Re: FreeBSD is now self-hosting on the UltraSPARC T1

On May 22, 2006, at 12:33 PM, m m wrote:

> On 5/22/06, Chad Leigh -- Shire.Net LLC <chad_at_shire.net> wrote:
>>
>> On May 21, 2006, at 7:54 PM, m m wrote:
>>
>> > While
>> > on topic, the Opterons aren't SMP either, and neither are the
>> > ht-Xeons...
>>
>> I would like t\o hear the rational for the Opterons (presumably the
>> dual core ones) not being SMP.  They have two independent operating
>> cores in one physical package.  Who cares how it is packaged?  I
>> would tend to agree with you on the ht-Xeon in terms of general
>> descriptions.  I do not know as well how the ht-xeon work as I don't
>> use any but it seems to me that the "SMP" moniker, at least in
>> FreeBSD, relate to how things are scheduled.
>
> SMP stands for "Symmetric MultiProcessing", which means that multiple
> processors have equal access latency to memory - typically
> accomplished by sitting the processors on a shared bus with memory.
> The MultiProcessor Opterons are _NOT_ SMP, they are _NUMA_ machines,
> "NonUniform Memory Access"; in the MP Opterons each processor has (or
> can have) its own "local" memory, which makes up only part of the
> shared address space.  When an Opteron accesses an address that is not
> in its "local" memory, it has to talk to a remote processor's memory,
> thereby incurring a different access latency.

Yes, but until FBSD differentiates and has different code for "SMP"  
and "NUMA" based systems,  then having support for them in the SMP  
branch seems appropriate.  In other words, even if not ideal, if  
there is no different code, there doesn't need to be a different  
section for the code.  I am no expert, but as far as I can tell, FBSD  
treats them like an SMP system and the NUMA stuff is done at the HW  
level.  Correct me if I am wrong.  Seems like the SMP moniker is a  
historical one, not a technical one :-)


---
Chad Leigh -- Shire.Net LLC
Your Web App and Email hosting provider
chad at shire.net