increase SIZE to work with 8-128 CPUs (in 8 CPU step size)

[brian90013]

Thank you for this repo. I use it as a sanity check when running on a new platform. Once it helped me detect a misconfigured server where the DIMMs were not in the optimal slots.

I recently obtained a server with 24 cores (48 threads) and the OpenMP SIZE assertion failed. This MR updates SIZE to work with all core counts from 8, 16, 24...112, 120, 128.

[hholst80]

@brian90013 Explain the rationale here? The new SIZE is (1*SIZE_OF_GB + 32504*1024LL). What is this residual term and how did you determine it?

[brian90013]

I want to support CPU counts from 8 to 128 so making SIZE divisible by LCM(8, 16...120, 128) will take care of assert(SIZE % omp_get_max_threads() == 0);. The code mentions "Have PAGE_SIZE buffering" so I also want SIZE % PAGE_SIZE == 0. A final LCM(8, 16...120, 128, 4096) = 184504320. Dividing that by 1024 * 1024 * 1024 = 5.8. Rounding to 6 yields 6 * 184504320 = 1107025920 which is my new proposed SIZE.

[hholst80]

I am not convinced that this should be a compile time constant. It would be better to allocate 1 GB and then cut down instead and set the size dynamically depending on the number of cores allocated to the test.

Let

10**3 >= chunksize * cores --> chunksize = floor(10**3 / cores).

Then we use that as the effective buffer size:

buffersize = floor(10**3 / cores) * cores

The buffer can stay at 1GB and we just use as much as we can to avoid the modulo assert.

This would work fine with also a odd processor count like 22 or even 21.

EDIT: I made an oversight that the chunksize must be divisible by PAGE_SIZE. I fished that in the PR.