* Peter Zijlstra [off-list ref] [2023-09-05 00:10:04]:
On Wed, Aug 30, 2023 at 05:56:14PM +0530, Srikar Dronamraju wrote:
quoted
Currently there are four powerpc specific sched topologies. These are
all statically defined. However not all these topologies are used by
all powerpc systems.
To avoid unnecessary degenerations by the scheduler , masks and flags
are compared. However if the sched topologies are build dynamically then
the code is simpler and there are greater chances of avoiding
degenerations.
Even x86 builds its sched topologies dynamically and new changes are
very similar to the way x86 is building its topologies.
Thanks Peter for taking a look.
You're not stating it explicitly, but you're doing this as a performance
optimization, right? The x86 thing didn't particularly care about
avoiding degenerate topologies -- it's just that the fixed tables method
grew unwieldy due to combinatorics.
Yes, its an optimization. On Powerpc, there is an utility ppc64, which users
would use to set their SMT mode, and whenever they do we end up recreating
the topology. Hence avoiding degenerates esp on large systems, should help.
Also dynamic add of CPUs is more common on Powerpc. Hence there also we
would avoid degenerating unnecessary domains.
And how does this patch relate to the other series touching this?
powerpc/smp: Shared processor sched optimizations
This patch will work independent of that patchset.
However Shared processor sched optimization patchset makes MC domain
avoid degeneration. Hence this patch will benefit from that patchset.
i.e without the Shared processor sched patchset, has_coregroup_support()
will return true on Power10 for even shared processor. And hence the
scheduler will create and destroy MC domains. If the patchset is already
present, on Power10 for shared processors, we will avoid MC domains.
Other that this there wont be any change.
--
Thanks and Regards
Srikar Dronamraju