More apple madness

[mvk]

The CPU affinity stuff has been on every box I have tried over the past 20 years, EXCEPT for mac os x. The process scheduler makes it very difficult to do any sort of performance measuring or testing since it doesn't understand the simple idea of one thread per physical core unless there are more threads than physical cores.

Last time I measured this with my program, OSX did it right: it first used the even logical cores, then added the odd ones when the load became larger than the number of physical cores.

[bob]

The CPU affinity stuff has been on every box I have tried over the past 20 years, EXCEPT for mac os x. The process scheduler makes it very difficult to do any sort of performance measuring or testing since it doesn't understand the simple idea of one thread per physical core unless there are more threads than physical cores.

Last time I measured this with my program, OSX did it right: it first used the even logical cores, then added the odd ones when the load became larger than the number of physical cores.

Not happening on my dual-core i7. two threads bounce around from core to core with no concept of physical vs logical. On this box, processors 0 and 1 are physical core 0, and 2 and 3 are physical core 2. I would expect to see 0 and 2 or 1 and 3, but it is pretty evenly balanced, 01, 02, 03, 12, 13 and 23. We have linux on another almost identical macbook and it does it right.

[mvk]

The CPU affinity stuff has been on every box I have tried over the past 20 years, EXCEPT for mac os x. The process scheduler makes it very difficult to do any sort of performance measuring or testing since it doesn't understand the simple idea of one thread per physical core unless there are more threads than physical cores.

Last time I measured this with my program, OSX did it right: it first used the even logical cores, then added the odd ones when the load became larger than the number of physical cores.

Not happening on my dual-core i7. two threads bounce around from core to core with no concept of physical vs logical. On this box, processors 0 and 1 are physical core 0, and 2 and 3 are physical core 2. I would expect to see 0 and 2 or 1 and 3, but it is pretty evenly balanced, 01, 02, 03, 12, 13 and 23. We have linux on another almost identical macbook and it does it right.

FWIW: I observed that on a quad core i7-3720QM 2.6GHz-3.6GHz (8 logical cores). But it was a while ago: end of 2012.

[bob]

The CPU affinity stuff has been on every box I have tried over the past 20 years, EXCEPT for mac os x. The process scheduler makes it very difficult to do any sort of performance measuring or testing since it doesn't understand the simple idea of one thread per physical core unless there are more threads than physical cores.

Last time I measured this with my program, OSX did it right: it first used the even logical cores, then added the odd ones when the load became larger than the number of physical cores.

Not happening on my dual-core i7. two threads bounce around from core to core with no concept of physical vs logical. On this box, processors 0 and 1 are physical core 0, and 2 and 3 are physical core 2. I would expect to see 0 and 2 or 1 and 3, but it is pretty evenly balanced, 01, 02, 03, 12, 13 and 23. We have linux on another almost identical macbook and it does it right.

FWIW: I observed that on a quad core i7-3720QM 2.6GHz-3.6GHz (8 logical cores). But it was a while ago: end of 2012.

Just got a new iMac i7 quad for my 2nd office machine. Interestingly, so far it seems to handle hyper-threading OK. But not my i7 dual macbook. Totally broken.