3 ms·
We still face a related class of problem today. The x86 PAUSE has wildly varying throughput. On most Intel parts it is 1/8 or so, but on Skylake Xeon it's 1/141
by jeffbee 6y ago
We still face a related class of problem today. The x86 PAUSE has wildly varying throughput. On most Intel parts it is 1/8 or so, but on Skylake Xeon it's 1/141. On Ryzen its 1/3. I've seen code that makes assumptions about how much real time must have passed based on PAUSE loops.
- acqq 6y ago> I've seen code that makes assumptions about how much real time must have passed based on PAUSE loops. Note: here, the PAUSE instruction is not the problem at all, but the "code that makes assumptions." Because the "seen" code is not named, I assume it's something internal for some company?
- jeffbee 6y agoYes, private code. Basically there was some mutex fairness thing that was written on a SKX and on a Zen CPU where PAUSE is 50x faster it didn't have good fairness, it was too tight.