[Spread-users] Spread on point-to-point links

[Jonathan Stanton]

On Thu, Nov 15, 2001 at 10:09:35PM -0800, Dave Parker wrote:
> I may have answered my own question here - here's what I did:
> 
> I have two nodes connected via a PLIP link (which is by definition
> ptp).  These look like this:
> 
> plip0     Link encap:Ethernet  HWaddr FC:FC:3F:79:88:FD
>           inet addr:A.B.C.Y  P-t-P:A.B.C.Z  Mask:255.255.255.255
>           UP POINTOPOINT RUNNING NOARP  MTU:1500  Metric:1
>           RX packets:1098 errors:0 dropped:0 overruns:0 frame:0
>           TX packets:1072 errors:0 dropped:0 overruns:0 carrier:0
>           collisions:0 txqueuelen:10
>           Interrupt:7 Base address:0x378
> 
> The problems are:
> 
> 1) The PLIP driver with the kernel I'm using has a problem with UDP
> broadcasts (as in won't do 'em)
> 2) I'm not even sure what a broadcast would look like on a network like
> this.

Generaly, if two machines are connected in such a way as they cannot do
broadcast to each other, then the right solution with spread is to put them
into separate segments (just like you did). The effective definition of a
Spread segment is a set of machines who can reach each other with one
multicast or broadcast.

 > 
> My solution was to set up each node as a segment:
> 
> Spread_Segment A.B.C.Y:4803 {
>         sat-ha1               A.B.C.Y  {
>            D A.B.C.Y
>            C A.B.C.Y
>         }
> }
> Spread_Segment A.B.C.Z:4803 {
>         sat-ha2               A.B.C.Z  {
>            D A.B.C.Z
>            C A.B.C.Z
>         }
> }
> Spread_Segment  127.0.0.255:4803 {
>         localhost               127.0.0.1  {
>            D 127.0.0.1
>            C 127.0.0.1
>         }
> }
> 
(with your corrected segments you sent in the next email) This looks fine,
my only question is if you really want to accept client connections on
a.b.c.z and a.b.c.y, or just 127.0.0.1? You don't have to have the C
a.b.c.z line if you don't want it.

> If I understand spread correctly, this means that I'll incurr some
> overhead as spread will assume it needs to use the hop protocol to get
> between the two nodes (?).

Actually, there are two different ways the networking can work. The hop
protocol is only currently implemented in a research/development version I
have on my laptop and is not in the released version. 

The way spread will work in this case is to form a ring of all of the
segments, so it will behave pretty much the same as if you had all the
machiens in one segment. The performance cost will be if the plip link is
'slow' latency wise, then the overall performance will take a hit because
the token has to rotate amoung all the machines and thus has to cross the
plip link twice during each rotation.

Jonathan

-- 
-------------------------------------------------------
Jonathan R. Stanton         jonathan at cs.jhu.edu
Dept. of Computer Science   
Johns Hopkins University    
-------------------------------------------------------