yahoo-eng-team team mailing list archive

Thread
Date
[Bug 1995153] Re: `socket` PCI NUMA policy doesn't work if another instance with a NUMA topology is booted first on the same host

To: yahoo-eng-team@xxxxxxxxxxxxxxxxxxx
From: OpenStack Infra <1995153@xxxxxxxxxxxxxxxxxx>
Date: Thu, 04 May 2023 10:27:31 -0000
Reply-to: Bug 1995153 <1995153@xxxxxxxxxxxxxxxxxx>
Sender: noreply@xxxxxxxxxxxxx
Reviewed:  https://review.opendev.org/c/openstack/nova/+/862964
Committed: https://opendev.org/openstack/nova/commit/04ebae9dc01ebd24552b5aacd1a0f8b129013a9e
Submitter: "Zuul (22348)"
Branch:    master

commit 04ebae9dc01ebd24552b5aacd1a0f8b129013a9e
Author: Artom Lifshitz <alifshit@xxxxxxxxxx>
Date:   Fri Oct 28 18:09:35 2022 -0400

    Save cell socket correctly when updating host NUMA topology
    
    Previously, in numa_usage_from_instance_numa(), any new NUMACell
    objects we created did not have the `socket` attribute. In some cases
    this was persisted all the way down to the database. Fix this by
    copying `socket` from the old_cell.
    
    Change-Id: I9ed3c31ccd3220b02d951fc6dbc5ea049a240a68
    Closes-Bug: 1995153


** Changed in: nova
       Status: In Progress => Fix Released

-- 
You received this bug notification because you are a member of Yahoo!
Engineering Team, which is subscribed to OpenStack Compute (nova).
https://bugs.launchpad.net/bugs/1995153

Title:
  `socket` PCI NUMA policy doesn't work if another instance with a NUMA
  topology is booted first on the same host

Status in OpenStack Compute (nova):
  Fix Released

Bug description:
  Disclaimer: I haven't reproduced this in a functional test, but based
  on the traceback that I gathered from a real environment as well as
  the fact that the proposed fix actually fixes this, I think my theory
  is correct

  Description
  ===========
  `socket` PCI NUMA policy doesn't work if another instance is booted first on the same host

  Steps to reproduce
  ==================
  1. Boot any instance.
  2. Boot an instance with the `socket` PCI NUMA policy on the same host.

  Expected result
  ===============
  `socket` instance boots.

  Actual result
  =============
  Instance creation fails with

  Details: Fault: {'code': 500, 'created': '2022-10-28T20:17:31Z',
  'message': 'NotImplementedError'}. Server boot request ID:
  req-e3fd15d7-fb79-440f-b2f3-e6b2a5505e56.

  
  Environment
  ===========
  Originally reported as part of QE verification of [1], so stable/wallaby.

  Additional info
  ===============
  Playing around with the whitebox test for the socket policy [2] on a wallaby deployment, I noticed that the `socket` field in the compute.numa_topology column was being switched to `null` then back to its correct value (0 or 1).

  I added logging of the stack trace to the resource tracker _udpate()
  method right before it calls compute_node.save(), and found that
  `null` was getting saved when an instance was being booted or deleted.
  Example of a traceback:

  File "/usr/lib/python3.9/site-packages/nova/utils.py", line 686, in context_wrapper\n    func(*args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/oslo_concurrency/lockutils.py", line 360, in inner\n    return f(*args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 2126, in _locked_do_build_and_run_instance\n    result = self._do_build_and_run_instance(*args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/exception_wrapper.py", line 63, in wrapped\n    return f(self, context, *args, **kw)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 154, in decorated_function\n    return function(self, context, *args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/utils.py", line 1434, in decorated_function\n    return function(self, context, *args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 200, in decorated_function\n    return function(self, context, *args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 2232, in _do_build_and_run_instance\n    self._build_and_run_instance(context, instance, image,\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 2383, in _build_and_run_instance\n    with self.rt.instance_claim(context, instance, node, allocs,\n'
  '  File "/usr/lib/python3.9/site-packages/oslo_concurrency/lockutils.py", line 360, in inner\n    return f(*args, **kwargs)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py", line 197, in instance_claim\n    self._update(elevated, cn)\n'
  '  File "/usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py", line 1247, in _update\n    LOG.debug(\'artom: %s\', traceback.format_stack())\n'] _update /usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py:1247

  Similarly for delete:

  2022-10-28 21:57:27.091 2 DEBUG nova.compute.resource_tracker [req-c9fa718c-983e-416c-bc87-9564b8747294 d6d16a793ab74fe6a0b5594d037d3165 599a6777a45d46a09a7e233a926b7675 - default default] artom: ['  File "/usr/lib/python3.9/site-packages
  /eventlet/greenpool.py", line 88, in _spawn_n_impl\n    func(*args, **kwargs)\n', '  File "/usr/lib/python3.9/site-packages/futurist/_green.py", line 71, in __call__\n    self.work.run()\n', '  File "/usr/lib/python3.9/site-packages/futur
  ist/_utils.py", line 49, in run\n    result = self.fn(*self.args, **self.kwargs)\n', '  File "/usr/lib/python3.9/site-packages/oslo_messaging/rpc/server.py", line 165, in _process_incoming\n    res = self.dispatcher.dispatch(message)\n', 
  '  File "/usr/lib/python3.9/site-packages/oslo_messaging/rpc/dispatcher.py", line 309, in dispatch\n    return self._do_dispatch(endpoint, method, ctxt, args)\n', '  File "/usr/lib/python3.9/site-packages/oslo_messaging/rpc/dispatcher.py"
  , line 229, in _do_dispatch\n    result = func(ctxt, **new_args)\n', '  File "/usr/lib/python3.9/site-packages/nova/exception_wrapper.py", line 63, in wrapped\n    return f(self, context, *args, **kw)\n', '  File "/usr/lib/python3.9/site-
  packages/nova/compute/manager.py", line 154, in decorated_function\n    return function(self, context, *args, **kwargs)\n', '  File "/usr/lib/python3.9/site-packages/nova/compute/utils.py", line 1434, in decorated_function\n    return fun
  ction(self, context, *args, **kwargs)\n', '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 200, in decorated_function\n    return function(self, context, *args, **kwargs)\n', '  File "/usr/lib/python3.9/site-packag
  es/nova/compute/manager.py", line 3072, in terminate_instance\n    do_terminate_instance(instance, bdms)\n', '  File "/usr/lib/python3.9/site-packages/oslo_concurrency/lockutils.py", line 360, in inner\n    return f(*args, **kwargs)\n', '
    File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 3060, in do_terminate_instance\n    self._delete_instance(context, instance, bdms)\n', '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 3024, 
  in _delete_instance\n    self._complete_deletion(context, instance)\n', '  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 828, in _complete_deletion\n    self._update_resource_tracker(context, instance)\n', '  File 
  "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line 596, in _update_resource_tracker\n    self.rt.update_usage(context, instance, instance.node)\n', '  File "/usr/lib/python3.9/site-packages/oslo_concurrency/lockutils.py", li
  ne 360, in inner\n    return f(*args, **kwargs)\n', '  File "/usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py", line 658, in update_usage\n    self._update(context.elevated(), self.compute_nodes[nodename])\n', '  File "/u
  sr/lib/python3.9/site-packages/nova/compute/resource_tracker.py", line 1247, in _update\n    LOG.debug(\'artom: %s\', traceback.format_stack())\n'] _update /usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py:1247

  On the other hand, the resource tracker's periodic resource update
  task was saving the socket correctly:

  2022-10-28 21:57:59.794 2 DEBUG nova.compute.resource_tracker
  [req-31329b8b-0de4-4b30-b2a1-dcd4d62369b4 - - - - -] artom: ['  File
  "/usr/lib/python3.9/site-packages/eventlet/greenthread.py", line 221,
  in main\n    result = function(*args, **kwargs)\n', '  File
  "/usr/lib/python3.9/site-packages/oslo_service/loopingcall.py", line
  150, in _run_loop\n    result = func(*self.args, **self.kw)\n', '
  File "/usr/lib/python3.9/site-packages/nova/service.py", line 307, in
  periodic_tasks\n    return self.manager.periodic_tasks(ctxt,
  raise_on_error=raise_on_error)\n', '  File "/usr/lib/python3.9/site-
  packages/nova/manager.py", line 104, in periodic_tasks\n    return
  self.run_periodic_tasks(context, raise_on_error=raise_on_error)\n', '
  File "/usr/lib/python3.9/site-packages/oslo_service/periodic_task.py",
  line 216, in run_periodic_tasks\n    task(self, context)\n', '  File
  "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line
  10026, in update_available_resource\n
  self._update_available_resource_for_node(context, nodename,\n', '
  File "/usr/lib/python3.9/site-packages/nova/compute/manager.py", line
  9935, in _update_available_resource_for_node\n
  self.rt.update_available_resource(context, nodename,\n', '  File
  "/usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py",
  line 896, in update_available_resource\n
  self._update_available_resource(context, resources,
  startup=startup)\n', '  File "/usr/lib/python3.9/site-
  packages/oslo_concurrency/lockutils.py", line 360, in inner\n
  return f(*args, **kwargs)\n', '  File "/usr/lib/python3.9/site-
  packages/nova/compute/resource_tracker.py", line 1003, in
  _update_available_resource\n    self._update(context, cn,
  startup=startup)\n', '  File "/usr/lib/python3.9/site-
  packages/nova/compute/resource_tracker.py", line 1247, in _update\n
  LOG.debug(\'artom: %s\', traceback.format_stack())\n'] _update
  /usr/lib/python3.9/site-packages/nova/compute/resource_tracker.py:1247

  Not included in the above for brevity is the log line showing what was
  actually being saved, you'll just have to trust me on this one ;)

  [1] https://bugzilla.redhat.com/show_bug.cgi?id=1883554
  [2] https://review.opendev.org/c/openstack/whitebox-tempest-plugin/+/851447

To manage notifications about this bug go to:
https://bugs.launchpad.net/nova/+bug/1995153/+subscriptions
References

[Bug 1995153] [NEW] `socket` PCI NUMA policy doesn't work if another instance is booted first on the same host
From: Artom Lifshitz, 2022-10-28