Re: [myrinet] Re:

wyx wrote:
  > > 
  > > It is said on myricom's website that GM has a very good performance.Can anyone give me GM's performance compared to
  > > other comunication layers such as FM¡¢BIP,VMMC and so on?
  > > Thank you very much.

From the keyboard of Andrew Robbie:
  >                             However, if you want the fastest performance
  > then Trapeze (FreeBSD), BIP, PM, or VMMC-2 may be better.


Two other issues are scalability and reliability. Just because
something is fast, doesn't mean it scales (to hundreds and thousands
of nodes). GM seems to do OK up to some hundred nodes although we
have not put it through its paces on hundreds of nodes using real
applications. An example of a scalability limitation is a token that
has to be passed around in order for node pairs to get permission to
send to each other. Some systems use tokens like that for flow and
congestion control. However, if the number of token does not go up
as the system size goes up, the system doesn't scale. Cross-sectional
bandwidth, for example, will drop.

Myrinet is highly reliable, but it is not perfect. Applications
that run for hours or days on large machines will see CRC errors
and dropped or truncated packets. For the system to be reliable,
these errors have to be corrected at some level in the communication
stack. GM does that while others, namely BIP, do not.

Rolf



+++-+--+----+-------+------------+--------------------+------------------------
Rolf Riesen                              Email: [email protected] [email protected]
Scalable Computing Systems               [email protected]
Sandia National Laboratories             WWW:   http://www.cs.sandia.gov/~rolf
Org. 9223,  MS 1110                      Work:  (505) 845-7363
Albuquerque, NM  87185-1110              Fax:   (505) 845-7442