By: Howard Chu (hyc.delete@this.symas.com), September 16, 2007 1:14 pm
Room: Moderated Discussions
David Kanter (dkanter@realworldtech.com) on 9/15/07 wrote:
---------------------------
>unknown (un@known.net) on 9/15/07 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 9/14/07 wrote:
>>>It's very straight forward - if different regions of memory have different latencies
>>>then the system is NUMA. If all memory has the same latency, then the system is UMA. It cannot be both.
>>
>>By that logic, turning on Node Interleave on the Opteron >would make it UMA (?)
Eh, I wouldn't go that far.
>Yes, that's right. I believe the idea behind node interleave is to have a mode
>that works effectively for operating systems with no, or poor NUMA support.
The main reason to use it is when you really have large data structures that need to be accessed by every thread in the system. E.g. a large multithreaded database server. In these situations, you don't want to pin data to a single node because it will cause hotspots/bottlenecks. I.e., node interleaving is a load balancing approach, and it has nothing to do with how good the OS's NUMA support is.
---------------------------
>unknown (un@known.net) on 9/15/07 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 9/14/07 wrote:
>>>It's very straight forward - if different regions of memory have different latencies
>>>then the system is NUMA. If all memory has the same latency, then the system is UMA. It cannot be both.
>>
>>By that logic, turning on Node Interleave on the Opteron >would make it UMA (?)
Eh, I wouldn't go that far.
>Yes, that's right. I believe the idea behind node interleave is to have a mode
>that works effectively for operating systems with no, or poor NUMA support.
The main reason to use it is when you really have large data structures that need to be accessed by every thread in the system. E.g. a large multithreaded database server. In these situations, you don't want to pin data to a single node because it will cause hotspots/bottlenecks. I.e., node interleaving is a load balancing approach, and it has nothing to do with how good the OS's NUMA support is.