Why spdlog-rs is so fast?

[nneesshh]

It is wonderful work, but about the benchmark:
"spdlog-rs (0.3.0)
Sync

[info] **********************************************************************
[info] Multi threaded: 1 threads, 250000 messages
[info] **********************************************************************
[info] basic_mt Elapsed: 0.09 secs 2881439/sec
[info] rotating_mt Elapsed: 0.09 secs 2810747/sec
[info] daily_mt Elapsed: 0.09 secs 2829965/sec
[info] level-off Elapsed: 0.00 secs 407365162/sec
...

C++ spdlog (1.11.0)
Sync

[info] **************************************************************
[info] Multi threaded: 1 threads, 250,000 messages
[info] **************************************************************
[info] basic_mt Elapsed: 0.13 secs 1,906,872/sec
[info] rotating_mt Elapsed: 0.13 secs 1,866,094/sec
[info] daily_mt Elapsed: 0.13 secs 1,981,681/sec
[info] level-off Elapsed: 0.00 secs 147,466,525/sec
...
"

It is amazing that spdlog-rs is much faster than spdlog-c++, how cound it be?

[SpriteOvO]

I'm not so quite remember of the details, but spdlog-rs is not "much" faster than C++ spdlog in some cases.

As you can see, the flexible-string version of spdlog-rs and fmt version of C++ spdlog from the benchmark results:

spdlog-rs (with flexible-string feature):

[info] **********************************************************************
[info] Multi threaded: 1 threads, 250000 messages
[info] **********************************************************************
[info] basic_mt                       Elapsed: 0.08 secs          2969371/sec
[info] rotating_mt                    Elapsed: 0.08 secs          2952681/sec
[info] daily_mt                       Elapsed: 0.08 secs          2992041/sec
[info] level-off                      Elapsed: 0.00 secs        514721021/sec
[info] **********************************************************************
[info] Multi threaded: 4 threads, 250000 messages
[info] **********************************************************************
[info] basic_mt                       Elapsed: 0.07 secs          3496977/sec
[info] rotating_mt                    Elapsed: 0.07 secs          3399459/sec
[info] daily_mt                       Elapsed: 0.07 secs          3364660/sec
[info] level-off                      Elapsed: 0.00 secs        632751202/sec

C++ spdlog (with fmt library):

[info] **************************************************************
[info] Multi threaded: 1 threads, 250,000 messages
[info] **************************************************************
[info] basic_mt                       Elapsed: 0.07 secs        3,430,498/sec
[info] rotating_mt                    Elapsed: 0.07 secs        3,481,225/sec
[info] daily_mt                       Elapsed: 0.07 secs        3,563,268/sec
[info] level-off                      Elapsed: 0.00 secs      138,757,839/sec
[info] **************************************************************
[info] Multi threaded: 4 threads, 250,000 messages
[info] **************************************************************
[info] basic_mt                       Elapsed: 0.11 secs        2,214,753/sec
[info] rotating_mt                    Elapsed: 0.12 secs        2,059,991/sec
[info] daily_mt                       Elapsed: 0.12 secs        2,058,245/sec
[info] level-off                      Elapsed: 0.00 secs      139,938,427/sec

Both of them are much faster than the corresponding standard version, the reasons are the same. fmt library provides an on-stack growable buffer type basic_memory_buffer for formatting string storage (used here), flexible-string provides a FlexibleString as well (used here).

For the on-stack string storage version, C++ spdlog is a little faster than spdlog-rs for _mt sink with a single thread. For the performance degradation of C++ spdlog in 4 threads, I guess it's caused by some concurrent data structure implementation, but it's just a guess, I didn't verify it.

For your question, why the standard version of spdlog-rs is much faster than C++ spdlog, if I remember correctly, the main reason is spdlog-rs do a heap pre-reservation (here) for formatting string storage, but C++ spdlog do not. A reservation can significantly improve performance by reducing fragmented heap allocation when formatting records.