On Wed, 29 Aug 2018 at 16:08, Qais Yousef <qais.you...@arm.com> wrote: >
> You beat me to it, I got confused by smt_gain recently when I stumbled > on it as I found out on ARM it's not used and had to spend sometime to > convince myself it's not really necessary to use it. > > It was hard to track the history of this and *why* it's needed. > > The only 'theoretical' case I found smt_gain can be useful is when you > have asymmetric system, for example: > > Node_A: 1 Core 2 Threads > Node_B: 1 Core 4 Threads > > Then with smt_gain the group_capacity at the core level will be limited > to smt_gain. But without smt_gain Node_B will look twice as powerful as > Node_A - which will affect balancing AFAICT causing Node_B's single core > to be oversubscribed as the 4 threads will still have to share the same > underlying hardware resources. I don't think in practice such systems > exists, or even make sense, though. Yes I came to the same conclusion > > So +1 from my side for the removal. Thanks > > > -- > Qais Yousef >