Re: [PATCH v2 0/5] mm: demotion: Introduce new node state N_DEMOTION_TARGETS

["ying.huang@intel.com" <ying.huang@intel.com>]

On Wed, 2022-04-27 at 09:27 -0700, Wei Xu wrote:
> On Wed, Apr 27, 2022 at 12:11 AM ying.huang@intel.com
> <ying.huang@intel.com> wrote:
> > 
> > On Mon, 2022-04-25 at 09:56 -0700, Wei Xu wrote:
> > > On Sat, Apr 23, 2022 at 8:02 PM ying.huang@intel.com
> > > <ying.huang@intel.com> wrote:
> > > > 
> > > > Hi, All,
> > > > 
> > > > On Fri, 2022-04-22 at 16:30 +0530, Jagdish Gediya wrote:
> > > > 
> > > > [snip]
> > > > 
> > > > > I think it is necessary to either have per node demotion targets
> > > > > configuration or the user space interface supported by this patch
> > > > > series. As we don't have clear consensus on how the user interface
> > > > > should look like, we can defer the per node demotion target set
> > > > > interface to future until the real need arises.
> > > > > 
> > > > > Current patch series sets N_DEMOTION_TARGET from dax device kmem
> > > > > driver, it may be possible that some memory node desired as demotion
> > > > > target is not detected in the system from dax-device kmem probe path.
> > > > > 
> > > > > It is also possible that some of the dax-devices are not preferred as
> > > > > demotion target e.g. HBM, for such devices, node shouldn't be set to
> > > > > N_DEMOTION_TARGETS. In future, Support should be added to distinguish
> > > > > such dax-devices and not mark them as N_DEMOTION_TARGETS from the
> > > > > kernel, but for now this user space interface will be useful to avoid
> > > > > such devices as demotion targets.
> > > > > 
> > > > > We can add read only interface to view per node demotion targets
> > > > > from /sys/devices/system/node/nodeX/demotion_targets, remove
> > > > > duplicated /sys/kernel/mm/numa/demotion_target interface and instead
> > > > > make /sys/devices/system/node/demotion_targets writable.
> > > > > 
> > > > > Huang, Wei, Yang,
> > > > > What do you suggest?
> > > > 
> > > > We cannot remove a kernel ABI in practice.  So we need to make it right
> > > > at the first time.  Let's try to collect some information for the kernel
> > > > ABI definitation.
> > > > 
> > > > The below is just a starting point, please add your requirements.
> > > > 
> > > > 1. Jagdish has some machines with DRAM only NUMA nodes, but they don't
> > > > want to use that as the demotion targets.  But I don't think this is a
> > > > issue in practice for now, because demote-in-reclaim is disabled by
> > > > default.
> > > > 
> > > > 2. For machines with PMEM installed in only 1 of 2 sockets, for example,
> > > > 
> > > > Node 0 & 2 are cpu + dram nodes and node 1 are slow
> > > > memory node near node 0,
> > > > 
> > > > available: 3 nodes (0-2)
> > > > node 0 cpus: 0 1
> > > > node 0 size: n MB
> > > > node 0 free: n MB
> > > > node 1 cpus:
> > > > node 1 size: n MB
> > > > node 1 free: n MB
> > > > node 2 cpus: 2 3
> > > > node 2 size: n MB
> > > > node 2 free: n MB
> > > > node distances:
> > > > node   0   1   2
> > > >   0:  10  40  20
> > > >   1:  40  10  80
> > > >   2:  20  80  10
> > > > 
> > > > We have 2 choices,
> > > > 
> > > > a)
> > > > node    demotion targets
> > > > 0       1
> > > > 2       1
> > > > 
> > > > b)
> > > > node    demotion targets
> > > > 0       1
> > > > 2       X
> > > > 
> > > > a) is good to take advantage of PMEM.  b) is good to reduce cross-socket
> > > > traffic.  Both are OK as defualt configuration.  But some users may
> > > > prefer the other one.  So we need a user space ABI to override the
> > > > default configuration.
> > > 
> > > I think 2(a) should be the system-wide configuration and 2(b) can be
> > > achieved with NUMA mempolicy (which needs to be added to demotion).
> > 
> > Unfortunately, some NUMA mempolicy information isn't available at
> > demotion time, for example, mempolicy enforced via set_mempolicy() is
> > for thread. But I think that cpusets can work for demotion.
> > 
> > > In general, we can view the demotion order in a way similar to
> > > allocation fallback order (after all, if we don't demote or demotion
> > > lags behind, the allocations will go to these demotion target nodes
> > > according to the allocation fallback order anyway).  If we initialize
> > > the demotion order in that way (i.e. every node can demote to any node
> > > in the next tier, and the priority of the target nodes is sorted for
> > > each source node), we don't need per-node demotion order override from
> > > the userspace.  What we need is to specify what nodes should be in
> > > each tier and support NUMA mempolicy in demotion.
> > 
> > This sounds interesting. Tier sounds like a natural and general concept
> > for these memory types. It's attracting to use it for user space
> > interface too. For example, we may use that for mem_cgroup limits of a
> > specific memory type (tier).
> > 
> > And if we take a look at the N_DEMOTION_TARGETS again from the "tier"
> > point of view. The nodes are divided to 2 classes via
> > N_DEMOTION_TARGETS.
> > 
> > - The nodes without N_DEMOTION_TARGETS are top tier (or tier 0).
> > 
> > - The nodes with N_DEMOTION_TARGETS are non-top tier (or tier 1, 2, 3,
> > ...)
> > 
> 
> Yes, this is one of the main reasons why we (Google) want this interface.
> 
> > So, another possibility is to fit N_DEMOTION_TARGETS and its overriding
> > into "tier" concept too.  !N_DEMOTION_TARGETS == TIER0.
> > 
> > - All nodes start with TIER0
> > 
> > - TIER0 can be cleared for some nodes via e.g. kmem driver
> > 
> > TIER0 node list can be read or overriden by the user space via the
> > following interface,
> > 
> >   /sys/devices/system/node/tier0
> > 
> > In the future, if we want to customize more tiers, we can add tier1,
> > tier2, tier3, .....  For now, we can add just tier0.  That is, the
> > interface is extensible in the future compared with
> > .../node/demote_targets.
> > 
> 
> This more explicit tier definition interface works, too.
> 

In addition to make tiering definition explicit, more importantly, this
makes it much easier to support more than 2 tiers.  For example, for a
system with HBM (High Bandwidth Memory), CPU+DRAM, DRAM only, and PMEM,
that is, 3 tiers, we can put HBM in tier 0, CPU+DRAM and DRAM only in
tier 1, and PMEM in tier 2, automatically, or via user space
overridding.  N_DEMOTION_TARGETS isn't natural to be extended to support
this.

Best Regards,
Huang, Ying

> > This isn't as flexible as the writable per-node demotion targets.  But
> > it may be enough for most requirements?
> 
> I would think so. Besides, it doesn't really conflict with the
> per-node demotion target interface if we really want to introduce the
> latter.
> 
> > Best Regards,
> > Huang, Ying
> > 
> > > Cross-socket demotion should not be too big a problem in practice
> > > because we can optimize the code to do the demotion from the local CPU
> > > node (i.e. local writes to the target node and remote read from the
> > > source node).  The bigger issue is cross-socket memory access onto the
> > > demoted pages from the applications, which is why NUMA mempolicy is
> > > important here.
> > > 
> > > > 3. For machines with HBM (High Bandwidth Memory), as in
> > > > 
> > > > https://lore.kernel.org/lkml/39cbe02a-d309-443d-54c9-678a0799342d@gmail.com/
> > > > 
> > > > > [1] local DDR = 10, remote DDR = 20, local HBM = 31, remote HBM = 41
> > > > 
> > > > Although HBM has better performance than DDR, in ACPI SLIT, their
> > > > distance to CPU is longer.  We need to provide a way to fix this.  The
> > > > user space ABI is one way.  The desired result will be to use local DDR
> > > > as demotion targets of local HBM.
> > > > 
> > > > Best Regards,
> > > > Huang, Ying
> > > > 
> > 
> > 
> >