[openib-general] ibv_get_device_list() question

Jeff Squyres jsquyres at cisco.com
Tue Jan 2 17:38:35 PST 2007


Ah!  That *would* be perfect, but I guess the RHEL4 kernel doesn't  
have this stuff (there's no "bus" or "local_cpus" entries in the  
device/ subdir).

Failing that, is <device_name> guaranteed to be the same across  
reboots, and/or the info returned by ibv_get_device_info() guaranteed  
to be in the same order?  (I suppose that I can use  
ibv_get_device_name() to match potentially unordered info returned by  
ibv_get_device_info(), if necessary -- as long as the name is  
guaranteed to be the same across reboots)


On Jan 2, 2007, at 6:57 PM, Robert Walsh wrote:

> Jeff Squyres wrote:
>> We're debating internally in the OMPI group about providing access  
>> to  specific NICs/HCAs, especially on NUMA machines where  
>> different NICs/ HCAs are local to different busses / processors.   
>> So I have the  following questions:
>> 1. Does the ibv API provide any information about devices'  
>> physical  location with regards to busses / processors?  (I doubt  
>> it, but I'm  just asking to be sure)
>
> Don't know whether you can get this directly, but you can get it  
> indirectly by looking in /sys/class/infiniband/<device_name>/ 
> device. bus points at the bus.  local_cpus contains the nearby cpu  
> mask.  This what you're looking for?


-- 
Jeff Squyres
Server Virtualization Business Unit
Cisco Systems





More information about the general mailing list