Re: [PATCH 1/7] node: Link memory nodes to their compute nodes

From: Dan Williams
Date: Fri Nov 16 2018 - 17:55:42 EST


On Thu, Nov 15, 2018 at 12:37 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Nov 15, 2018 at 07:59:20AM -0700, Keith Busch wrote:
> > On Thu, Nov 15, 2018 at 05:57:10AM -0800, Matthew Wilcox wrote:
> > > On Wed, Nov 14, 2018 at 03:49:14PM -0700, Keith Busch wrote:
> > > > Memory-only nodes will often have affinity to a compute node, and
> > > > platforms have ways to express that locality relationship.
> > > >
> > > > A node containing CPUs or other DMA devices that can initiate memory
> > > > access are referred to as "memory iniators". A "memory target" is a
> > > > node that provides at least one phyiscal address range accessible to a
> > > > memory initiator.
> > >
> > > I think I may be confused here. If there is _no_ link from node X to
> > > node Y, does that mean that node X's CPUs cannot access the memory on
> > > node Y? In my mind, all nodes can access all memory in the system,
> > > just not with uniform bandwidth/latency.
> >
> > The link is just about which nodes are "local". It's like how nodes have
> > a cpulist. Other CPUs not in the node's list can acces that node's memory,
> > but the ones in the mask are local, and provide useful optimization hints.
>
> So ... let's imagine a hypothetical system (I've never seen one built like
> this, but it doesn't seem too implausible). Connect four CPU sockets in
> a square, each of which has some regular DIMMs attached to it. CPU A is
> 0 hops to Memory A, one hop to Memory B and Memory C, and two hops from
> Memory D (each CPU only has two "QPI" links). Then maybe there's some
> special memory extender device attached on the PCIe bus. Now there's
> Memory B1 and B2 that's attached to CPU B and it's local to CPU B, but
> not as local as Memory B is ... and we'd probably _prefer_ to allocate
> memory for CPU A from Memory B1 than from Memory D. But ... *mumble*,
> this seems hard.
>
> I understand you're trying to reflect what the HMAT table is telling you,
> I'm just really fuzzy on who's ultimately consuming this information
> and what decisions they're trying to drive from it.

The singular "local" is a limitation of the HMAT, but I would expect
the Linux translation of "local" would allow for multiple initiators
that can achieve some semblance of the "best" performance. Anything
less than best is going to have a wide range of variance and will
likely devolve to looking at the platform firmware data table
directly. The expected 80% case is software wants to be able to ask
"which CPUs should I run on to get the best access to this memory?"