[openib-general] in need of a simple ulp

Bernhard Fischer blist at aon.at
Fri May 20 14:33:17 PDT 2005


On Fri, May 20, 2005 at 01:40:42PM -0700, Sean Hefty wrote:
>Jeff Carr wrote:
>>There must also be some limit to how many cqe's you can allocate with 
>>ib_post_recv(). (?)
>
>Each receive posted will generate a cqe.  That limit is controlled by the 
>message_count variable.
>
>>There must be some way to delete/free these? They don't get re-used I 
>>take it? Surely it wasn't intended that ib_post_recv() be initially run 
>>for each transfer expected in the lifetime of the connection. :)
>
>There needs to be one cqe reserved for each possible completion that can be 
>outstanding at one time.  If you know that the number of completions will 
>be high initially, and then drop to a smaller value, there is a verb that 
>can be used to resize the CQ.  (I don't believe it is implemented in the 
>openib stack at this time.)
>
>>There must also be some information about what is known about these 
>>cqe's. How do we know if one of them was used for a transfer from the 
>>server to the client or from the client letting the server know the 
>>transfer was recieved?
>
>In this case you won't know what's on a cqe until you poll the completion. 
>An alternative way of structuring the code is to create two CQs, and tie 
>one to the send queue, the other to the receive queue.

hmz, yes, see "Quick unrelated question:" in
http://openib.org/pipermail/openib-general/2005-May/006796.html

I do have ...s_cq and r_cq. For two nodes, using cqe=64 does obviously
work, but when i've got, say, a setup where 786 nodes each have r and s?

This currently w{on't,ill in no way way} work (for me)!? Help!!

Also, Sean, any thought on the qp limit question referenced there?

TIA



More information about the general mailing list