- Logs messages in a compact binary format
- Fast
- Hundreds of millions of logs per second
- Average latency of 1-3 ns for basic data types
- See benchmarks
- Provides an unpacker to deflate the log messages
- Uses fmtlib to format the logs
- Synchronous logging - not thread safe
- Header-only library
- Single header file version available here
- Requires C 20
- MIT License
The following code logs 1 billion integers to file.
#include <binary_log/binary_log.hpp>
int main()
{
binary_log::binary_log log("log.out");
for (int i = 0; i < 1E9; i)
BINARY_LOG(log, "Hello logger, msg number: {}", i);
}
On a modern workstation desktop, the above code executes in ~2s
.
Type | Value |
---|---|
Time Taken | 1.935 s |
Throughput | 2.06 Gb/s |
Performance | 516 million logs/s |
Average Latency | 1.73 ns |
File Size | ~4 GB |
foo@bar:~/dev/binary_log$ time ./build/examples/billion_integers/billion_integers
real 0m1.935s
user 0m0.906s
sys 0m1.000s
foo@bar:~/dev/binary_log$ ls -lart log.out*
-rw-r--r-- 1 pranav pranav 10 Sep 20 11:46 log.out.runlength
-rw-r--r-- 1 pranav pranav 33 Sep 20 11:46 log.out.index
-rw-r--r-- 1 pranav pranav 4000000002 Sep 20 11:46 log.out
These binary log files can be deflated using the provided unpacker app:
foo@bar:~/dev/binary_log$ time ./build/tools/unpacker/unpacker log.out > log.deflated
real 2m19.853s
user 1m16.078s
sys 0m50.969s
foo@bar:~/dev/binary_log$ ls -lart log.deflated
-rw-r--r-- 1 pranav pranav 35888888890 Dec 6 08:09 log.deflated
foo@bar:~/dev/binary_log$ wc -l log.deflated
1000000000 log.deflated
foo@bar:~/dev/binary_log$ $ head log.deflated
Hello logger, msg number: 0
Hello logger, msg number: 1
Hello logger, msg number: 2
Hello logger, msg number: 3
Hello logger, msg number: 4
Hello logger, msg number: 5
Hello logger, msg number: 6
Hello logger, msg number: 7
Hello logger, msg number: 8
Hello logger, msg number: 9
foo@bar:~/dev/binary_log$ tail log.deflated
Hello logger, msg number: 999999990
Hello logger, msg number: 999999991
Hello logger, msg number: 999999992
Hello logger, msg number: 999999993
Hello logger, msg number: 999999994
Hello logger, msg number: 999999995
Hello logger, msg number: 999999996
Hello logger, msg number: 999999997
Hello logger, msg number: 999999998
Hello logger, msg number: 999999999
Type | Value |
---|---|
Time Taken | 2m 19s |
Throughput | 258 MB/s |
Original File Size | ~5 GB |
Deflated File Size | ~35 GB |
Log Compression | 7x |
See benchmarks section for more performance metrics.
- Implement a single-threaded, synchronous logger - Do not provide thread safety
- If the user wants multi-threaded behavior, the user can choose and implement their own queueing solution
- There are numerous well-known lock-free queues available for this purpose (moody::concurrentqueue, atomic_queue etc.) - let the user choose the technology they want to use.
- The latency of enqueuing into a lock-free queue is large enough to matter
- Users who do not care about multi-threaded scenarios should not suffer the cost
- Looking at the atomic_queue benchmarks, the average round-trip latency across many state-of-the-art multi-producer, multi-consumer queues, to send and receive a 4-byte integer (between 2 threads, using 2 queues) is around 150-250 ns.
- Avoid writing static information more than once
- Examples of static information: the format string, the number of format args, and type of each format arg
- Store the static information in an "index" file
- Store the dynamic information in the log file (refer to the index file where possible)
- Do as little work as possible in the runtime hot path
- No formatting of any kind
- All formatting will happen offline using an unpacker that deflates the binary logs
binary_log
splits the logging into three files:
- Index file contains all the static information from the logs, e.g., format string, number of args, type of each arg etc.
- If a format argument is marked as constant using
binary_log::constant
, the value of the arg is also stored in the index file
- If a format argument is marked as constant using
- Log file contains two pieces of information per log call:
- An index into the index table (in the index file) to know which format string was used
- If runlength encoding is working, this index might not be written, instead the final runlength will be written to the runlengths file
- The value of each argument
- An index into the index table (in the index file) to know which format string was used
- Runlength file contains runlengths - If a log call is made 5 times, this information is stored here (instead of storing the index 5 times in the log file)
- NOTE: Runlengths are only stored if the runlength > 1 (to avoid the inflation case with RLE)
One can specify a log format argument as a constant by wrapping the value with binary_log::constant(...)
. When this is detected, the value is stored in the index file instead of the log file as it is now considered "static information" and does not change between calls.
for (auto i = 0; i < 1E9; i) {
BINARY_LOG(log, "Joystick {}: x_min={}, x_max={}, y_min={}, y_max={}",
binary_log::constant("Nintendo Joycon"),
binary_log::constant(-0.6),
binary_log::constant( 0.65),
binary_log::constant(-0.54),
binary_log::constant( 0.71));
}
The above loop runs in under 500 ms
. The final output is compact at just 118 bytes
and contains all the information needed to deflate the log (if needed).
File | Size |
---|---|
log.out | 1 byte |
log.out.runlength | 6 bytes |
log.out.index | 111 bytes |
foo@bar:~/dev/binary_log$ ls -lart log.out*
-rw-r--r-- 1 pranav pranav 6 Dec 5 08:41 log.out.runlength
-rw-r--r-- 1 pranav pranav 111 Dec 5 08:41 log.out.index
-rw-r--r-- 1 pranav pranav 1 Dec 5 08:41 log.out
foo@bar:~/dev/binary_log$ hexdump -C log.out.index
00000000 33 4a 6f 79 73 74 69 63 6b 20 7b 7d 3a 20 78 5f |3Joystick {}: x_|
00000010 6d 69 6e 3d 7b 7d 2c 20 78 5f 6d 61 78 3d 7b 7d |min={}, x_max={}|
00000020 2c 20 79 5f 6d 69 6e 3d 7b 7d 2c 20 79 5f 6d 61 |, y_min={}, y_ma|
00000030 78 3d 7b 7d 05 0c 0b 0b 0b 0b 01 0f 4e 69 6e 74 |x={}........Nint|
00000040 65 6e 64 6f 20 4a 6f 79 63 6f 6e 01 33 33 33 33 |endo Joycon.3333|
00000050 33 33 e3 bf 01 cd cc cc cc cc cc e4 3f 01 48 e1 |33..........?.H.|
00000060 7a 14 ae 47 e1 bf 01 b8 1e 85 eb 51 b8 e6 3f |z..G.......Q..?|
0000006f
Type | Value |
---|---|
Processor | 11th Gen Intel(R) Core(TM) i9-11900KF @ 3.50GHz 3.50 GHz |
Installed RAM | 32.0 GB (31.9 GB usable) |
SSD | ADATA SX8200PNP |
OS | Ubuntu 20.04 LTS running on WSL in Windows 11 |
C Compiler | g (Ubuntu 10.3.0-1ubuntu1~20.04) 10.3.0 |
foo@bar:~/dev/binary_log$ ./build/benchmark/binary_log_benchmark
2022-09-20T12:59:39-05:00
Running ./build/benchmark/binary_log_benchmark
Run on (16 X 3504 MHz CPU s)
Load Average: 0.52, 0.58, 0.59
------------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
------------------------------------------------------------------------------------------------------------------
BM_binary_log_static_integer<uint8_t>/42 1.22 ns 1.20 ns 560000000 Latency=1.19978ns Logs/s=833.488M/s
BM_binary_log_static_integer<uint16_t>/395 1.43 ns 1.43 ns 448000000 Latency=1.42997ns Logs/s=699.317M/s
BM_binary_log_static_integer<uint32_t>/3123456789 1.89 ns 1.84 ns 373333333 Latency=1.84152ns Logs/s=543.03M/s
BM_binary_log_static_integer<uint64_t>/9876543123456789 5.45 ns 2.76 ns 248888889 Latency=2.76228ns Logs/s=362.02M/s
BM_binary_log_static_integer<int8_t>/-42 1.25 ns 1.26 ns 560000000 Latency=1.25558ns Logs/s=796.444M/s
BM_binary_log_static_integer<int16_t>/-395 1.54 ns 1.57 ns 448000000 Latency=1.56948ns Logs/s=637.156M/s
BM_binary_log_static_integer<int32_t>/-123456789 1.94 ns 1.97 ns 373333333 Latency=1.96708ns Logs/s=508.369M/s
BM_binary_log_static_integer<int64_t>/-9876543123456789 4.11 ns 2.92 ns 235789474 Latency=2.91574ns Logs/s=342.967M/s
BM_binary_log_static_float 1.82 ns 1.84 ns 407272727 Latency=1.84152ns Logs/s=543.03M/s
BM_binary_log_static_double 3.29 ns 2.73 ns 263529412 Latency=2.7274ns Logs/s=366.65M/s
BM_binary_log_static_string 4.93 ns 2.92 ns 235789474 Latency=2.91574ns Logs/s=342.967M/s
BM_binary_log_random_integer<uint8_t> 5.75 ns 5.72 ns 112000000 Latency=5.71987ns Logs/s=174.829M/s
BM_binary_log_random_integer<uint16_t> 6.08 ns 6.14 ns 112000000 Latency=6.13839ns Logs/s=162.909M/s
BM_binary_log_random_integer<uint32_t> 7.51 ns 7.67 ns 89600000 Latency=7.67299ns Logs/s=130.327M/s
BM_binary_log_random_integer<uint64_t> 15.0 ns 15.0 ns 44800000 Latency=14.9972ns Logs/s=66.6791M/s
BM_binary_log_random_integer<int8_t> 5.70 ns 5.72 ns 112000000 Latency=5.71987ns Logs/s=174.829M/s
BM_binary_log_random_integer<int16_t> 5.84 ns 5.86 ns 112000000 Latency=5.85938ns Logs/s=170.667M/s
BM_binary_log_random_integer<int32_t> 7.89 ns 7.67 ns 89600000 Latency=7.67299ns Logs/s=130.327M/s
BM_binary_log_random_integer<int64_t> 14.9 ns 15.0 ns 44800000 Latency=14.9972ns Logs/s=66.6791M/s
BM_binary_log_random_real<float> 6.29 ns 6.25 ns 100000000 Latency=6.25ns Logs/s=160M/s
BM_binary_log_random_real<double> 11.6 ns 11.7 ns 64000000 Latency=11.7188ns Logs/s=85.3333M/s
BM_binary_log_billion_integers 2320246800 ns 1765625000 ns 1 Latency=1.76562ns Logs/s=566.372M/s
binary_log
supports a limited number of types of format arguments. They are:
bool,
char,
uint8_t, uint16_t, uint32_t, uint64_t
int8_t, int16_t, int32_t, int64_t,
float, double,
const char*,
std::string,
std::string_view
See the BUILDING document.
python3 utils/amalgamate/amalgamate.py -c single_include.json -s .
See the CONTRIBUTING document.
The project is available under the MIT license.