[comp.parallel] Networking PC-ATs

mays@PYR.GATECH.EDU (Dick Mays) (07/15/88)

 
 I am currently interested in networking 20 to 30, PC-AT's, 
 (286 or 386 boards) with rather large interboard communications
 requirements, as a part of an experimental project investigating
 fault tolerant processing.  Suppose I were to have 6, 386 boxes,
 functioning as database servers, (perhaps with replicated data),
 and 24 286 boxes as clients.  Each 286 box requires 50K bytes/sec
 of data transfer, (to and from the servers), in say 10K byte chunks.
 (Strictly hypothetical mind you).  
  
  Any suggestions (other than changing architectural platform)
  as to how this mission could be accomplished would be appreciated.
  Of particular interest would be:
   
      1) Suggestions for network topology.
         (currently leaning toward broad-band R/F).
	     
      2) Reccomendations for exisiting AT-bus compatible 
         controllers, (Ethernet or Broad-band).
		       
   
      3) Any general comments concerning feasibility, and possibile
         alternative solutions for smaller problem: (8 clients
         and 2 or 3 servers).
				       
 This is my first net posting so please forgive me if I have 
 violated any sacrosanct rules of Net etiquette.
				
Thanks in advance.
					 
 -- Dick Mays
					  
  Georgia Insitute of Technology, Atlanta Georgia, 30332
..!{akgua,allegra,amd,hplabs,ihnp4,seismo,ut-ngp}!gatech!gitpyr!mays

vandys@hpisoa1.hp.com (Andrew Valencia) (07/19/88)

>I am currently interested in networking 20 to 30, PC-AT's, 
>(286 or 386 boards) with rather large interboard communications
>requirements, as a part of an experimental project investigating
>fault tolerant processing.  Suppose I were to have 6, 386 boxes,
>functioning as database servers, (perhaps with replicated data),
>and 24 286 boxes as clients.  Each 286 box requires 50K bytes/sec
>of data transfer, (to and from the servers), in say 10K byte chunks.

    Depending on how risky you want to be, you might check out Gary
Delp's Memnet work at UDel.  He used a 80 Mb/S token ring with a shared
memory paradigm.  They had two cards working and a third being wrapped
when I visited.  His performance parameters made interesting reading.

				Andy Valencia