Re: [RFC PATCH v2 0/5] mm/vmscan: adaptive multi-threaded kswapd for NUMA-aware reclaim

[Shakeel Butt]

On Tue, Nov 03, 2026 at 04:15:44PM -0800, Dave Hansen wrote:
> On 11/3/26 15:41, Ababneh, Ehab wrote:
> > I would appreciate your thoughts on how we might achieve a similar
> > performance improvement without increasing the number of kswapd
> > threads.
> 
> It also just plain doesn't seem like folks are very interested in
> optimizations in this area at the moment. Feedback that's basically, "go
> optimize something else" would also be very welcome.

Hi Ehab and Dave,

My appologies for late response. I have not gotten the chance to look
deeper into the series as I am stuck in something else but I am very
much interested in this (line of) work.

I have discussed with Johannes about multithreaded kswapds but the
motivation is more about scaling kswapd threads corresponding to the
number of cores on the node. We are getting CPU with hundred of cores per
numa code and kswapd is already too late or slow to keep up with
allocations happening on all of those cores. So, I definitely think we
need multiple workers doing the background reclaim.

Please give me sometime, I will organize my thoughts on how we should
aim to tackle this challenge. I will also ask Johannes to provide his
take on the issue.

Thanks for pushing this work.
Shakeel