Re: [PATCH v7 00/13] fold per-CPU vmstats remotely

[Andrew Theurer]

> On Apr 19, 2023, at 6:15 AM, Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote:
> 
> On Wed, Apr 19, 2023 at 08:14:09AM -0300, Marcelo Tosatti wrote:
>> On Tue, Apr 18, 2023 at 03:02:00PM -0700, Andrew Morton wrote:
>>> On Mon, 20 Mar 2023 15:03:32 -0300 Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote:
>>> 
>>>> This patch series addresses the following two problems:
>>>> 
>>>> 1. A customer provided evidence indicating that a process
>>>>   was stalled in direct reclaim:
>>>> 
>>>> ...
>>>> 
>>>> 2. With a task that busy loops on a given CPU,
>>>>    the kworker interruption to execute vmstat_update
>>>>    is undesired and may exceed latency thresholds
>>>>    for certain applications.
>>>> 
>>> 
>>> I don't think I'll be sending this upstream in the next merge window. 
>>> Because it isn't clear that the added complexity in vmstat handling is
>>> justified.
>> 
>> From my POV this is an incorrect statement (that the complexity in
>> vmstat handling is not justified).
>> 
>> Andrew, this is the 3rd attempt to fix this problem:
>> 
>> First try:  https://lore.kernel.org/lkml/20220127173037.318440631@fedora.localdomain/
>> 
>> Second try: https://patchew.org/linux/20230105125218.031928326@xxxxxxxxxx/
>> 
>> Third try: syncing vmstats remotely from vmstat_shepherd (this
>> patchset).
>> 
>> And also, can you please explain: what is so complicated about the
>> vmstat handling? cmpxchg has been around and is used all over the
>> kernel, and nobody considers "excessively complicated".
>> 
>>> - Michal's request for more clarity on the end-user requirements
>>>  seems reasonable.
>> 
>> And i explained to Michal in great detail where the end-user 
>> requirements come from. For virtualized workloads, there are two
>> types of use-cases:
>> 
>> 1) For example, for the MAC scheduler processing must occur every 1ms,
>> and a certain amount of computation takes place (and must finish before
>> the next 1ms timeframe). A > 50us latency spike as observed by cyclictest
>> is considered a "failure".
>> 
>> I showed him a 7us trace caused by, and explained that will extend to >
>> 50us in the case of virtualized vCPU.
>> 
>> 2) PLCs. These workloads will also suffer > 50us latency spikes
>> which is undesirable.
>> 
>> Can you please explain what additional clarity is required?
>> 
>> RH's performance team, for example, has been performing packet
>> latency tests and waiting for this issue to be fixed for about 2
>> years now.
>> 
>> Andrew Theurer, can you please explain what problem is the vmstat_work
>> interruption causing in your testing?
> 
> +CC Andrew.

Nearly every telco we work with for 5G RAN is demanding <20 usec CPU latency as measured by cyclictest & oslat.  We cannot achieve under 20 usec with the vmstats interruption.

-Andrew