Hi Yinghai,
2013/8/24 Yinghai Lu <yinghai@xxxxxxxxxx>:
......
>> Do you guys think this is reasonable and acceptable?
>
> current boot flow that need to have all cpu and mem and pci discovered
> are not scalable.
>
> for numa system, we should boot system with cpu/mem/pci in PXM(X) only.
> and assume that PXM are not hot-removed later.
> Later during booting late stage hot add other PXM in parallel.
>
> That case, we could reduce boot time, and also could solve other PXM
> hotplug problem.
>
This is a good point, I think. Actually, I had a similar thinking
before. This can
solve the hotplug issue, and also the local node page table problem.
I found that the current kernel will do the memory hot-add procedure too in
later boot sequence. And we could get the following message:
System RAM resource ...... cannot be added
This message was from :
add_memory()
|->register_memory_resource()
because we have found and mapped all the memory in the system at early time.
But it is not easy to solve.
1. We still have to know how much memory and cpus the boot PXM has.
How could we know that ? SRAT again ?
2. The boot PXM could have little memory, and we meet the kexec and kdump
problem again.
3. I'm not quite sure is there any important benefit that the kernel
initializes all
the memory at beginning ?
And also, the memory hotplug schedule is very tough for us. We really want the
movablenode functionality could be available soon. And this idea could
be a long
way to go. So I also think this would be the next step.
Thanks.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>