David Hess (davidwhess.delete@this.gmail.com) on September 10, 2021 8:39 pm wrote:
> Andrey (andrey.semashev.delete@this.gmail.com) on September 10, 2021 5:12 pm wrote:
> >
> > I don't know exactly how Thread Director works but as I
> > understand it monitors the kind of instructions being
> > executed in a thread. I doubt this can be done purely in software without major performance losses. I mean,
> > the kernel can know that a certain class of instructions have been used, but it can't know how many of
> > them have been used in a given time slice or, say, how much power said instructions consumed.
>
> Wouldn't the number of cycles spent waiting on memory accesses
> be more important than the mix of instructions?

That could be another metric that is being monitored. I wouldn't say it's more important though, as arguably any core would spend roughly the same time waiting for DRAM under the same conditions (and moving the workload to another core would only make it worse due to cold caches). But it would reflect on the amount of power consumed by the core, so if Thread Director monitors power then it would indirectly monitor DRAM stalls.