Re: [PATCH 3/9] x86/intel_rdt: Cache Allocation documentation and cgroup usage guide

[Marcelo Tosatti <mtosatti@redhat.com>]

On Sun, Aug 02, 2015 at 05:48:07PM +0200, Martin Kletzander wrote:
> On Thu, Jul 30, 2015 at 05:08:13PM -0300, Marcelo Tosatti wrote:
> >On Thu, Jul 30, 2015 at 10:47:23AM -0700, Vikas Shivappa wrote:
> >>
> >>
> >>Marcello,
> >>
> >>
> >>On Wed, 29 Jul 2015, Marcelo Tosatti wrote:
> >>>
> >>>How about this:
> >>>
> >>>desiredclos (closid  p1  p2  p3 p4)
> >>>	     1       1   0   0  0
> >>>	     2	     0	 0   0  1
> >>>	     3	     0   1   1  0
> >>
> >>#1 Currently in the rdt cgroup , the root cgroup always has all the
> >>bits set and cant be changed (because the cgroup hierarchy would by
> >>default make this to have all bits as all the children need to have
> >>a subset of the root's bitmask). So if the user creates a cgroup and
> >>not put any task in it , the tasks in the root cgroup could be still
> >>using that part of the cache. Thats the reason i say we can have
> >>really 'exclusive' masks.
> >>
> >>Or in other words - there is always a desired clos (0) which has all
> >>parts set which acts like a default pool.
> >>
> >>Also the parts can overlap.  Please apply this for all the below
> >>comments which will change the way they work.
> >>
> >>>
> >>>p means part.
> >>
> >>I am assuming p = (a contiguous cache capacity bit mask)
> >
> >Yes.
> >
> >>>closid 1 is a exclusive cgroup.
> >>>closid 2 is a "cache hog" class.
> >>>closid 3 is "default closid".
> >>>
> >>>Desiredclos is what user has specified.
> >>>
> >>>Transition 1: desiredclos --> effectiveclos
> >>>Clean all bits of unused closid's
> >>>(that must be updated whenever a
> >>>closid1 cgroup goes from empty->nonempty
> >>>and vice-versa).
> >>>
> >>>effectiveclos (closid  p1  p2  p3 p4)
> >>>	       1       0   0   0  0
> >>>	       2       0   0   0  1
> >>>	       3       0   1   1  0
> >>
> >>>
> >>>Transition 2: effectiveclos --> expandedclos
> >>>expandedclos (closid  p1  p2  p3 p4)
> >>>	       1       0   0   0  0
> >>>	       2       0   0   0  1
> >>>	       3       1   1   1  0
> >>>Then you have different inplacecos for each
> >>>CPU (see pseudo-code below):
> >>>
> >>>On the following events.
> >>>
> >>>- task migration to new pCPU:
> >>>- task creation:
> >>>
> >>>	id = smp_processor_id();
> >>>	for (part = desiredclos.p1; ...; part++)
> >>>		/* if my cosid is set and any other
> >>>	   	   cosid is clear, for the part,
> >>>		   synchronize desiredclos --> inplacecos */
> >>>		if (part[mycosid] == 1 &&
> >>>		    part[any_othercosid] == 0)
> >>>			wrmsr(part, desiredclos);
> >>>
> >>
> >>Currently the root cgroup would have all the bits set which will act
> >>like a default cgroup where all the otherwise unused parts (assuming
> >>they are a set of contiguous cache capacity bits) will be used.
> >>
> >>Otherwise the question is in the expandedclos - who decides to
> >>expand the closx parts to include some of the unused parts.. - that
> >>could just be a default root always ?
> >
> >Right, so the problem is for certain closid's you might never want
> >to expand (because doing so would cause data to be cached in a
> >cache way which might have high eviction rate in the future).
> >See the example from Will.
> >
> >But for the default cache (that is "unclassified applications"
> >i suppose it is beneficial to expand in most cases, that is,
> >use maximum amount of cache irrespective of eviction rate, which
> >is the behaviour that exists now without CAT).
> >
> >So perhaps a new flag "expand=y/n" can be added to the cgroup
> >directories... What do you say?
> >
> >Userspace representation of CAT
> >-------------------------------
> >
> >Usage model:
> >1) measure application performance without L3 cache reservation.
> >2) measure application perf with L3 cache reservation and
> >X number of cache ways until desired performance is attained.
> >
> >Requirements:
> >1) Persistency of CLOS configuration across hardware. On migration
> >of operating system or application between different hardware
> >systems we'd like the following to be maintained:
> >       - exclusive number of bytes (*) reserved to a certain CLOSid.
> >       - shared number of bytes (*) reserved between a certain group
> >         of CLOSid's.
> >
> >For both code and data, rounded down or up in cache way size.
> >
> >2) Reasoning:
> >Different CBM masks in different hardware platforms might be necessary
> >to specify the same CLOS configuration, in terms of exclusive number of
> >bytes and shared number of bytes. (cache-way rounded number of bytes).
> >For example, due to L3 allocation by other hardware entities in certain parts
> >of the cache it might be necessary to relocate CBM mask to achieve
> >the same CLOS configuration.
> >
> >3) Proposed format:
> >
> 
> Few questions from a random listener, I apologise if some of them are
> in a wrong place due to me missing some information from past threads.
> 
> I'm not sure whether the following proposal to the format is the
> internal structure or what's going to be in cgroups.  If this is
> user-visible interface, I think it could be a little less detailed.

User visible interface. The idea is to have userspace code that performs

[ user visible specification ]  ----> [ cbm bitmasks on present hardware
				       platform ]

In systemd, probably (or whatever is between the user and the cgroup
interface).

> >sharedregionK.exclusive - Number of exclusive cache bytes reserved for
> >			shared region.
> >sharedregionK.excl_data - Number of exclusive cache data bytes reserved for
> >	 		shared region.
> >sharedregionK.excl_bytes - Number of exclusive cache code bytes reserved for
> >			shared region.
> >sharedregionK.round_down - Round down to cache way bytes from respective number
> >		     specification (default is round up).
> >sharedregionK.expand - y/n - Expand shared region to more cache ways
> >			when available (default N).
> >
> >cgroupN.exclusive - Number of exclusive L3 cache bytes reserved
> >		    for cgroup.
> >cgroupN.excl_data - Number of exclusive L3 data cache bytes reserved
> >		    for cgroup.
> >cgroupN.excl_code - Number of exclusive L3 code cache bytes reserved
> >		    for cgroup.
> 
> By exclusive, you mean that it's exclusive to the tasks in this
> cgroup?

Correct.

> The thing is that we must differentiate between limiting some
> process's from hogging the memory (like example 2 below) and making
> some part of the cache exclusive for particular application (example 1
> below).

AFAICS there is no difference because: both require exclusive cache
access: the hog wants exclusive access between any other user of its
cachelines will be penalized. the high performance application wants 
exclusive cache access because any other user of its cachelines will
penalize it.

Where do you see the need to differentiate? 

> I just hope we won't need to add something similar to 'isolcpus=' just
> so we can make sure none of the tasks in the root cgroup can spoil the
> part of the cache we need to have exclusive.
> 
> I'm not sure creating a new subgroup and moving all the tasks there
> would work, It certainly is not possible with other cgroups, like the
> cpuset cgroup mentioned beforehand.

Why not? Should be able to place all tasks in a given cgroup? (trying
to setup systemd to do that now...).

> I also don't quite fully understand how the co-mounting with the
> cpuset cgroup should work, but that's not design-related.

Neither do I.

> One more question, how does this work on systems with multiple L3
> caches (e.g. large NUMA node systems)?  I'm guessing if the process is
> running only on some CPUs, the wrmsr() will be called on that
> particular CPU(s), right?

Not in the current patchset, that has to be fixed...

> >cgroupN.round_down - Round down to cache way bytes from respective number
> >		     specification (default is round up).
> >cgroupN.expand - y/n - Expand shared region to more cache ways when
> >		       available (default N).
> >cgroupN.shared = { sharedregion1, sharedregion2, ... } (list of shared
> >regions)
> >
> >Example 1:
> >One application with 2M exclusive cache, two applications
> >with 1M exclusive each, sharing an expansive shared region of 1M.
> >
> >cgroup1.exclusive = 2M
> >
> >sharedregion1.exclusive = 1M
> >sharedregion1.expand = Y
> >
> >cgroup2.exclusive = 1M
> >cgroup2.shared = sharedregion1
> >
> >cgroup3.exclusive = 1M
> >cgroup3.shared = sharedregion1
> >
> >Example 2:
> >3 high performance applications running, one of which is a cache hog
> >with no cache locality.
> >
> >cgroup1.exclusive = 8M
> >cgroup2.exclusive = 8M
> >
> >cgroup3.exclusive = 512K
> >cgroup3.round_down = Y
> >
> >In all cases the default cgroup (which requires no explicit
> >specification) is expansive and uses the remaining cache
> >ways, including the ways shared by other hardware entities.
> >
> >--
> >To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> >the body of a message to majordomo@vger.kernel.org
> >More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >Please read the FAQ at  http://www.tux.org/lkml/