Date: Mon, 4 Mar 2019 12:58:14 -0800 From: Mark Millard <marklmi@yahoo.com> To: Bruce Evans <brde@optusnet.com.au> Cc: Konstantin Belousov <kostikbel@gmail.com>, freebsd-hackers Hackers <freebsd-hackers@freebsd.org>, FreeBSD PowerPC ML <freebsd-ppc@freebsd.org> Subject: Re: powerpc64 head -r344018 stuck sleeping problems: th->th_scale * tc_delta(th) overflows unsigned 64 bits sometimes [patched failed] Message-ID: <E473E08C-5EAD-4414-96B3-2EDF7B671974@yahoo.com> In-Reply-To: <20190305031010.I4610@besplex.bde.org> References: <20190301194217.GB68879@kib.kiev.ua> <20190302071425.G5025@besplex.bde.org> <20190302105140.GC68879@kib.kiev.ua> <20190302225513.W3408@besplex.bde.org> <20190302142521.GE68879@kib.kiev.ua> <20190303041441.V4781@besplex.bde.org> <20190303111931.GI68879@kib.kiev.ua> <20190303223100.B3572@besplex.bde.org> <20190303161635.GJ68879@kib.kiev.ua> <20190304043416.V5640@besplex.bde.org> <20190304114150.GM68879@kib.kiev.ua> <20190305031010.I4610@besplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 2019-Mar-4, at 10:17, Bruce Evans <brde at optusnet.com.au> wrote: >> . . . > > I think it is just a kludge that reduced the problem before it was fixed > properly using fences. > > Cross-socket latency is over 100 cycles according to jhb's tscskew > benchmark: on Haswell 4x2: > > CPU | TSC skew (min/avg/max/stddev) > ----+------------------------------ > 0 | 0 0 0 0.000 > 1 | 24 49 84 14.353 > 2 | 164 243 308 47.811 > 3 | 164 238 312 47.242 > 4 | 168 242 332 49.593 > 5 | 168 243 324 48.722 > 6 | 172 242 320 52.596 > 7 | 172 240 316 53.014 > > freefall is similar. Latency is apparently measured relative to CPU 0. > It is much lower to CPU 1 since that is on the same core. > You may want to look at: https://lists.freebsd.org/pipermail/freebsd-hackers/2019-March/054218.html for cruder, but somewhat related, information for the old Powermac G5 2-socket with 2 cores each, given how FreeBSD tries to synchronize the tbr's across cores as it starts up the CPUs. It may give some idea of a ball-park scale involved for such context, especially the reports of what happened for varying one figure in the source code. As stands, I've only done the experiments with a debug kernel build. I built using devel/powerpc64-xtoolchain-gcc related infrastructure, not gcc 4.2.1 . (This is typical for me.) === Mark Millard marklmi at yahoo.com ( dsl-only.net went away in early 2018-Mar)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?E473E08C-5EAD-4414-96B3-2EDF7B671974>