On Mon, Oct 04, 2021 at 04:05:23PM +0200, Christoph Lameter wrote:
> On Sun, 3 Oct 2021, Hyeonggon Yoo wrote:
>
> > As remote node allocation isn't that frequent and better avoided,
> > this patch tries to compromise some execution time for memory usage.
>
> Remote node allocation is typical for large memory loads and essential to
> balance the memory allocation over multiple nodes. Large NUMA systems
> exist because the memory requirements of the application cannot be
> satisfied by what a single node has to offer and usually code threads from
> multiple nodes may access remote memory.
>
> > Remove init_reap_node, next_reap_node and related variables
> > introduced by commit 8fce4d8e3b9e ("[PATCH] slab: Node rotor for freeing
> > alien caches and remote per cpu pages."), Because it has only one tree
> > for all remote nodes.
>
> Rotors to spread the memory allocated are essential for the performance
> of large memory loads! Memory must be spread evenly otherwise memory
> accesses will overload a single node.
Sorry for making a noise. It was because I misunderstood NUMA.
Thought it should be not frequent because kernel's default memory allocation policy was
allocating from local node.
It would have been better If I asked why it uses much memory first, not writing code.
I'll be more careful next time. Thank you for reviewing!
Thanks,
Hyeonggon
