I am using FC32 on a AMD 3970X with HT off. I have a program which maps a System V shm segment using huge pages, then references random locations for 10 seconds to find what the latency of memory is.
The funny thing is, the latency is better when printing to a terminal (xterm) under Xwayland, and worse when redirecting to a file or piping through less: $ test -m sysv2m:shm.test -c 8 -p 100 -x int -o find -n 10 ival tot op/s ns 10.1 510000000 50605422 19.8 $ test -m sysv2m:shm.test -c 8 -p 100 -x int -o find -n 10 | less ival tot op/s ns 10.1 350000000 34621166 28.9 The difference printing to stdout is significant, but it only happens under XWayland in a terminal, not on the console. The above use 8 MB shm of 2M page size. I've looked at the CPU frequency, different nice priorities, tried different cpupower governors, I can't explain this weirdness. This does not occur under CentOS 8, even when using the same binary compiled under FC32. When I use the prefetch instruction, prefetching the next 2 references, a command line option, the performance difference disappears: $ test -m sysv2m:shm.test -c 8 -p 100 -x int -o find -n 10 -f 2 ival tot op/s ns 10.2 550000000 54143996 18.5 $ test -m sysv2m:shm.test -c 8 -p 100 -x int -o find -n 10 -f 2 | less ival tot op/s ns 10.1 550000000 54260358 18.4 _______________________________________________ users mailing list -- users@lists.fedoraproject.org To unsubscribe send an email to users-le...@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/users@lists.fedoraproject.org