Date: Thu, 29 Jan 2026 12:27:13 -0500 From: "Drew Gallatin" <gallatin@freebsd.org> To: "Olivier Certner" <olce@freebsd.org>, "Gleb Smirnoff" <glebius@freebsd.org> Cc: freebsd-current@freebsd.org, src-committers@freebsd.org, "ShengYi Hung" <aokblast@freebsd.org> Subject: Re: January 2026 stabilization week Message-ID: <6197e075-212f-4260-b437-902e06d1002b@app.fastmail.com> In-Reply-To: <2329920.sMrx5ctUpN@ravel> References: <aXctKTNm_CEau0yB@cell.glebi.us> <aXp2VvOLn063IcUs@cell.glebi.us> <2329920.sMrx5ctUpN@ravel>
index | next in thread | previous in thread | raw e-mail
[-- Attachment #1 --] On Thu, Jan 29, 2026, at 3:47 AM, Olivier Certner wrote: > > For now, it's either CPPC or powerd(8) (as for Intel), and a priori I'd recommend switching to CPPC entirely (once we have finished with the potential problems here). For pressing production issues, in the meantime you can set the 'machdep.hwpstate_amd_cppc_enable' tunable to 0 to force using the "manual" P-states that powerd(8) can manipulate. > On Intel, and now on AMD, it seems that CPPC is far worse than powerd (*). By that I mean that low power settings use the same amount of power as powerd when ramped down, but limit the performance when a server is busy. Conversely, midrange settings use more power than powerd and provide less performance. And high settings use more power. (*) At Netflix, we use a modified powerd, similar to powerdxx that takes multi-core load into account. So the ability to let software control frequency is something that I don't want to loose. Drew [-- Attachment #2 --] <!DOCTYPE html><html><head><title></title></head><body><div><br></div><div><br></div><div>On Thu, Jan 29, 2026, at 3:47 AM, Olivier Certner wrote:</div><blockquote type="cite" id="qt" style=""><div><br></div><div>For now, it's either CPPC or powerd(8) (as for Intel), and a priori I'd recommend switching to CPPC entirely (once we have finished with the potential problems here). For pressing production issues, in the meantime you can set the '<a href="http://machdep.hwpstate">machdep.hwpstate</a>_amd_cppc_enable' tunable to 0 to force using the "manual" P-states that powerd(8) can manipulate.</div><div><br></div></blockquote><div><br></div><div>On Intel, and now on AMD, it seems that CPPC is far worse than powerd (*). By that I mean that low power settings use the same amount of power as powerd when ramped down, but limit the performance when a server is busy. Conversely, midrange settings use more power than powerd and provide less performance. And high settings use more power.</div><div><br></div><div>(*) At Netflix, we use a modified powerd, similar to powerdxx that takes multi-core load into account.</div><div><br></div><div>So the ability to let software control frequency is something that I don't want to loose.</div><div><br></div><div>Drew</div></body></html>home | help
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?6197e075-212f-4260-b437-902e06d1002b>