[comp.parallel] Broadcasting on the Hypercube Computers

wangjw@usceast.cs.scarolina.edu (Jingwen Wang) (11/28/90)

Dear Networkers,
   As you may know, many applications need multicast communications and
broadcast communications. However, todays hypercube computer, N-cube/10,
for example, does not have such functions. You have to effect this by 
sending a separate message to every other nodes. This, in fact, includes
too many duplicated transmissons, since some nodes have to forward this
same information many times. 
   I know some people have done some research in this area, such as 
the Yale University, but I wonder why these algorithms are not implemented
in the commercial products now, although such algorithms were ready several
years ago. Can anyone out there give an answer to this? Does it make sense
if we try to implement these algorithms now?

Jingwen Wang

Department of Electrical and Computer Engineering
University of South Carolina
Columbia, SC 29208

E-mail:wangjw@usceast.cs.scarolina.edu

berryman-harry@CS.YALE.EDU (Harry Berryman) (11/29/90)

In article <11912@hubcap.clemson.edu> wangjw@usceast.cs.scarolina.edu (Jingwen Wang) writes:
>Dear Networkers,
>   As you may know, many applications need multicast communications and
>broadcast communications. However, todays hypercube computer, N-cube/10,
>for example, does not have such functions. You have to effect this by 
  [ Stuff deleted ]
>Jingwen Wang

The Intel iPSC/2 and iPSC/860 hypercubes have somewhat better libraries
than the old NCUBE. Included are functions for global syncs, multicast,
global reductions, interupt driven communication, and some other
useful stuff. I cannot speak for the NCUBE2, but the old NCUBE was
a real bear to work with because of the lack of any solid software support.

This is not to imply that Intel has done a very good job supplying system
software, just that NCUBE's is (or at least was) much worse. 

Scott Berryman
Yale University CS Dept.       (berryman@cs.yale.edu)
or 
ICASE/NASA Langley Research Center (berryman@icase.edu)

joes@ncube.com (Joe Sikorski) (11/30/90)

In article <11917@hubcap.clemson.edu> berryman-harry@CS.YALE.EDU (Harry Berryman) writes:
>In article <11912@hubcap.clemson.edu> wangjw@usceast.cs.scarolina.edu (Jingwen Wang) writes:
>>Dear Networkers,
>>   As you may know, many applications need multicast communications and
>>broadcast communications. However, todays hypercube computer, N-cube/10,
>>for example, does not have such functions. You have to effect this by 
>  [ Stuff deleted ]
>>Jingwen Wang
>
>The Intel iPSC/2 and iPSC/860 hypercubes have somewhat better libraries
>than the old NCUBE. Included are functions for global syncs, multicast,
>global reductions, interupt driven communication, and some other
>useful stuff. I cannot speak for the NCUBE2, but the old NCUBE was
[ remainder of message deleted ]

I've discussed with Dr. Erik DeBenedictis, Staff Scientist 
at nCUBE, the technical issues mentioned in the messages 
above. He made these comments about those issues:

  The nCUBE 2 hardware has a broadcast mode as part of
  the hypercube wormhole router.  Now, use of this
  hardware is somewhat restricted due to the possibility
  of broadcast deadlocks (as first discovered on the
  Cosmic Cube). In current software, when the host or any
  node sends a message to destination -1, the message is
  broadcast to all nodes in the subcube (including the
  sending node, but not including the host even if sent
  from the host).  Broadcast messages are relayed by
  the kernel to node 0 (unless they originate on node 0),
  which then uses the broadcast hardware to send to all
  nodes.

  Multicast is currently under consideration
  for the nCUBE 2.  Uniformly efficient methods
  of specifying the set of target processors have not
  been found.  Suggestions are welcome.

  The nCUBE 2 software library includes global
  synchronization, and global operations.  Some reactive
  kernel primitives are available on an experimental
  basis.  The Express library from Parasoft is bundled
  with all software releases.

His comments are based on Release 2.0 of nCUBE's software 
for the nCUBE 2 series of massively parallel supercomputers.
Release 2.0 will be generally available at the end of this year.

--Joseph Sikorski, nCUBE Product Marketing
nCUBE; 20 Davis Drive; Belmont, CA 94002
joes@ncube.com