OSDev.org

If someone want to create new operating system kernel (ie for Linux, BSD), is there any difference between Kernel source code for:

1. A motherboard with 2 sockets and 2 cores.
2. A motherboard with 2 sockets and 4 cores (2 cores + 2 cores)
3. A motherboard with 2 sockets and 3 cores (2 cores + 1 core)
4. A motherboard with 1 socket and 4 cores.

Hi,

How you handle this depends on your personal preference. If you treat each core as a resource, there is no reason why you need a different kernel for each system.

I wonder if you are asking this because of the core limit on various versions of Windows? If so, as far as I am aware, that is a purely commercial decision (not sure what Vista does, but it used to be that the 'high end' server OS's supported up to 32 cores, whereas the lower end desktop OS's only supported a single core. NT had various stages in-between these extremes).

I have not added multi-core support for my OS, but I understand that all you need for multi-core support is a different scheduler instance for each core (and care with using locks when appropriate!).

Cheers,
Adam

Multiple core processors are basically multiple processors sharing the same buss and therefor act like buss starved multiple socket systems.
Even hyperthreading looks like 2 processors even though its one cpu pretending to be 2.
There is no difference other than performance.

You can try to specifically support NUMA architectures (google it). From what I know of processor architecture, they only occur with multi-socketed systems. basically, its just a performance hit if you optimize for NUMA when your system isn't or when your system is NUMA and you don't support it. That'd be the closest to "different code" I can think of. However if you choose at run time which method to use, there is no difference in code at all - its just that a part of it will not be used.

AFAIK, linux has it as a compile-time option.

Hi,



For all possible 80x86 computers (with any mixture of hyper-threading, multi-core, multi-socket and NUMA) an OS can treat all logical CPUs as identical with no special treatment. It's only when you start optimizing that things get tricky.

For NUMA, the main thing to optimise is memory accesses. For example, you might want to use processor affinity to restrict processes to NUMA domains, and then allocate physical pages for those processes that are "close" to the CPU/s in that NUMA domain. The idea is to reduce the chance of processes accessing memory that isn't "close". There's also other things you can do though (e.g. load balancing, and optimizing the kernel's memory usage).

For hyper-threading, the load on one logical CPU effects the performance of other logical CPUs in the same core. Because of this you might want to optimise CPU usage. For example, if you've got 2 cores with 2 logical CPUs per core and 2 of the logical CPUs are idle, then it's better to have one logical CPU per core idle than it is to have both logical CPUs in one core idle and both logical CPUs in another core busy.

In all cases, you might prefer to run threads on the same CPU that they ran on last time, as the thread's data might still be in that CPUs cache.

For both hyper-threading and multi-core, some caches may be shared between CPUs. For an example, with a multi-core CPU that supports hyper-threading the L1 caches might be shared by logical CPUs in the same core and the L3 cache might be shared by all logical CPUs in the chip. In this case you don't necessarily need to run the thread on the same CPU that it ran on last time to get the same benefits from cached data - you might get the same advantage running the thread on a different logical CPU in the same core (or any CPU in the chip). In a similar way, there may also be advantages in running different threads that belong to the same process on the same chip or core (so that several threads running on seperate logical CPUs in the same chip can share data in that chip's cache). The same thinking can apply IPC (it might better for threads that communicate with each other to run on the same core or chip).

Of course there's other things you could do in certain situations, like not bothering to set MTTRs on all CPUs (the MTTRs are shared for hyper-threading, so you only need to set one logical CPU's MTTRs in each core and can skip the other logical CPUs), or optimizing IRQ handling.

Also, I'm not sure if any operating system actually does all of the above (I'd assume most might do some of it).


Cheers,

Brendan