Mixed instruction set schedulers? Intel 12th gen issues

3 points by doommius 4 years ago | 5 comments

I've been looking a bit at the news recently about the intel 12th gen CPUs, and how they disable the AVX-512 instruction set if the efficiency cores are enabled. which seemed like a weird and poor implementation that bottlenecks some workloads and leaves quite a bit on the table in terms of performance. Why isn't this handled in the scheduler ? all the major operating systems already support the big/small core design. Why isn't the instruction set requirements simply passed to the scheduler via program headers to dynamically schedule tasks in the most efficient/fastest manner?

I read a bit up on the topics and there is already research into the topics. so I wonder why the approach intel used is even there. http://web.cs.ucla.edu/~tjn/papers/pact2018-hybrid-sched.pdf

Why isn't a better solution used? and why just disable the functionally as it seems like low hanging performance gains from intels side? And isn't this a approach to also look into more as modern hardware utilize ASIC style or dedicated hardware for more operations and will so going forward?

ayende 4 years ago |

There is no facility to say "this program runs AVX512 instructions". For that matter, a program may not know that it does. If it uses a JIT, loads a dynamic library, etc.

And if you mess it up, you get an invalid instruction, so that is something that you really want to avoid.

ant6n 4 years ago | |

The invalid instruction can be caught by the kernel, which can inspect the instruction, see that it's avx512, mark the process as avx enabled, and only run it on p cores.

Some sort of settings or policy could be set up so that processes don't force themselves onto the p cores.

doommius 4 years ago | |

Agreed, but there has to be a better solution then just outright disabling the features.

PaulHoule 4 years ago | | |

Intel is not 100% serious about quality and performance.

The Whole AVX1, AVX2, AVX4, …, AVX8192 means the SIMD instructions will only get used by missile labs. Since you have to recompile all your code to use them and because end users will have old machines, devs will either stick to something 10 years old or just skip them entirely.

It is self-defeating for Intel because their greatest competition is not AMD or ARM but sticking with the CPU you already have. Hypothetically SIMD is the major source of performance improvement in new generations but if the software doesn’t support it, it might as well not exist, so why upgrade? If users don’t upgrade why should devs use the new SIMD instructions.

Try to tell them this common sense and it is just ‘talk to the hand’.