[openib-general] Re: Re: [PATCH] [RFC] - example user mode rdmaping/pongprogram using CMA

[Caitlin Bestler]

openib-general-bounces at openib.org wrote:
>>> By the way, in case its not clear:  The SEND/RECV exchanges are done
>>> just to advertise source and sink memory regions, and to indicate
>>> completion of rdma read and write operations to the peer.  The
>>> "ping/pong" data is transferred with rdma read and write operations.
>>> 
>>> Thanks for the feedback!
>>> 
>> 
>> Code tends to get copied around ... its easy to imagine someone
>> copying this and measuring the send latency. Just posting many WRs in
>> the initialization sequence, with no other code changes, will fix
>> this problem. 
>> 
> 
> Each "ping/pong" iteration with rping is composed of 2 sends
> on the client side, 2 sends on the server side, plus 1 rdma
> read and 1 rdma write on the server side.
> 
> Again, latency performance (or any performance) isn't a goal
> of this program.  Testing CMA, CQ and CMA event
> notifications, and send/recv/rr/rw are the goals.
> 
> 
> snipit from the patch:
> 
> +/*
> + * rping "ping/pong" loop:
> + *     client sends source rkey/addr/len
> + *     server receives source rkey/add/len
> + *     server rdma reads "ping" data from source
> + *     server sends "go ahead" on rdma read completion
> + *     client sends sink rkey/addr/len
> + *     server receives sink rkey/addr/len
> + *     server rdma writes "pong" data to sink
> + *     server sends "go ahead" on rdma write completion + *    
> <repeat loop> + */
> 

Why does the server send "go ahead" after rdma write completion?
It should be able to just post the send after posting the rdma
write without waiting. When the rdma write completes has no
device/transport independent meaning.