[llvm-dev] Efficient Green Thread Context-Switching
Joshua Thomas Wise via llvm-dev
llvm-dev at lists.llvm.org
Fri Mar 27 12:58:03 PDT 2020
Sorry, I certainly didn't mean to be dishonest. I was just repeating one of the comparisons given by the research paper. Regardless, even setjmp() uses a structure of 148 bytes in size (on my machine). The main point is that with compiler support, many context switches can be easily reduced to just a few instructions and only 8 byte of memory, and even in the worst case they will outperform setjmp() by a factor of 2. That’s very significant for modern servers which could expect millions of contexts to exist simultaneously. It would also really help small IOT processors to multitask, which is something they’re normally bad at. I think the use-cases and efficiency achieved is compelling (especially as computing becomes more and more parallel/concurrent).
> On Mar 27, 2020, at 2:29 PM, Joerg Sonnenberger via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> On Fri, Mar 27, 2020 at 01:51:48PM -0500, Joshua Thomas Wise via llvm-dev wrote:
>> The paper shows the results of many benchmarks, all of which show this
>> model heavily outperforming existing models of context switching, in
>> both speed and memory usage. For example, it mentions that the POSIX
>> functions makecontext() and swapcontext() save a structure that’s 936
>> bytes in size (768 on my machine), but their model generates context
>> switches that save only 8-80 bytes of CPU state, depending on which
>> existing optimizations are able to apply to the specific context switch.
>> In some cases, the entire context switch is reduced to 1 or 2 instructions.
> Take a look at the the setjmp/longjmp intrinsics. IMO it is dishonest to
> compare with makecontext/swapcontext, which generally have to save the
> full FPU state and that's where the majority of the data is.
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
More information about the llvm-dev