Message-ID: <41cc8a35-81fb-b890-f963-8dc9524a54b0@intel.com>
Date:   Tue, 12 Sep 2023 11:31:50 -0700
From:   Reinette Chatre <reinette.chatre@...el.com>
To:     Tony Luck <tony.luck@...el.com>
CC:     Fenghua Yu <fenghua.yu@...el.com>,
        Peter Newman <peternewman@...gle.com>,
        Jonathan Corbet <corbet@....net>,
        Shuah Khan <skhan@...uxfoundation.org>, <x86@...nel.org>,
        Shaopeng Tan <tan.shaopeng@...itsu.com>,
        James Morse <james.morse@....com>,
        Jamie Iles <quic_jiles@...cinc.com>,
        Babu Moger <babu.moger@....com>,
        Randy Dunlap <rdunlap@...radead.org>,
        <linux-kernel@...r.kernel.org>, <linux-doc@...r.kernel.org>,
        <patches@...ts.linux.dev>
Subject: Re: [PATCH v5 0/8] Add support for Sub-NUMA cluster (SNC) systems

Hi Tony,

On 9/12/2023 10:45 AM, Tony Luck wrote:
> On Tue, Sep 12, 2023 at 10:13:31AM -0700, Reinette Chatre wrote:
>> Hi Tony,
>>
>> On 9/12/2023 9:01 AM, Tony Luck wrote:
>>> On Mon, Sep 11, 2023 at 01:23:35PM -0700, Reinette Chatre wrote:
>>>> Hi Tony,
>>>>
>>>> On 8/29/2023 4:44 PM, Tony Luck wrote:
>>>>> The Sub-NUMA cluster feature on some Intel processors partitions
>>>>> the CPUs that share an L3 cache into two or more sets. This plays
>>>>> havoc with the Resource Director Technology (RDT) monitoring features.
>>>>> Prior to this patch Intel has advised that SNC and RDT are incompatible.
>>>>>
>>>>> Some of these CPU support an MSR that can partition the RMID
>>>>> counters in the same way. This allows for monitoring features
>>>>> to be used (with the caveat that memory accesses between different
>>>>> SNC NUMA nodes may still not be counted accuratlely.
>>>>
>>>> Same typo as in V4.
>>>
>>> Sorry. Will fix and re-post.
>>>
>>>>>
>>>>> Note that this patch series improves resctrl reporting considerably
>>>>> on systems with SNC enabled, but there will still be some anomalies
>>>>> for processes accessing memory from other sub-NUMA nodes.
>>>>
>>>> I have the same question as with V4 that was not answered in that email
>>>> thread nor in this new version.
>>>> https://lore.kernel.org/lkml/e350514e-76ed-14ea-3e74-c0852658182f@intel.com/
>>>
>>> Non-SNC systems already have an issue when reporting memory bandwidth
>>> for a task that Linux may migrate the task to a CPU on a different node
>>> which means that logging for that task will also move to different files
>>> in the mon_data/mon_L3_*/ for the new node.
>>
>> It is not obvious to me that this is an issue. From what I understand
>> the data remains accurate.
>>
>> How does this map to the earlier "may still not be counted
>> accurately"? 
> 
> Yes, the data is accurate. But a naive application reading the wrong
> files from mon_data will not see the accurate data. Without SNC users
> may only see this issue rarely as Linux tries hard to not migrate
> processes to other NUMA nodes or L3 cache domains. But with SNC enabled
> this is no longer the case. the ACPI SLIT distance of 0xC is below the
> threshold that Linux checks for "is the target CPU for a migration far away"
> so migration to other SNC nodes may be quite common.

I would like to recommend that you take care not to present this work
using uncertain terms like "may not be counted accurately" or "there will
still be some anomalies". If I understand correctly there is no uncertainty.
When that "naive application" reads the wrong files then I think it can be
considered a usage error and should not be documented as an issue with the
counters. I understand that the user space requirements are not obvious,
and there should be guidance. 

> I can move this out of the cover letter and provide guidance/warnings
> in the patch to Documentation/arch/x86/resctrl.rst

Yes, I do think this will be very helpful.

> 
>>
>>>
>>> With SNC enabled, migration between NUMA nodes on the same socket may happen
>>> much more frequently because:
>>> 1) The CPUs on the other NUMA nodes in the socket are in the same Linux
>>>    L3 cache domain. So Linux regard the migration as "cheap".
>>> 2) The ACPI SLIT table on SNC enabled systems may also report the
>>>    latency for remote access to another NUMA node on the same socket
>>>    as significantly lower than the latency for cross-socket access. On
>>>    my test system the SLIT distance for same socket nodes is 0xC,
>>>    compared to 0x15 for cross-socket distance. This will also lead
>>>    to Linux being more likely to migrate a task to a CPU on another
>>>    SNC NUMA node in the same socket.
>>>
>>> To avoid migration issues, users may use sched_setaffinity(2) to bind
>>> tasks to the subset of CPUs that share an SNC NUMA node.
>>>
>>> I can write this up in a new cover letter.
>>>
>>>> I stop my review of this series here.
>>>
>>> Reinette
>>>
>>> Should I repost the whole series as v6 with the new cover letter. The
>>> only change to the patches so far is to the selftest reported by
>>> Shaopeng Tan[1].
>>>
>>
>> Is this an assurance that the cover letter in no way reflects how 
>> feedback was addressed in the rest of this series?
> 
> My track record here is far from perfect. I believe I addressed all
> the issues you raised. But it's very possible that I may have missed
> some, or misintepreted the concerns you raised.

This is a familiar response from you that just puts the burden back
on me to go dig out previous discussions. This will be the last time.

Reinette

Powered by blists - more mailing lists