[gpfsug-discuss] swapped_warn event

[Jonathan Buzzard]

On 17/02/2026 08:39, Mathias Dietz wrote:

> Hi Jonathan,
> 
> Thank you for your question regarding the swapped_warn event.
> 
> In IBM Spectrum Scale, different categories of health events serve 
> different purposes:
> 
> *State-change events (such as degraded, failed, etc.)* indicate actual 
> problems or failures. These events typically resolve automatically; only 
> a small number require manual resolution using the mmhealth event 
> resolve command.
> *TIP events* provide recommendations or highlight minor issues, 
> best‑practice deviations, or configuration optimizations. These events 
> do not indicate a failure. If desired, TIP events can be permanently 
> hidden using mmhealth event hide, and once hidden they will not appear 
> again.
> 
> The *swapped_warn *event falls into the *TIP event category*.
> 
> You are correct that a small amount of swap usage is not necessarily 
> problematic, especially when plenty of RAM is available and the system 
> is not actively swapping. However, we have seen multiple real-world 
> cases where even moderate swap usage negatively impacted system 
> responsiveness and overall Scale performance.
 > > Because of this, the TIP is intended to help users who want to extract
> maximum performance from their systems. If swap usage is not a concern 
> in your environment, you can safely hide the TIP and continue operating 
> normally.
>

I agree with everything you have said, but the problem is that GPFS is 
warning about swap usage even though no swap usage is occurring.

That there is usage of swap space is not, and has not been, for over 
quarter of a century now, a valid measure of whether the Linux kernel is 
paging memory in/out of disk.

You can determine if the kernel is actually paging memory to and from 
disk with the vmstat command. Checking the nodes reporting the 
swapped_warn tip shows that they have not paged a single memory page 
since the last reboot, despite swap usage.

As mentioned earlier, the Linux kernel will preemptively write memory 
pages to swap space as a precaution, so swap usage does not reflect what 
the GPFS developers think it does.

I don't want to hide the tip because it would be useful if it actually 
did what it said on the tin. The problem is that, at the moment, due to 
faulty assumptions, the tip is as useful as a chocolate teapot.

The TL;DR is that GPFS needs to switch to using a valid measure of paging.


JAB.

-- 
Jonathan A. Buzzard                         Tel: +44141-5483420
HPC System Administrator, ARCHIE-WeSt.
University of Strathclyde, John Anderson Building, Glasgow. G4 0NG