Re: "True" Broadcasting for Myrinet

Hi Wilson,

On Mon, 2003-01-06 at 20:02, Wilson Fung (Hotmail) wrote:
> Does "True" Broadcasting really exist for Myrinet?

What is "True" broadcasting ? Hardware supported ? At which level
(switch, NIC) ?

> I mean 
> broadcasting that is not accomplished by sending the same 
> message to all the members one by one.

Gm does not expose a multicast functionality in its API. Myrinet does
not provide hardware support for broadcast at the switch level. However,
the GM driver already implement a simple broadcast support in the
firmware for IP, when the same payload is sent to each destination in a
loop, but from the NIC itself, not from the host. It does not scale
Large, but it's low host overhead.

Now, the right way to do broadcast in Myrinet is to implement a spanning
tree between NICs: it scales very well and it's low host overhead. There
is a lot of literature on the subject, a lot of work from Folks in
Amsterdam (Henri Bal's group) in LFC and other results from Panda's
group at Ohio State.

I would also recommend my Phd thesis (chapter 3), but it's in French and
I sure will never take the time to translate it :-)

Hope it helps.

Patrick
-- 

Patrick Geoffray, Phd
Myricom, Inc.
http://www.myri.com