor version of NanoLog requires a tighter integration with the user build chain and is only for advanced/extreme users.

It requires the user's GNUmakefile to include the NanoLogMakeFrag, declare USR_SRCS and USR_OBJS variables to list all app’s source and object files respectively, and use the pre-defined run-cxx macro to compile ALL the user .cc files into .o files instead of g++. See the preprocessor sample GNUmakefile for more details.

Internally, the run-cxx invocation will run a Python script over the source files and generate library code that is specific to each compilation of the user application. In other words, the compilation builds a version of the NanoLog library that is non-portable, even between compilations of the same application and each make invocation rebuilds this library.

Additionally, the compilation should also generate a ./decompressor executable in the app directory and this can be used to reconstitute the full human-readable log file (instructions below).

Sample Applications

The sample applications are intended as a guide for how users are to interface with the NanoLog library. Users can modify these applications to test NanoLog's various API and functionality. The C++17 and Preprocessor versions of these applications reside in ./sample and ./sample_preprocessor respectively. One can modify main.cc in each directory, build/run the application, and execute the decompressor to examine the results.

Below is an example for C++17 NanoLog's sample application.

cd sample

# Modify the application
nano main.cc

make clean-all
make
./sampleApplication
./decompressor decompress /tmp/logFile

Note: The sample application sets the log file to /tmp/logFile.

NanoLog API

To use the NanoLog system in the code, one just has to include the NanoLog header (either NanoLogCpp17.h for C++17 NanoLog or NanoLog.h for Preprocessor NanoLog) and invoke the NANO_LOG() function in a similar fashion to printf, with the exception of a log level before it. Example below:

#include "NanoLogCpp17.h"
using namespace NanoLog::LogLevels;

int main() 
{
  NANO_LOG(NOTICE, "Hello World! This is an integer %d and a double %lf\r\n", 1, 2.0);
  return 0;
}

Valid log levels are DEBUG, NOTICE, WARNING, and ERROR and the logging level can be set via NanoLog::setLogLevel(...)

The rest of the NanoLog API is documented in the NanoLog.h header file.

Post-Execution Log Decompressor

The execution of the user application should generate a compressed, binary log file (default locations: ./compressedLog or /tmp/logFile). To make the log file human-readable, simply invoke the decompressor application with the log file.

./decompressor decompress ./compressedLog

After building the NanoLog library, the decompressor executable can be found in either the ./runtime directory (for C++17 NanoLog) or the user app directory (for Preprocessor NanoLog).

Unit Tests

The NanoLog project contains a plethora of tests to ensure correctness. Below is a description of each and how to access/build/execute them.

Integration Tests

The integration tests build and test the Nanolog system end-to-end. For both C++17 NanoLog and Preprocessor NanoLog, it compiles a client application with the NanoLog library, executes the application, and runs the resulting log file through the decompressor. It additionally compares the output of the decompressor to ensure that the log contents match the expected result.

One can execute these tests with the following commands:

cd integrationTest
./run.sh

Preprocessor and Library Unit Tests

The NanoLog Library and Preprocessor engine also contain a suit of their own unit tests. These will test the inner-workings of each component by invoking individual functions and checking their returns match the expected results.

To run the NanoLog preprocessor unit tests, execute the following commands:

cd preprocessor
python UnitTests.py

To build and run the NanoLog library unit tests, execute the following commands:

git submodule update --init

cd runtime
make clean
make test
./test --gtest_filter=-*assert*

Note: The gtest filter is used to removed tests with assert death statements in them.