From: Jonathan Cameron <jonathan.cameron@huawei.com> To: Michal Hocko <mhocko@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>, Andrew Morton <akpm@linux-foundation.org>, Linux Memory Management List <linux-mm@kvack.org>, kvm@vger.kernel.org, LKML <linux-kernel@vger.kernel.org>, Fan Du <fan.du@intel.com>, Yao Yuan <yuan.yao@intel.com>, Peng Dong <dongx.peng@intel.com>, Huang Ying <ying.huang@intel.com>, Liu Jingqi <jingqi.liu@intel.com>, Dong Eddie <eddie.dong@intel.com>, Dave Hansen <dave.hansen@intel.com>, Zhang Yi <yi.z.zhang@linux.intel.com>, Dan Williams <dan.j.williams@intel.com>, Mel Gorman <mgorman@suse.de>, Andrea Arcangeli <aarcange@redhat.com>, linux-accelerators@lists.ozlabs.org Subject: Re: [RFC][PATCH v2 00/21] PMEM NUMA node and hotness accounting/migration Date: Wed, 2 Jan 2019 12:21:10 +0000 [thread overview] Message-ID: <20190102122110.00000206@huawei.com> (raw) In-Reply-To: <20181228195224.GY16738@dhcp22.suse.cz> On Fri, 28 Dec 2018 20:52:24 +0100 Michal Hocko <mhocko@kernel.org> wrote: > [Ccing Mel and Andrea] > > On Fri 28-12-18 21:31:11, Wu Fengguang wrote: > > > > > I haven't looked at the implementation yet but if you are proposing a > > > > > special cased zone lists then this is something CDM (Coherent Device > > > > > Memory) was trying to do two years ago and there was quite some > > > > > skepticism in the approach. > > > > > > > > It looks we are pretty different than CDM. :) > > > > We creating new NUMA nodes rather than CDM's new ZONE. > > > > The zonelists modification is just to make PMEM nodes more separated. > > > > > > Yes, this is exactly what CDM was after. Have a zone which is not > > > reachable without explicit request AFAIR. So no, I do not think you are > > > too different, you just use a different terminology ;) > > > > Got it. OK.. The fall back zonelists patch does need more thoughts. > > > > In long term POV, Linux should be prepared for multi-level memory. > > Then there will arise the need to "allocate from this level memory". > > So it looks good to have separated zonelists for each level of memory. > > Well, I do not have a good answer for you here. We do not have good > experiences with those systems, I am afraid. NUMA is with us for more > than a decade yet our APIs are coarse to say the least and broken at so > many times as well. Starting a new API just based on PMEM sounds like a > ticket to another disaster to me. > > I would like to see solid arguments why the current model of numa nodes > with fallback in distances order cannot be used for those new > technologies in the beginning and develop something better based on our > experiences that we gain on the way. > > I would be especially interested about a possibility of the memory > migration idea during a memory pressure and relying on numa balancing to > resort the locality on demand rather than hiding certain NUMA nodes or > zones from the allocator and expose them only to the userspace. This is indeed a very interesting direction. I'm coming at this from a CCIX point of view. Ignore the next bit of you are already familiar with CCIX :) Main thing CCIX brings is that memory can be fully coherent anywhere in the system including out near accelerators, all via shared physical address space, leveraging ATS / IOMMUs / MMUs to do translations. Result is a big and possibly extremely heterogenous NUMA system. All the setup is done in firmware so by the time the kernel sees it everything is in SRAT / SLIT / NFIT / HMAT etc. We have a few usecases that need some more fine grained control combined with automated balancing. So far we've been messing with nasty tricks like hotplugging memory after boot a long way away, or the original CDM zone patches (knowing they weren't likely to go anywhere!) Userspace is all hand tuned which is not great in the long run... Use cases (I've probably missed some): * Storage Class Memory near to the host CPU / DRAM controllers (pretty much the same as this series is considering). Note that there isn't necessarily any 'pairing' with host DRAM as seen in this RFC. A typical system might have a large single pool with similar access characteristics from each host SOC. The paired approach is probably going to be common in early systems though. Also not necessarily Non Volatile, could just be a big DDR expansion board. * RAM out near an accelerator. Aim would be to migrate data to that RAM if the access patterns from the accelerator justify it being there rather than near any of the host CPUs. In a memory pressure on host situation anything could be pushed out there as probably still better than swapping. Note that this would require some knowledge of 'who' is doing the accessing which isn't needed for what this RFC is doing. * Hot pages may not be hot just because the host is using them a lot. It would be very useful to have a means of adding information available from accelerators beyond simple accessed bits (dreaming ;) One problem here is translation caches (ATCs) as they won't normally result in any updates to the page accessed bits. The arm SMMU v3 spec for example makes it clear (though it's kind of obvious) that the ATS request is the only opportunity to update the accessed bit. The nasty option here would be to periodically flush the ATC to force the access bit updates via repeats of the ATS request (ouch). That option only works if the iommu supports updating the accessed flag (optional on SMMU v3 for example). We need the explicit placement, but can get that from existing NUMA controls. More of a concern is persuading the kernel it really doesn't want to put it's data structures in distant memory as it can be very very distant. So ideally I'd love this set to head in a direction that helps me tick off at least some of the above usecases and hopefully have some visibility on how to address the others moving forwards, Good to see some new thoughts in this area! Jonathan > > > On the other hand, there will also be page allocations that don't care > > about the exact memory level. So it looks reasonable to expect > > different kind of fallback zonelists that can be selected by NUMA policy. > > > > Thanks, > > Fengguang >
WARNING: multiple messages have this Message-ID (diff)
From: Jonathan Cameron <jonathan.cameron@huawei.com> To: Michal Hocko <mhocko@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>, Andrew Morton <akpm@linux-foundation.org>, Linux Memory Management List <linux-mm@kvack.org>, <kvm@vger.kernel.org>, LKML <linux-kernel@vger.kernel.org>, Fan Du <fan.du@intel.com>, Yao Yuan <yuan.yao@intel.com>, Peng Dong <dongx.peng@intel.com>, Huang Ying <ying.huang@intel.com>, Liu Jingqi <jingqi.liu@intel.com>, Dong Eddie <eddie.dong@intel.com>, Dave Hansen <dave.hansen@intel.com>, Zhang Yi <yi.z.zhang@linux.intel.com>, Dan Williams <dan.j.williams@intel.com>, "Mel Gorman" <mgorman@suse.de>, Andrea Arcangeli <aarcange@redhat.com>, <linux-accelerators@lists.ozlabs.org> Subject: Re: [RFC][PATCH v2 00/21] PMEM NUMA node and hotness accounting/migration Date: Wed, 2 Jan 2019 12:21:10 +0000 [thread overview] Message-ID: <20190102122110.00000206@huawei.com> (raw) Message-ID: <20190102122110.lh-xv9i-xHWvvhi32KSjbs9k3yWWjgFoDYhoSgrjecE@z> (raw) In-Reply-To: <20181228195224.GY16738@dhcp22.suse.cz> On Fri, 28 Dec 2018 20:52:24 +0100 Michal Hocko <mhocko@kernel.org> wrote: > [Ccing Mel and Andrea] > > On Fri 28-12-18 21:31:11, Wu Fengguang wrote: > > > > > I haven't looked at the implementation yet but if you are proposing a > > > > > special cased zone lists then this is something CDM (Coherent Device > > > > > Memory) was trying to do two years ago and there was quite some > > > > > skepticism in the approach. > > > > > > > > It looks we are pretty different than CDM. :) > > > > We creating new NUMA nodes rather than CDM's new ZONE. > > > > The zonelists modification is just to make PMEM nodes more separated. > > > > > > Yes, this is exactly what CDM was after. Have a zone which is not > > > reachable without explicit request AFAIR. So no, I do not think you are > > > too different, you just use a different terminology ;) > > > > Got it. OK.. The fall back zonelists patch does need more thoughts. > > > > In long term POV, Linux should be prepared for multi-level memory. > > Then there will arise the need to "allocate from this level memory". > > So it looks good to have separated zonelists for each level of memory. > > Well, I do not have a good answer for you here. We do not have good > experiences with those systems, I am afraid. NUMA is with us for more > than a decade yet our APIs are coarse to say the least and broken at so > many times as well. Starting a new API just based on PMEM sounds like a > ticket to another disaster to me. > > I would like to see solid arguments why the current model of numa nodes > with fallback in distances order cannot be used for those new > technologies in the beginning and develop something better based on our > experiences that we gain on the way. > > I would be especially interested about a possibility of the memory > migration idea during a memory pressure and relying on numa balancing to > resort the locality on demand rather than hiding certain NUMA nodes or > zones from the allocator and expose them only to the userspace. This is indeed a very interesting direction. I'm coming at this from a CCIX point of view. Ignore the next bit of you are already familiar with CCIX :) Main thing CCIX brings is that memory can be fully coherent anywhere in the system including out near accelerators, all via shared physical address space, leveraging ATS / IOMMUs / MMUs to do translations. Result is a big and possibly extremely heterogenous NUMA system. All the setup is done in firmware so by the time the kernel sees it everything is in SRAT / SLIT / NFIT / HMAT etc. We have a few usecases that need some more fine grained control combined with automated balancing. So far we've been messing with nasty tricks like hotplugging memory after boot a long way away, or the original CDM zone patches (knowing they weren't likely to go anywhere!) Userspace is all hand tuned which is not great in the long run... Use cases (I've probably missed some): * Storage Class Memory near to the host CPU / DRAM controllers (pretty much the same as this series is considering). Note that there isn't necessarily any 'pairing' with host DRAM as seen in this RFC. A typical system might have a large single pool with similar access characteristics from each host SOC. The paired approach is probably going to be common in early systems though. Also not necessarily Non Volatile, could just be a big DDR expansion board. * RAM out near an accelerator. Aim would be to migrate data to that RAM if the access patterns from the accelerator justify it being there rather than near any of the host CPUs. In a memory pressure on host situation anything could be pushed out there as probably still better than swapping. Note that this would require some knowledge of 'who' is doing the accessing which isn't needed for what this RFC is doing. * Hot pages may not be hot just because the host is using them a lot. It would be very useful to have a means of adding information available from accelerators beyond simple accessed bits (dreaming ;) One problem here is translation caches (ATCs) as they won't normally result in any updates to the page accessed bits. The arm SMMU v3 spec for example makes it clear (though it's kind of obvious) that the ATS request is the only opportunity to update the accessed bit. The nasty option here would be to periodically flush the ATC to force the access bit updates via repeats of the ATS request (ouch). That option only works if the iommu supports updating the accessed flag (optional on SMMU v3 for example). We need the explicit placement, but can get that from existing NUMA controls. More of a concern is persuading the kernel it really doesn't want to put it's data structures in distant memory as it can be very very distant. So ideally I'd love this set to head in a direction that helps me tick off at least some of the above usecases and hopefully have some visibility on how to address the others moving forwards, Good to see some new thoughts in this area! Jonathan > > > On the other hand, there will also be page allocations that don't care > > about the exact memory level. So it looks reasonable to expect > > different kind of fallback zonelists that can be selected by NUMA policy. > > > > Thanks, > > Fengguang >
next prev parent reply other threads:[~2019-01-02 12:22 UTC|newest] Thread overview: 99+ messages / expand[flat|nested] mbox.gz Atom feed top 2018-12-26 13:14 [RFC][PATCH v2 00/21] PMEM NUMA node and hotness accounting/migration Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 01/21] e820: cheat PMEM as DRAM Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-27 3:41 ` Matthew Wilcox 2018-12-27 4:11 ` Fengguang Wu 2018-12-27 5:13 ` Dan Williams 2018-12-27 5:13 ` Dan Williams 2018-12-27 19:32 ` Yang Shi 2018-12-27 19:32 ` Yang Shi 2018-12-28 3:27 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 02/21] acpi/numa: memorize NUMA node type from SRAT table Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 03/21] x86/numa_emulation: fix fake NUMA in uniform case Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 04/21] x86/numa_emulation: pass numa node type to fake nodes Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 05/21] mmzone: new pgdat flags for DRAM and PMEM Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 06/21] x86,numa: update numa node type Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 07/21] mm: export node type {pmem|dram} under /sys/bus/node Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 08/21] mm: introduce and export pgdat peer_node Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-27 20:07 ` Christopher Lameter 2018-12-27 20:07 ` Christopher Lameter 2018-12-28 2:31 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 09/21] mm: avoid duplicate peer target node Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 10/21] mm: build separate zonelist for PMEM and DRAM node Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2019-01-01 9:14 ` Aneesh Kumar K.V 2019-01-01 9:14 ` Aneesh Kumar K.V 2019-01-07 9:57 ` Fengguang Wu 2019-01-07 14:09 ` Aneesh Kumar K.V 2018-12-26 13:14 ` [RFC][PATCH v2 11/21] kvm: allocate page table pages from DRAM Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2019-01-01 9:23 ` Aneesh Kumar K.V 2019-01-01 9:23 ` Aneesh Kumar K.V 2019-01-02 0:59 ` Yuan Yao 2019-01-02 16:47 ` Dave Hansen 2019-01-07 10:21 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 12/21] x86/pgtable: " Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:14 ` [RFC][PATCH v2 13/21] x86/pgtable: dont check PMD accessed bit Fengguang Wu 2018-12-26 13:14 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 14/21] kvm: register in mm_struct Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2019-02-02 6:57 ` Peter Xu 2019-02-02 10:50 ` Fengguang Wu 2019-02-04 10:46 ` Paolo Bonzini 2018-12-26 13:15 ` [RFC][PATCH v2 15/21] ept-idle: EPT walk for virtual machine Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 16/21] mm-idle: mm_walk for normal task Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 17/21] proc: introduce /proc/PID/idle_pages Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 18/21] kvm-ept-idle: enable module Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 19/21] mm/migrate.c: add move_pages(MPOL_MF_SW_YOUNG) flag Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 20/21] mm/vmscan.c: migrate anon DRAM pages to PMEM node Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-26 13:15 ` [RFC][PATCH v2 21/21] mm/vmscan.c: shrink anon list if can migrate to PMEM Fengguang Wu 2018-12-26 13:15 ` Fengguang Wu 2018-12-27 20:31 ` [RFC][PATCH v2 00/21] PMEM NUMA node and hotness accounting/migration Michal Hocko 2018-12-28 5:08 ` Fengguang Wu 2018-12-28 8:41 ` Michal Hocko 2018-12-28 9:42 ` Fengguang Wu 2018-12-28 12:15 ` Michal Hocko 2018-12-28 13:15 ` Fengguang Wu 2018-12-28 13:15 ` Fengguang Wu 2018-12-28 19:46 ` Michal Hocko 2018-12-28 13:31 ` Fengguang Wu 2018-12-28 18:28 ` Yang Shi 2018-12-28 18:28 ` Yang Shi 2018-12-28 19:52 ` Michal Hocko 2019-01-02 12:21 ` Jonathan Cameron [this message] 2019-01-02 12:21 ` Jonathan Cameron 2019-01-08 14:52 ` Michal Hocko 2019-01-10 15:53 ` Jerome Glisse 2019-01-10 15:53 ` Jerome Glisse 2019-01-10 16:42 ` Michal Hocko 2019-01-10 17:42 ` Jerome Glisse 2019-01-10 17:42 ` Jerome Glisse 2019-01-10 18:26 ` Jonathan Cameron 2019-01-10 18:26 ` Jonathan Cameron 2019-01-28 17:42 ` Jonathan Cameron 2019-01-28 17:42 ` Jonathan Cameron 2019-01-29 2:00 ` Fengguang Wu 2019-01-03 10:57 ` Mel Gorman 2019-01-10 16:25 ` Jerome Glisse 2019-01-10 16:25 ` Jerome Glisse 2019-01-10 16:50 ` Michal Hocko 2019-01-10 18:02 ` Jerome Glisse 2019-01-10 18:02 ` Jerome Glisse 2019-01-02 18:12 ` Dave Hansen 2019-01-08 14:53 ` Michal Hocko
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=20190102122110.00000206@huawei.com \ --to=jonathan.cameron@huawei.com \ --cc=aarcange@redhat.com \ --cc=akpm@linux-foundation.org \ --cc=dan.j.williams@intel.com \ --cc=dave.hansen@intel.com \ --cc=dongx.peng@intel.com \ --cc=eddie.dong@intel.com \ --cc=fan.du@intel.com \ --cc=fengguang.wu@intel.com \ --cc=jingqi.liu@intel.com \ --cc=kvm@vger.kernel.org \ --cc=linux-accelerators@lists.ozlabs.org \ --cc=linux-kernel@vger.kernel.org \ --cc=linux-mm@kvack.org \ --cc=mgorman@suse.de \ --cc=mhocko@kernel.org \ --cc=yi.z.zhang@linux.intel.com \ --cc=ying.huang@intel.com \ --cc=yuan.yao@intel.com \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).