Re: [RFC PATCH 0/4] CXL Hotness Monitoring Unit perf driver

[Gregory Price <gourry@gourry.net>]

On Thu, Nov 21, 2024 at 02:58:52PM +0000, Jonathan Cameron wrote:
> On Thu, 21 Nov 2024 09:24:43 -0500
> Gregory Price <gourry@gourry.net> wrote:
> 
> > On Thu, Nov 21, 2024 at 10:18:41AM +0000, Jonathan Cameron wrote:
> > > The CXL specification release 3.2 is now available under a click through at
> > > https://computeexpresslink.org/cxl-specification/ and it brings new
> > > shiny toys.
> > > 
> > > RFC reason
> > > - Whilst trace capture with a particular configuration is potentially useful
> > >   the intent is that CXL HMU units will be used to drive various forms of
> > >   hotpage migration for memory tiering setups. This driver doesn't do this
> > >   (yet), but rather provides data capture etc for experimentation and
> > >   for working out how to mostly put the allocations in the right place to
> > >   start with by tuning applications.
> > > 
> > > CXL r3.2 introduces a CXL Hotness Monitoring Unit definition. The intent
> > > of this is to provide a way to establish which units of memory (typically
> > > pages or larger) in CXL attached memory are hot. The implementation details
> > > and algorithm are all implementation defined. The specification simply
> > > describes the 'interface' which takes the form of ring buffer of hotness
> > > records in a PCI BAR and defined capability, configuration and status
> > > registers.
> > > 
> > > The hardware may have constraints on what it can track, granularity etc
> > > and on how accurately it tracks (e.g. counter exhaustion, inaccurate
> > > trackers). Some of these constraints are discoverable from the hardware
> > > registers, others such as loss of accuracy have no universally accepted
> > > measures as they are typically access pattern dependent. Sadly it is
> > > very unlikely any hardware will implement a truly precise tracker given
> > > the large resource requirements for tracking at a useful granularity.
> > > 
> > > There are two fundamental operation modes:
> > > 
> > > * Epoch based. Counters are checked after a period of time (Epoch) and
> > >   if over a threshold added to the hotlist.
> > > * Always on. Counters run until a threshold is reached, after that the
> > >   hot unit is added to the hotlist and the counter released.
> > > 
> > > Counting can be filtered on:
> > > 
> > > * Region of CXL DPA space (256MiB per bit in a bitmap).
> > > * Type of access - Trusted and non trusted or non trusted only, R/W/RW
> > > 
> > > Sampling can be modified by:
> > > 
> > > * Downsampling including potentially randomized downsampling.
> > > 
> > > The driver presented here is intended to be useful in its own right but
> > > also to act as the first step of a possible path towards hotness monitoring
> > > based hot page migration. Those steps might look like.
> > > 
> > > 1. Gather data - drivers provide telemetry like solutions to get that
> > >    data. May be enhanced, for example in this driver by providing the
> > >    HPA address rather than DPA Unit Address. Userspace can access enough
> > >    information to do this so maybe not.
> > > 2. Userspace algorithm development, possibly combined with userspace
> > >    triggered migration by PA. Working out how to use different levels
> > >    of constrained hardware resources will be challenging.  
> > 
> > FWIW this is what i was thinking about for this extension:
> > 
> > https://lore.kernel.org/all/20240319172609.332900-1-gregory.price@memverge.com/
> 
> Yup. I had that in mind. Forgot to actually add a link.
> 
> > 
> > At least for testing CHMU stuff. So if anyone is poking at testing such
> > things, they can feel free to use that for prototyping. However, I think
> > there is general discomfort around userspace handling HPA/DPA.
> > 
> > So it might look more like
> > 
> > echo nr_pages > /sys/.../tiering/nodeN/promote_pages
> > 
> > rather than handling the raw data from the CHMU to make decisions.
> 
> Agreed, but I think we are far away from a point where we can implement that.
> 
> Just working out how to tune the hardware to grab useful data is going
> to take a while to figure out, let alone doing anything much with it.
> 
> Without care you won't get a meaningful signal for what is actually
> hot out of the box. Lots of reasons why including:
> a) Exhaustion of tracking resources, due to looking at too large a window
>    or for too long.  Will probably need some form of auto updating of
>    what is being scanning (coarse to fine might work though I'm doubtful,
>    scanning across small regions maybe).
> b) Threshold too high, no detections.
> c) Threshold too low, everything hot.
> d) Wrong timescales. Hot is not a well defined thing.
> e) Hardware that won't do tracking at fine enough granularity.
> 

f) How does this even work with interleaving on larger pools :B
   It's pretend-addressing all the way down :D

Lots of conceptually complex and fun questions here.

~Gregory