@jedisct1 How much speedup did you get in recent libsodium due to MAP_POPULATE, presumably for Argon2 & scrypt when memory (de)allocation happens each time? I've just tried it for yescrypt on RHEL6 & 7, getting ~+3% on average (but more stable speeds), up to ~+10% for some sizes.
-
-
Thanks. My tests were also on bare metal. Now I wish we had an equivalent flag for use along with shmget() IPC_CREAT, and ideally also a flag (or more?) to spread the allocation evenly across all NUMA nodes vs. keep it local to the current NUMA node (vs. indicate no preference?)
-
Just found out that MAP_POPULATE results in a 7% regression for yescrypt at 2 MiB in multi-threaded benchmarks on RHEL6 when the memory allocation would be reused anyway - thus, where MAP_POPULATE was expected to neither help nor hurt. Now think how to skip it when it's unneeded.
- 6 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.