Kernel : MAX_SMP vs NR_CPUS

[krumpf]

Hi all,
here's something I've been wondering about for a while : what's the difference between setting the numbers of CPUs manually with NR_CPUS, and using the MAX_SMP option ?
The NR_CPUS help says "This is purely to save memory: each supported CPU adds about 8KB to the kernel image.", while the MAX_SMP says "If unsure, say N.". From what I've read here & there, MAX_SMP is made for really large number of cpus (up to 8192), which I guess applies for servers/datacenters.
So I'd like to know what are the pros and cons for each option ? Is it bad practice to use MAX_SMP on a desktop computer instead of NR_CPUS ?
_________________
Dragon Princess Music Games Heroes and villains

[pietinger]

Usually you can choose between 2 and 512 CPUs (see <Help> of CONFIG_NR_CPUS) ... IF you need more THEN you must enable CONFIG_MAXSMP.

If you enable CONFIG_MAXSMP you have automatically the maximum of 8192 CPUs ... AND ... you cannot select CONFIG_NR_CPUS futhermore, because you will see CONFIG_NR_CPUS only if ->

[krumpf]

Thanks pietinger, but you didn't really answer my question.

How to explain... My CPU is a 8cores/16threads, so I suppose the optimal NR_CPUS would be 16. Let's say I set it on 64, as I understand it, it will inflate the kernel image by (64-16)*8KB = 384 KB, and during boot the kernel will "discard" the cpus that are not detected.
But what are the consequences of using MAX_SMP instead of NR_CPUS ? Would that increase the kernel size by (8192-16)*8KB almost 64MB ? Or does the compiler detects the correct number of cpus when building, and automatically set it ?
_________________
Dragon Princess Music Games Heroes and villains

[Hu]

A kernel built for N CPUs can be run on a system with M CPUs, regardless of whether N < M, N = M, or N > M. However, if N < M, you only get use of N CPUs, meaning the (M - N) CPUs go to waste. If N > M, you get use of all M CPUs that you actually have, but the kernel sets aside some extra memory as if you had N CPUs. Thus, setting N > M is a waste of memory, and the greater the delta, the greater the waste. Per pietinger's post, MAXSMP forces N = 8192, thus giving you maximal memory waste. You should always pick N such that it is at least M, but preferably no bigger. If you want a single kernel that runs on multiple different systems, you might set N to max(system1, system2, ...), so that all systems can use all their CPUs, and the less capable systems just waste a bit of memory.

[krumpf]

I think I finally understand, it affects the kernel image once it's decompressed in RAM, not the image written on disk.
Thanks for the explanations.

(Feels like I have the hardest time to figure out the simplest things. Now if you excuse me, my shame and me are going to hide in some deep hole.)
_________________
Dragon Princess Music Games Heroes and villains

[pietinger]