Re: [myrinet] Problems sending 1k and 2k messages

sounds to me like you are seeing some problems 
with cache-coherency and the receive queue.

try running a large "make" (like of the linux kernel)
or other big memory-hog program to see if that
makes the modified gm_debug program
work fine.

If so, then I'd suspect that your OS/hardware are not
making the DMAs from the LANai into the receive queue
flush the cached values.



=> We have modified gm_basic in order to send messages of user defined
=> length (and user defined number of messages). We experienced problems
=> when sending 1024bytes (or 2048bytes) messages (i.e. 8 messages).
=> Sometimes the host receives all the GM_PEER_RECV_EVENT (i.e. 8) but not
=> all the GM_SENT_EVENT (i.e. 6) and then it receives GM_NO_RECV_EVENT all
=> the time. The behaviour of the test program changes when we activate
=> DEBUG (or when we include more printfs), in this case it runs perfectly.
=> 
=> Now gm_allsize works fine.
=> 
=> We understand that since changing gm_basic program can introduce some
=> race conditions, solving the problem can be very difficult. If you are
=> interested in, we can pass you the modifications we have done in order
=> to detect the possible bug.
=> 
=> Jose Flich and Salva Coll
=> 
=> P.S. We are runing GM version 1.2pre16 on a SuSE 6.1 Linux distribution
=> with kernel 2.0.36 with PII processors.
=> 
=>