does having two xeon quad channel processors effectively give you 2x quad channel meaning eight channel ram? an entirely extra order of magnitude from the processing side as compared to single quad channel? can it be thought of that way? I don't see why not.
>>52349655
No.
>>52349655
Yes.
>>52349655
Maybe.
>>52349655
Maybe.
>>52349655
Probably, maybe.
>>52349655
In theory.
One way or the other.
>>52349655
Gentoo Does this natively
>>52349655
Might do.
>>52349655
Sometimes.
>>52349655
Can you repeat the question?
>>52349655
It worked once
>>52349655
No, you have two separate pools of RAM. If one CPU has to access data on the other CPU's RAM it has to go through the QPI interconnect.
Look up NUMA in wikipedia.
>>52349655
Seasonic will release this technology in 2004. Can't wait.
>>52349809
You're not the boss of me now, you're not the boss of me now.
You get 4 channels per socket. So each socket is addressing ITS memory in quad channel. NUMA doesn't actually share ram, its weird. The sockets talk on a NUMA link thing between each other which is a different speed.
Each socket is fully asynchronous form the other.
The MMU's just copy memory from one node to the other across the link, but all memory shares the same mapping, so the pointers never change.
So no 1 CPU can use all 8 channels, each is using 4 on their own, each with their own tasks. This is why memory separation between processes is really important. If you share to much memory you slow down a lot, and a HUGE amount in a multi-CPU configuration.
>>52349655
Yes, but having that much more bandwidth may not help you. For an individual core, there is plenty of bandwidth provided by its attached quad channels. If all other cores in a cpu package get busy, yes maybe your core in question can use the memory bandwidth from the other cpu die's memory controllers. But at that point you are helping to starve the second cpu's cores from memory bandwidth.
It is my understanding that the googles & facebooks of the world are now moving back to single cpu motherboards. For them it is purely a question of cost per cpu and bit of ram in their cloud; they will customize their software as required to leverage their hardware.
Where these dual cpu motherboards really come into their own is if you have software that can use lots of cores, so you can launch your program and it will spread itself across the machine and there is no need for distributed computing. We do some data processing on these machines and they scream.
Also the higher number of ram slots in the machine lets you get more GB (yes even TB) of ram into a single system image.
I used to run dual processor workstations exclusively; I may move back to them as performance increases per generation start to flatten out.
>>52349886
This is not entirely true. You can (or at least used to be able to) configure the machine to be non-NUMA, striping the ram across all cpu's memory controllers. Basically exactly like RAID-0. It was a bios option, available for performance tuning.
But even on a NUMA machine, where each cpu has its own local address space, if you malloc a huge chuck bigger than that, you will get data spread across both cpu's memory banks.
At some point I remember reading about migrating pages and threads to move data and threads around the NUMA machine to be closer to the ram/cpu currently being used in execution.
