Re: [RFC PATCH v3 0/4] Node Weights and Weighted Interleave

[Johannes Weiner]

On Tue, Oct 31, 2023 at 04:56:27PM +0100, Michal Hocko wrote:
> On Tue 31-10-23 11:21:42, Johannes Weiner wrote:
> > On Tue, Oct 31, 2023 at 10:53:41AM +0100, Michal Hocko wrote:
> > > On Mon 30-10-23 20:38:06, Gregory Price wrote:
> > > > This patchset implements weighted interleave and adds a new sysfs
> > > > entry: /sys/devices/system/node/nodeN/accessM/il_weight.
> > > > 
> > > > The il_weight of a node is used by mempolicy to implement weighted
> > > > interleave when `numactl --interleave=...` is invoked.  By default
> > > > il_weight for a node is always 1, which preserves the default round
> > > > robin interleave behavior.
> > > > 
> > > > Interleave weights may be set from 0-100, and denote the number of
> > > > pages that should be allocated from the node when interleaving
> > > > occurs.
> > > > 
> > > > For example, if a node's interleave weight is set to 5, 5 pages
> > > > will be allocated from that node before the next node is scheduled
> > > > for allocations.
> > > 
> > > I find this semantic rather weird TBH. First of all why do you think it
> > > makes sense to have those weights global for all users? What if
> > > different applications have different view on how to spred their
> > > interleaved memory?
> > > 
> > > I do get that you might have a different tiers with largerly different
> > > runtime characteristics but why would you want to interleave them into a
> > > single mapping and have hard to predict runtime behavior?
> > > 
> > > [...]
> > > > In this way it becomes possible to set an interleaving strategy
> > > > that fits the available bandwidth for the devices available on
> > > > the system. An example system:
> > > > 
> > > > Node 0 - CPU+DRAM, 400GB/s BW (200 cross socket)
> > > > Node 1 - CPU+DRAM, 400GB/s BW (200 cross socket)
> > > > Node 2 - CXL Memory. 64GB/s BW, on Node 0 root complex
> > > > Node 3 - CXL Memory. 64GB/s BW, on Node 1 root complex
> > > > 
> > > > In this setup, the effective weights for nodes 0-3 for a task
> > > > running on Node 0 may be [60, 20, 10, 10].
> > > > 
> > > > This spreads memory out across devices which all have different
> > > > latency and bandwidth attributes at a way that can maximize the
> > > > available resources.
> > > 
> > > OK, so why is this any better than not using any memory policy rely
> > > on demotion to push out cold memory down the tier hierarchy?
> > > 
> > > What is the actual real life usecase and what kind of benefits you can
> > > present?
> > 
> > There are two things CXL gives you: additional capacity and additional
> > bus bandwidth.
> > 
> > The promotion/demotion mechanism is good for the capacity usecase,
> > where you have a nice hot/cold gradient in the workingset and want
> > placement accordingly across faster and slower memory.
> > 
> > The interleaving is useful when you have a flatter workingset
> > distribution and poorer access locality. In that case, the CPU caches
> > are less effective and the workload can be bus-bound. The workload
> > might fit entirely into DRAM, but concentrating it there is
> > suboptimal. Fanning it out in proportion to the relative performance
> > of each memory tier gives better resuls.
> > 
> > We experimented with datacenter workloads on such machines last year
> > and found significant performance benefits:
> > 
> > https://lore.kernel.org/linux-mm/YqD0%2FtzFwXvJ1gK6@xxxxxxxxxxx/T/
> 
> Thanks, this is a useful insight.
>  
> > This hopefully also explains why it's a global setting. The usecase is
> > different from conventional NUMA interleaving, which is used as a
> > locality measure: spread shared data evenly between compute
> > nodes. This one isn't about locality - the CXL tier doesn't have local
> > compute. Instead, the optimal spread is based on hardware parameters,
> > which is a global property rather than a per-workload one.
> 
> Well, I am not convinced about that TBH. Sure it is probably a good fit
> for this specific CXL usecase but it just doesn't fit into many others I
> can think of - e.g. proportional use of those tiers based on the
> workload - you get what you pay for.
> 
> Is there any specific reason for not having a new interleave interface
> which defines weights for the nodemask? Is this because the policy
> itself is very dynamic or is this more driven by simplicity of use?

A downside of *requiring* weights to be paired with the mempolicy is
that it's then the application that would have to figure out the
weights dynamically, instead of having a static host configuration. A
policy of "I want to be spread for optimal bus bandwidth" translates
between different hardware configurations, but optimal weights will
vary depending on the type of machine a job runs on.

That doesn't mean there couldn't be usecases for having weights as
policy as well in other scenarios, like you allude to above. It's just
so far such usecases haven't really materialized or spelled out
concretely. Maybe we just want both - a global default, and the
ability to override it locally. Could you elaborate on the 'get what
you pay for' usecase you mentioned?