"HPHPCHH for short"
Anyway, "all of this monitoring and automated management has a cost in terms of lost cycles. And that overhead, even if only a few percent, is probably unacceptable to the typical HPC customer. Moreover, they’ll definitely balk at paying more for the hardware to support these features, mainly because that money could be better used to buy more/faster CPUs or memory "
Disclaimer: I'm no HPC guy, never touched it. However... I've found the absolute quintessence of tuning anything I've ever sped up (sometimes by orders of magnitude) is feedback, and the relatively low cost of gathering that is multiply repaid in increased speed and/or reduced memory. I'd be amazed if that's different here.
but... <repeats disclaimer>