Hi,
On 18/01/21 11:25, Song Bao Hua wrote: >> -----Original Message----- >> From: Vincent Guittot [mailto:vincent.guit...@linaro.org] >> Sent: Tuesday, January 19, 2021 12:14 AM >> To: Song Bao Hua (Barry Song) <song.bao....@hisilicon.com> >> Cc: Ingo Molnar <mi...@kernel.org>; Peter Zijlstra <pet...@infradead.org>; >> Dietmar Eggemann <dietmar.eggem...@arm.com>; Morten Rasmussen >> <morten.rasmus...@arm.com>; Valentin Schneider <valentin.schnei...@arm.com>; >> linux-kernel <linux-kernel@vger.kernel.org>; Mel Gorman <mgor...@suse.de>; >> linux...@openeuler.org >> Subject: Re: [RFC PATCH] sched/fair: first try to fix the scheduling impact >> of NUMA diameter > 2 >> >> On Fri, 15 Jan 2021 at 21:42, Barry Song <song.bao....@hisilicon.com> wrote: >> > >> > This patch is a follow-up of the 3-hops issue reported by Valentin >> > Schneider: >> > [1] https://lore.kernel.org/lkml/jhjtux5edo2.mog...@arm.com/ >> > [2] >> https://lore.kernel.org/lkml/20201110184300.15673-1-valentin.schneider@arm >> .com/ >> > >> > Here is a brief summary of the background: >> > For a NUMA system with 3-hops, sched_group for NUMA 2-hops could be not a >> > subset of sched_domain. >> > For example, for a system with the below topology(two cpus in each NUMA >> > node): >> > node 0 1 2 3 >> > 0: 10 12 20 22 >> > 1: 12 10 22 24 >> > 2: 20 22 10 12 >> > 3: 22 24 12 10 >> > >> > For CPU0, domain-2 will span 0-5, but its group will span 0-3, 4-7. >> > 4-7 isn't a subset of 0-5. >> > >> > CPU0 attaching sched-domain(s): >> > domain-0: span=0-1 level=MC >> > groups: 0:{ span=0 cap=989 }, 1:{ span=1 cap=1016 } >> > domain-1: span=0-3 level=NUMA >> > groups: 0:{ span=0-1 cap=2005 }, 2:{ span=2-3 cap=2028 } >> > domain-2: span=0-5 level=NUMA >> > groups: 0:{ span=0-3 cap=4033 }, 4:{ span=4-7 cap=3909 } >> > ERROR: groups don't span domain->span >> > domain-3: span=0-7 level=NUMA >> > groups: 0:{ span=0-5 mask=0-1 cap=6062 }, 6:{ span=4-7 mask=6-7 >> > cap=3928 } >> > >> > All other cpus also have the same issue: sched_group could be not a subset >> > of sched_domain. >> > >> > Here I am trying to figure out the scheduling impact of this issue from >> > two aspects: >> > 1. find busiest cpu in load_balance >> > 2. find idlest cpu in fork/exec/wake balance >> >> Would be better to fix the error in the sched domain topology instead >> of hacking the load balance to compensate the topology problem > > I think Valentin Schneider has tried to do that before, but failed. This will > add some new groups which won't be managed by current update_group_capacity()? > @Valentine, would you like to share more details? > Sorry for being late to the party, this is gnarly stuff and I can't dive back into it without spending some time staring at my notes and diagrams... I did indeed try to fix the group construction, thinking it would "just" be a matter of changing one mask into another, but it turned out to be quite trickier. Let's go back to https://lore.kernel.org/lkml/jhjtux5edo2.mog...@arm.com/ Right now, for that #Case study w/ QEMU platform, we get: CPU0-domain1: span=0-2 group0: span=0-2, mask=0 group2: span=1-3, mask=2 # CPU3 shouldn't be included CPU1-domain1: span=0-3 group1: span=0-2, mask=1 group3: span=2-3, mask=3 CPU2-domain1: span=0-3 group2: span=1-3, mask=2 group0: span=0-1, mask=0 CPU3-domain1: span=0-2 group3: span=2-3, mask=3 group1: span=0-2, mask=1 # CPU0 shouldn't be included We would want to "fix" this into: CPU0-domain1 group0: span=0-2, mask=0 - group2: span=1-3, mask=2 + group?: span=1-2, mask=?? CPU1-domain1 group1: span=0-2, mask=1 group3: span=2-3, mask=3 CPU2-domain1 group2: span=1-3, mask=2 group0: span=0-1, mask=0 CPU3-domain1 group3: span=2-3, mask=3 - group1: span=0-2, mask=1 + group?: span=1-2, mask=?? Note the '?' for the group ID and for the group balance mask. What I tried to hint at when writing this is that, right now, there is no sane ID nor balance mask to give to those "new" groups. The group ID is supposed to be the CPU owning the sched_group_capacity structure for the unique group span, which right now is the first CPU in the balance mask - I recommend reading the comments atop group_balance_cpu(), build_balance_mask() and get_group(). Here, we would end up with 5 unique group spans despite only having 4 CPUs: our allocation scheme doesn't hold up anymore. This stems from the way we allocate our topology data: we have a percpu slot per topology level. Furthermore, these "new" groups won't be the local group of any CPU, which means update_group_capacity() will never visit them - their capacity will never be updated. Here are some possible ways forward: - Have extra storage in struct sd_data for sched_group_capacity of those new, non-local groups. There might be topologies where you'll need to store more than one such group per CPU in there. - During load balance stats update, update the local group *and* all of those new, non-local ones. Conceptually I think this is what would be required, but it does feel very duct-tape-y... > On the other hand, another purpose of this RFC is that I also want to dig into > more details about how the 3-hops issue could affect the behavior of > scheduler. > In Valentine's original thread, I think we haven't figured out how the issue > will really impact scheduling. > I think the long story short was that since you can end up with groups spanning CPUs farther away than what the domain represents, the load balance stats computation (to figure out which busiest group to pull from) can and will be skewered. There are safeguards to prevent pulling from outside the domain span, but that doesn't protect the stats.