Meeting Details

...

Meeting notes 

• Lorenzo Pieralisi : Arm : Scheduler topology: cluster scheduler awareness
  • Song Bao Hua
    • talking to slides
    • Better performance in single cluster vs multiple - lead to the investigation, used Jonathon patch to test Linux scheduler.
    • hackbench better with select_idle_sibling.  it collates the waking up of a cluster
  • Vincent, Morten there are multiple topics
    • Morten
      • how to get information
      • scheduling policy - this first
      • Need a wider discussion. Made some comments on Friday
    • Vincent 
      • Scheduler policy - adding scheduling level,  I don't see any pushback and there is perf improvement
      • Sometimes you want to spread and other times cluster - HB there are 40 threads, there are few idle states, it is about spreading tasks. Patches should fit on Peter's wakeup patches, NUMA scheduling will be most challenging.
      • Dietmar - will this scan and fall back to standard LNC? (V3)  VG: cant loop, but in a new approach, it is OK as you hit each CPU only once.
      • Morten, when is it good to pack vs spread them?  What is the measure to understand which approach to take.  And example two tasks with a lot of shared data will be a problem.
      • Vincent, need to look at previous CPU being idle. migrating may happen to move tasks from one cluster to another, need to complete task before load balance takes effect
      • JC:  if this is spread across NUMA nodes we have many more than 40 CPUs
        • VG: There will not be  be any on the longer one.
      • DE: we don't have an example where we chose a policy based on the workload
    • Lorenzo - we have covered topic one of this 35 mins spent, let's move to the list where we can
    • Tim Chen
      • X86 will also need this concept
      • could have tasks that share memory without a wakeup relationship
      • SBH - userspace may need to provide information 
    • JC: we can gather the info even if it is not accurate initially, we perhaps need to gather data separately
      • X86 does not use PPTT - but there are extra levels of topology information
      • TC we are using L2 
      • JC we don't have a way to represent L3 tag cache
      • Die is no available to userspace
      • Any concerns on extra complexity?
      • LP: userspace should be architectural - JC: sort of agree, with generic naming, MR: why don't you just dump PPTT, JC: kernel is not always enabled to get dump MC: we could patch this.
    • SBH will send a V4 at this time

LP - AOB

• Lorenzo Pieralisi : Arm : ACPI CPU hotplug
  • Progress with discussions - 2  PSCI solutions, QEmu changes or use ACPI as it is in X86 world
  • Working inside Arm as things firm up
  • For hotplug in a VM, suspect it relies on userspace - is actually align in userspace,   Salili  Salil Mehta that is correct. 
    • SM: at QEMU level you need an event, LP: in guest, there is something to be done, SM yes, we are using sysfs - Justing Hey Justin He was using PSoc VSOCK event -, we are doing it manually for testing
    • RMR - we sorted it out, need to get it done specs will follow
• JC - don't re-enumerate PCI - discussion with Jean-Phillip

Action items