[Question] Why does the total value decrease when the "_mi_stats_print" function is executed for each TLD?

[Lee-HyeongSeok]

Hi, thank you for your work on this project.

In our proxy server project, which uses mimalloc as the memory manager, we maintain statistics for each TLD using a dedicated structure when the server process runs. We have implemented an interface function that allows us to print stats for each TLD using _mi_stats_print, as shown below. However, as you can see in the attached screenshot, we have observed that the value of "total" decreases when this is done.

void mi_custom_thread_stats_print_out(mi_stats_t* tid_stat, mi_output_fun* out, void* arg) mi_attr_noexcept {
  _mi_stats_print(tid_stat, out, arg);
}

This issue did not occur with mimalloc version 2.1.2, but started happening after upgrading to version 3.0.8. Do you have any idea what might be causing this?
For reference, the screenshot only shows the output for the “pssv#1” thread, filtered from the _mi_stats_print results using grep.

[Lee-HyeongSeok]

Additionally, at the timing of mi_process_init in init.c, the tld structure of heap_main manages the stats information via a pointer.
At the timing of _mi_thread_heap_init, the stats information of the currently created TLD is being managed via a pointer.

Is it the intention of mimalloc 3 development to periodically merge the statistics of heap_main (defined in init.c) into the statistics of subproc_main?

[daanx]

Hi -- yes, you are right, in v3 the thread-local stats are periodically merged into the main stats.
It is best to not use v3.0.x versions -- those are meant for internal use. The latest v3 is best to use (now v3.2.7 but I hope to update later this week). In the v3.2.x versions there are now real first-class heaps that can allocate from any thread:

• each mi_heap_t now consists of many thread-local heaps (mi_theap_t)
• the stats in each theap are added periodically into their owning heap, and then the stats in the thread-local theap are reset (and this is why the total is now decreasing for an individual theap).
• it is best to only show/get stats for the owning heap as those are the correct counts (well, at some point in the past as not all threads add stats only periodically for performance). The new API's are in include/mimalloc-stats.h. Probably, using mi_stats_get or mi_stats_print_out will suffice?

Let me know if this works for you.

ps. mimalloc v2 essentially also works the same except it merges the thread local stats in the main stats. However, the merging is not periodic but only at thread termination or if called by the user (mi_stats_merge).

[Lee-HyeongSeok]

Thank you for the detailed explanation.
As you mentioned, I confirmed that memory reset is performed after merging at the theap part.
When I used mi_stats_print_out, I was able to see the overall statistics of the process printed correctly.
Thank you.