Page 2 of 2

Re:Will your OS support multi-CPU?

Posted: Wed Apr 20, 2005 6:01 am
by smiddy
Well, since I am a hobby-ist at best, I had to vote with the reality of single CPU with the caveat that I will make available to best use of the cpu option that the OS is installed on...single user workstation-ish. That means if it is a hyper-threaded system it will take advantage of it, when I am able to get a handle on task scheduling, et al. So, that being said, I'm currently only past the hurdle of getting into protected mode, without interrupts (lots of polling going on). Some day I'll publish my criteria for it on it's own website, but alas that is down the road some.

Re:Will your OS support multi-CPU?

Posted: Tue Apr 26, 2005 9:39 pm
by Crazed123
I'm voting SMP only because I want to be simple and realistic. I'm currently working on exception handling, but when I get to a scheduler I want to have support for multiple CPUs if that's going to become the standard.

Re:Will your OS support multi-CPU?

Posted: Tue Apr 26, 2005 11:57 pm
by Colonel Kernel
I've been thinking about this some more and this questions popped up:

How exactly do you define NUMA? Is it NUMA in the sense of the Opteron, where there is still a single physical address space, but access times to memory not directly connected to a processor will be slower...? Or is it loosely coupled NUMA, where each "node" is a distinct physical address space with one processor, and processors need to communicate with each other via (hardware) message-passing?

Since most OS dev done by people on these forums is for x86, I'd assume the former and not the latter... In which case maybe I want to change my vote. :)

Re:Will your OS support multi-CPU?

Posted: Wed Apr 27, 2005 6:10 am
by Brendan
Hi,
Colonel Kernel wrote:How exactly do you define NUMA? Is it NUMA in the sense of the Opteron, where there is still a single physical address space, but access times to memory not directly connected to a processor will be slower...? Or is it loosely coupled NUMA, where each "node" is a distinct physical address space with one processor, and processors need to communicate with each other via (hardware) message-passing?

Since most OS dev done by people on these forums is for x86, I'd assume the former and not the latter... In which case maybe I want to change my vote. :)
I assumed any form of NUMA - either opteron style and/or otherwise.

I'm only familiar with opteron style (cache coherent) NUMA. I'm not aware of any computer system that uses non-cache coherent NUMA.

IMHO systems where physical address spaces are completely independant (e.g. blade servers AFAIK) aren't NUMA, but instead can be considered as multiple seperate computers within a single enclosure. These systems would use an OS that supports "clusters" (or a distrubuted OS), or could concievably use several completely different OSs.

In any case my knowledge of these systems (systems where physical address spaces are completely independant) is insufficient to know for sure. I would however appreciate any further information as my OS is intended to be a distributed OS - the initial design is for conventional 80x86/PCs but I hope to eventually extend this to other platforms (including blade servers and clusters of mixed architecture desktop machines).


Cheers,

Brendan