Question regarding pagesize definition

[green-br]

I am trying to understand whether OpenBLAS handles systems with different pagesize and how to change it if different to any values in header files. I see in common_arm64.h that #define PAGESIZE ( 4 << 10) which is 4kb. It is common to see larger page sizes on Arm such as 64kb. Any thoughts how best to configure the larger pagesize? Is it a matter of defining this on the command line with other options using Make? How would it work with Cmake? At the moment I see:

OpenBLAS : munmap failed:: Invalid argument
error code=22,	release->address=0x40024a171000

When OpenBLAS is called from Open3d in Python on Arm.

[martin-frbg]

Good question, I'm not aware of this having come up before, but I'd expect it to work if you use either make PAGESIZE=65536 or cmake -DPAGESIZE=65536 as appropriate

[martin-frbg]

Actually for CMake you'd need to use cmake -DCCOMMON_OPT="-DPAGESIZE=65536" as the PAGESIZE variable itself is unknown to the CMakeLists.
OTOH I am curious if this is actually the root cause of what you are seeing, as at least on the one Arm system with 64k pagesize that I have access to, both 4k and 64k PAGESIZE builds of OpenBLAS appear to work fine.

[green-br]

Hi - I agree - not encountered issues with OpenBLAS regarding pagesize before but hit similar mmap issues previously with pagesize differences so thought I would check. It seems to be coming from openblas that comes with a version of numpy e.g. python3.10/site-packages/numpy/core/../../numpy.libs/libopenblas64_p-r0-17488984.3.23.dev.so if I use LD_DEBUG=libs when running the python code. The error occurs when the Python code exits.

[martin-frbg]

Hmm, that's some rather old Numpy and OpenBLAS, I recall some people reporting wild (and irreproducible) behaviour with that particular binary package (#4365).

[giordano]

Any chance of detecting the pagesize during init? Hardcoding the value at compile time would be unfortunate for downstream packagers who would like to provide generic builds.

[imciner2]

In Julia, we call sysconf(_SC_PAGESIZE) to get the pagesize information at runtime (https://github.com/JuliaLang/julia/blob/514adf882c911c84487feb054a7a0d3d90fce3ba/src/sys.c#L588). The thing I don't know though is how the HUGE_PAGESIZE would need to work in the case the other pagesize comes dynamically at runtime.

[martin-frbg]

I guess one could parse /proc/meminfo for HUGE_PAGESIZE where available, but I'll need to look into if/how/where the two are interrelated in OpenBLAS. But at this point I'd prefer to remain unconvinced that PAGESIZE is actually a problem...