By: Jon Masters (jcm.delete@this.jonmasters.org), November 19, 2020 12:31 am
Room: Moderated Discussions
Maynard Handley (name99.delete@this.name99.org) on November 18, 2020 4:46 pm wrote:
> There is a right way and a wrong way of doing things. The wrong way, in this case, is treating
> SMT as a fake core, rather than something like a "co-routine acceleration engine".
Yes, exposing it as a fake core for marketing reasons was always a poor choice. It had the benefit of quickly being enabled by software (that then had to be educated about it to not actively hurt performance) but this went on far too long. When the cloud happened, folks handed out vCPUs (aka 1.3 cores per 2 "core" VM) and the number of people using the cloud who actually even know these are often HT threads within a shared single core is probably a very shockingly small number. In that respect, I actually thought the SMT side-channels in 2018/19 that drew attention to this were net positive for the overall industry.
Jon.
> There is a right way and a wrong way of doing things. The wrong way, in this case, is treating
> SMT as a fake core, rather than something like a "co-routine acceleration engine".
Yes, exposing it as a fake core for marketing reasons was always a poor choice. It had the benefit of quickly being enabled by software (that then had to be educated about it to not actively hurt performance) but this went on far too long. When the cloud happened, folks handed out vCPUs (aka 1.3 cores per 2 "core" VM) and the number of people using the cloud who actually even know these are often HT threads within a shared single core is probably a very shockingly small number. In that respect, I actually thought the SMT side-channels in 2018/19 that drew attention to this were net positive for the overall industry.
Jon.
