On May 15, 2008, at 5:55 AM, Torje Henriksen wrote:
> 1. In the BTL SM component (both in "ompi_info --param btl sm" and the
> source), there is mention of RDMA. Is RDMA used in ths component? Is
> any special hardware needed for using it?
No and no. The SM BTL uses a copyout/copyin mechanism (copy out to
shared memory, copy in to the target process) -- it does not directly
copy from one process to another.
The OB1 PML uses "RDMA-like" semantics; I'm guessing that's what it's
referring to.
> 2. As for eager vs non-eager messages. Is the only difference between
> these the amount of memory allocated for a fragment? Is there any
> rendevouz protocol used for large messages or anything like that?
The rendesvouz protocol is enforced by the upper level PML (OB1) -- it
is basically the lower-layer-independent "protocol engine". OB1
decides the exact protocol to use (short, long, pipelined, etc.), and
the lower later BTLs decide how to get the bits there.
> The module based architecture is great, but sometimes I find it hard
> to figure out what is happening :)
Indeed. This is definitely a tradeoff when you only see small pieces
of the puzzle at a time. :-\
--
Jeff Squyres
Cisco Systems
|
![[Flag]](../../../../../community/mirrors/flag-us.gif){kind=small}