Antoine Pitrou
Antoine Pitrou
@github-actions crossbow submit -g cpp
CI looks good to me. Is this ready for review again @zanmato1984 ?
> Would it address most of your concerns if I replaced the "SHOULD" language, I'll revise to do so and we can hopefully iterate from there. I think it could...
I'm getting a similar error when trying to use it on Apache Arrow's CI. Weirdly, it seems that the problem may be dependent on the exact Python version that `pipx`...
See discussion in https://bugs.python.org/issue33444
You probably don't want 0 for `MALLOC_MMAP_THRESHOLD_` but a round number of 4kB pages (for example 16384 bytes). Smaller allocations should probably still go through the normal allocator.
I think you couid add a documentation section about memory consumption issues and mention those environment variables there. It's not Numpy's or Python's business to make global decisions about the...
From `/usr/include/malloc.h`: ```c /* mallopt options that actually do something */ #define M_TRIM_THRESHOLD -1 #define M_TOP_PAD -2 #define M_MMAP_THRESHOLD -3 #define M_MMAP_MAX -4 #define M_CHECK_ACTION -5 #define M_PERTURB -6 #define...
Interesting. According to the `mallopt` man page: > Arenas are thread safe and therefore may have multiple concurrent memory requests. The trade-off is between the number of threads and the...
One downside of `mmap` is that it can cost more CPU (for example because it always returns zero-initialized memory). That's why the libc is cautious about using it.