Lock contention bottleneck (jemalloc vs tcmalloc) about tcmalloc HOT 2 OPEN

If you sync past 3e01aa2, do you see the same pattern?

Actually I did not see the problem even before 3e01aa2 for allocations <=256KB

But, just out of curiosity decided to check after 3e01aa2 (synced up to 299c1cc):
No significant changes (but maybe user/sys/real times is just too simple).

BTW I've added another size class, that uses frontend allocator for allocations up to 4MB (but I think that 2MB is enough and number of cached pages are just some random numbers so should not be considered anyhow) for testing purpose:

However this does not looks like a long-term way to me.
Any thoughts on this?

from tcmalloc.

If you sync past 3e01aa2, do you see the same pattern?

The 1MB allocations will be problematic, but that should help with <=256KB.

from tcmalloc.