New: Bug from the mm subsystem involving X (fwd)
Quoted 3 lines · Martin J. Bligh
If we really want to do good testing, we should make a fake NUMA config that can run a 4x SMP box as fake NUMA, with half the memory in each "node" and half the processors ... but I never got around to coding that ;-)
I have such a patch for x86-64 if anybody is interested in that.
x86-64 low level NUMA is quite different from IA32 NUMA though so it would be a bit difficult to port.
-Andi
--Andi Kleen <[email protected]> wrote (on Saturday, February 07, 2004 04:54:03 +0100):
Quoted 10 lines
Quoted 3 lines · Martin J. Bligh
If we really want to do good testing, we should make a fake NUMA config that can run a 4x SMP box as fake NUMA, with half the memory in each "node" and half the processors ... but I never got around to coding that ;-)
I have such a patch for x86-64 if anybody is interested in that.
x86-64 low level NUMA is quite different from IA32 NUMA though so it would be a bit difficult to port.
Not quite sure what you mean ... I was driving at pretending an SMP box was NUMA ... but the x86_64 is already NUMA ... are you grouping nodes together into single nodes with 2 cpus each?
What might be intriguing is to use Nick's domains stuff to create a heirarchy for the scheduler where we have 1 cpu nodes and 2 cpu nodes above that, but still keep the normal NUMA stuff flat for mem allocation. What might be interesting is a heirarchy where if this is the HT connections of cpu layouts:
1 --- 2
| | | | | |
3 --- 4
then domains of (1,2,3) (2,3,4) (1,3,4) (1 2 4), with a view to restricting the "double hop" traffic as much as possible. But I'm not sure the domains code copes with multiple overlapping domains - Nick?
Andi, do you already set up the mem allocation fallback zonelists like that?
M.
Quoted 3 lines · On Fri, 06 Feb 2004 20:49:40 -0800 "Martin J. Bligh" <[email protected]> wrote:
Not quite sure what you mean ... I was driving at pretending an SMP box was NUMA ... but the
x86_64is already NUMA ... are you grouping nodes together into single nodes with 2 cpus each?
There are Opteron boxes which are not NUMA. Or rather they are NUMA, but only have a single node. Some of the cheaper boards only connect the DIMM slots to a single CPU, which gives you only a single node even with two CPUs. One of the test machines I have here is of this type.
It's also useful for testing on simulators.
Quoted 1 line
Andi, do you already set up the mem allocation fallback zonelists like that?
I don't do anything special, it's all generic page_alloc.c logic.
-Andi
Quoted 23 lines · Martin J. Bligh
--Andi Kleen <[email protected]> wrote (on Saturday, February 07, 2004 04:54:03 +0100):
Quoted 11 lines
Quoted 4 lines · Martin J. Bligh
If we really want to do good testing, we should make a fake NUMA config that can run a 4x SMP box as fake NUMA, with half the memory in each "node" and half the processors ... but I never got around to coding that ;-)
I have such a patch for x86-64 if anybody is interested in that.
x86-64 low level NUMA is quite different from IA32 NUMA though so it would be a bit difficult to port.
Not quite sure what you mean ... I was driving at pretending an SMP box was NUMA ... but the
x86_64is already NUMA ... are you grouping nodes together into single nodes with 2 cpus each?What might be intriguing is to use Nick's domains stuff to create a heirarchy
for the scheduler where we have 1 cpu nodes and 2 cpu nodes above that, but
still keep the normal NUMA stuff flat for mem allocation. What might be interesting is a heirarchy where if this is the HT connections of cpu layouts:
Quoted 10 lines
1 --- 2
| | | | | |3 --- 4
then domains of (1,2,3) (2,3,4) (1,3,4) (1 2 4), with a view to restricting the "double hop" traffic as much as possible. But I'm not sure the domains code copes with multiple overlapping domains - Nick?
Yes it can do ring topologies like this. I'm pretty sure it can do just about any sort of topology although this is one that I sat down and drew when designing it.
You can technically restrict a double hop, but after you move, say, clockwise once, you might just as easily be moved clockwise again. The only way to restrict this is with some kind of home domain thing.
Quoted 12 lines
Quoted 11 lines
1 --- 2
| | | | | |3 --- 4
then domains of (1,2,3) (2,3,4) (1,3,4) (1 2 4), with a view to restricting the "double hop" traffic as much as possible. But I'm not sure the domains code copes with multiple overlapping domains - Nick?
Quoted 7 lines
Yes it can do ring topologies like this. I'm pretty sure it can do just about any sort of topology although this is one that I sat down and drew when designing it.
You can technically restrict a double hop, but after you move, say, clockwise once, you might just as easily be moved clockwise again. The only way to restrict this is with some kind of home domain thing.
Well, this doesn't ban it, but requiring the double migrate will curtail it somewhat, which is better than nothing.
M.