memory latency SIZE overflow #2
Comments
|
That's interesting. This tool had been primarily developed for use on the embedded ARM/MIPS systems to spot the oddities/misconfiguration in the memory subsystem. I never imagined that anyone would run it on big servers :) About the SIZE variable limitation. There is one more problem to solve: currently 32-bit http://en.wikipedia.org/wiki/Linear_congruential_generator is used for generating random offsets inside the buffer. This also needs to be reworked a bit. I'll try to see what can be done. |
|
looking at the wikipedia article of the it looks like the current constants ( 1103515245 and 12345 ) can just be replaced by 6364136223846793005 and 1442695040888963407 respectively. thanks for the awesome tool by the way. In large server systems memory access is not uniform and I want to use your tool measure the latency between different nodes in the system. |
* coopy SSE2 assembly to avx2 no code changes yet * remove non-AVX archs remove i386 and win32 * update aligned_block_copy update aligned_block_copy to AVX2 * convert aligned_block_copy_nt to AVX2 * convert aligned_block_copy_pf32 to AVX2 * block_copy_pf64 and block_copy_nt_pf32 * convert aligned_block_copy_nt_pf64 * convert aligned_block_fill * convert aligned_block_fill_nt * define __X86_AVX2_H__ * delete movsd and movsb functions These don't seem to have a place in AVX2 routines * update Makefile with x86-avx2 * add avx2 benchmarks (incomplete) Still needs to: - properly check AVX2 support - run both SSE2 and AVX2 if AVX2 supported * avx2 benchmarks include sse2 If AVX2, append the SSE2 benchmarks to the AVX2 benchmarks and run all of them * simplify array sizing calcs * separate AVX2 bi from SSE2 * update copyright notice * define a constant for the register size * add AVX512 assembly code * replave vmovdqa with vmovdqa64 (?) * add AVX512 benchmarks * use getopt_long to add args to run SSE/AVX tests Co-authored-by: Joel Luth <joel.luth@emc.com> Co-authored-by: root <root@mpl037a.west.isilon.com>
I have a system with lots of ram (1.5TB) and I want to use this tool to test the memory latency when accessing different NUMA nodes. unfortunately the block SIZE variable overflows as it is stored in a integer instead of a 64bit value.
The text was updated successfully, but these errors were encountered: