[Beowulf] many cores and ib

Patrick Geoffray patrick at myri.com
Mon May 12 18:44:20 PDT 2008


Gilad Shainer wrote:
> My apologizes. I meant the MPI includes an option to collect several MPI
> messages into one network message. For applications cases, sometimes it
> helps with performance and sometimes it does not. OSU have shown both
> cases, and every user can decide what works best for him.

Please, point me to a document/paper/report that shows that message 
coalescing improves real applications. Even from Panda, that's fine.

> As long as you use the same code and the same benchmark (and the same
> platform), you can use any benchmark to compare between different
> network devices. You can claim that if you want to count the network
> messages and not the MPI messages, using this way to do it will give you
> the wrong number, but if you want to compare between 2 interconnects,
> this is valid. 

The message rate micro-benchmark tries to measure the capacity of the 
interconnect to send messages to a large number of nodes, a large number 
of local processes (multi-core), or both. In short, it tries measures 
the gap between two messages on the wire (if you don't know about the 
LogP model, read a book).

If the MPI implementation changes the intent of the micro-benchmark, and 
abuse the fact that it runs on 2 nodes only to be convenient, then it 
makes the metric irrelevant because it cannot be used to estimate the 
performance at scale.

If you want to compare, do it with Open-MPI then, it does not yet 
implement message coalescing. Same benchmark, same platform, same MPI. 
That should work, no ?

> This becomes personal now... :-) you don't need to be angry all the
> time. Bad for your health. I could have replay back mentioning all the

People often mistake anger for passion. I have to say, your case 
(marketing abusing technical info) is a very special treat. I handle it 
very much like chocolate: a little bit at a time, only once a week. 
Don't worry for my health, though I appreciate your concern.

> FUD and misleading information you provide wherever you go, and we saw
> couple of examples just few weeks ago, but I wont.   

I guess you are referring to the technical panel at the IDC forum in 
Norfolk and my correction about the (in)compatibility between iWARP and 
IB. If so, please explain how the last-byte-written-last semantic in 
iWARP and IB are the same. Please, give me a sweet one, with something 
chewy/gooey in the middle. I just love those :-)


More information about the Beowulf mailing list