Max. number of hosts in segment

Max. number of hosts in segment

Post by Dave Pasc » Wed, 17 Nov 1993 07:09:13



I'm interested in finding out what others' experiences have been with
the maximum number of UNIX machines per Ethernet segment (on a
10-BaseT concentrator), especially where most of the machines are
high-throughput machines (SPARC-10's, etc.) talking primarily NFS with
a file server.

I'm just trying to get a sanity check on a network plan and I'm
interested in getting a sampling from the net.


==== Tel: 508-653-2452 (x362) ==== Fax: 508-653-2997 =================

 
 
 

Max. number of hosts in segment

Post by Dave Pasc » Thu, 18 Nov 1993 12:22:50


: This is not clear to me - the maximum number of devices on a 10baseT segment
: is "1" - period. 10BaseT hubs are repeaters, and every cable coming out of
: a hub is an ethernet segment on its own.

I should have phrased this more carefully......I meant to say
subnet, not segment.  I guess I tend to think of thick and thin-net
and hence 'segments'.  I hope you get the idea.....

This problem has too many variables, as I later realized and someone
pointed out, but perhaps some of you can share some data anyway.


     The MathWorks, Inc.
     24 Prime Park Way, Natick, MA 01760
==== Tel: 508-653-2452 (x362) ==== Fax: 508-653-2997 =================

 
 
 

Max. number of hosts in segment

Post by Ric » Fri, 19 Nov 1993 03:20:59




>|I'm interested in finding out what others' experiences have been with
>|the maximum number of UNIX machines per Ethernet segment (on a
>|10-BaseT concentrator), especially where most of the machines are
>|high-throughput machines (SPARC-10's, etc.) talking primarily NFS with
>|a file server.
>This is not clear to me - the maximum number of devices on a 10baseT segment
>is "1" - period. 10BaseT hubs are repeaters, and every cable coming out of
>a hub is an ethernet segment on its own.

You're equating a segment with a wire, which not everybody does (including,
I think, the original poster). "Segment" when referring to 10BaseT is usually
taken to mean a collision domain, in which case the theoretical limit is
1024 (one can have multiple hubs in a collision domain).

As for the number of unix hosts that can be run, when are people going to
learn that there is no easy answer to this? It's entirely dependent on a
variety of factors not evident from the post and a subjective evaluation of
performance that only the original poster could judge.

 
 
 

Max. number of hosts in segment

Post by ober.. » Fri, 19 Nov 1993 05:27:55





>>This is not clear to me - the maximum number of devices on a 10baseT segment
>>is "1" - period. 10BaseT hubs are repeaters, and every cable coming out of
>>a hub is an ethernet segment on its own.
>You're equating a segment with a wire, which not everybody does (including,
>I think, the original poster). "Segment" when referring to 10BaseT is usually
>taken to mean a collision domain, in which case the theoretical limit is
>1024 (one can have multiple hubs in a collision domain).

Sorry, but this is simply not true. A segment is a physical run of the Ethernet
medium. Any other definition is simply wrong. A segment is defined in 802.3 as
a physical entity and most certainly NOT the collision domain. This is not only
referred to in the base document, but in Section 14.1.1 on 10Base-T.

Network is confusing enough to people. Let's not make it worse by redefining
terms on the fly.

On the other hand, you are probably right about whet the original poster was
asking. And your answer, if that was the question, was not unreasonable.

R. Kevin Oberman                        Lawrence Livermore National Laboratory

Disclaimer: Being a know-it-all isn't easy. It's especially tough when you
don't know that much. But I'll keep trying. (Both)

 
 
 

Max. number of hosts in segment

Post by Ric » Fri, 19 Nov 1993 17:26:02






>>>This is not clear to me - the maximum number of devices on a 10baseT segment
>>>is "1" - period. 10BaseT hubs are repeaters, and every cable coming out of
>>>a hub is an ethernet segment on its own.
>>You're equating a segment with a wire, which not everybody does (including,
>>I think, the original poster). "Segment" when referring to 10BaseT is usually
>>taken to mean a collision domain, in which case the theoretical limit is
>>1024 (one can have multiple hubs in a collision domain).
>Sorry, but this is simply not true. A segment is a physical run of the Ethernet
>medium. Any other definition is simply wrong. A segment is defined in 802.3 as
>a physical entity and most certainly NOT the collision domain. This is not only
>referred to in the base document, but in Section 14.1.1 on 10Base-T.
>Network is confusing enough to people. Let's not make it worse by redefining
>terms on the fly.

I should have made myself more clear. I was referring to the meaning the
original poster attributed to segment only. Thus the answer of "1" was going
to be pretty confusing to said poster.

-Rich

 
 
 

Max. number of hosts in segment

Post by Vivek Khindr » Mon, 22 Nov 1993 14:29:18


Dave,

   You can start by making some assumptions and some samples.

   This is very crude and simple but its a start:  Determine
   the typical I/O block size for your NFS servers.  Estimate
   what the I/O demand might be on your server.  Find out what your
   server can support as sustained I/O.     Estimate
   the total capacity of your network by picking a working load (say
   30%) to represent the running saturation of your network.
   This give you a capacity of 3 Mbits/sec.  Determine the amount
   of NFS activity generated by each server and its clients.

   Divide the total capacity by the average network load per server/client
   group, and you will end up with a rough number describing max groups per
   segment.

   This gets more complicated when individual machines are both clients and
   servers, and you through in terminal traffic, email, and other traffic.

   I hope someone can give a more formal approach, but I am afraid this can
   be a very complicated thing to model.

   Some of my calculations have come up with  2 server/client groups for
   basic office automation scenerios.

Vivek


>I'm interested in finding out what others' experiences have been with
>the maximum number of UNIX machines per Ethernet segment (on a
>10-BaseT concentrator), especially where most of the machines are
>high-throughput machines (SPARC-10's, etc.) talking primarily NFS with
>a file server.

>I'm just trying to get a sanity check on a network plan and I'm
>interested in getting a sampling from the net.


>==== Tel: 508-653-2452 (x362) ==== Fax: 508-653-2997 =================


| McMaster University                            Computation Services Unit |
| Faculty of Health Sciences                         (905) 529-7070 x22439 |
v 1200 Main St W, Hamilton, Ontario,Canada L8N3Z5                          v
 
 
 

Max. number of hosts in segment

Post by Kevin - Steve » Wed, 24 Nov 1993 08:08:54


Well, I'll just call it a "collision domain" to avoid all the terminology
flames...

Within one of 'em, as has been pointed out, your mileage will vary considerable.
I've never found a more appropriate way than starting with the fewest practical
machines, hooking up a protocal analyzer or traffic monitor, and watching the
numbers as you do REAL WORK with the network.  There are a variety of software
PAs available now under $1K, and they will all work fine for this sort of thing.

For what it's worth, here are some personal experiences:
I have a 132 port 10BT hub, loaded with light IPX traffic.  It works fine with
single-digit utilization, and peaks of 40% or so for a few seconds at most.
However, I consider this to be Too Damn Many nodes on one segment, and it makes
me nervous, because if this light traffic ever changes to "moderate traffic",
so many stations trying to talk will send the collisions through the roof.

I'd rather have utilization be my limiting factor than collisions any day.

I have another collision domain with exactly two nodes on it.  It is way
overloaded because those nodes are Novell servers, and everybody is trying
to talk to them.  Exactly the opposite situation:  very high utilization
(40% AVERAGE), but relatively low collision rate because there aren't many
stations to fight over the wire.  I'm moving those systems to FDDI shortly.

I have a segment with 12 Sun stations on it.  There is practically no traffic
on it at all because all the stations are set up as local hosts and rarely
talk to each other.

                                    KeS

 
 
 

1. (Shared Memory Limits) Raise max number of segments per process

I would like to know how to raise the  max number of
segments(IPC V shared memory) per process ?
I found on /proc/sys/kernel the shmall and shmmax.

I adjust successfully the Maximum size in bytes for a shared memory
segment
echo 1073741824 > /proc/sys/kernel/shmmax

But how to set on the flight the max number of shm segment.
I have only 128 segments max on my Linux machine while I need
1000 segments !
SHMMNI     System wide maximum number of shared memory segments

Thanks for your help

  freygagne.vcf
< 1K Download

2. what is a library ending by .so?

3. Apache max number of virtual hosts?

4. how to unsubscribe

5. Q: max number of hosts in netgroup

6. IDE drive necessary ?

7. How can I tuning Max process number and open file number on Sun Solaris 7(Intel)???

8. Stats comp.os.linux.misc (last 7 days)

9. why max ethernet segment length is 185 meters

10. max size of shared memory segment

11. max bomb segment tuning with read latency 2 patch in contest

12. TCP Max. Segment Size

13. TCP max segment size on HP-UX 10.20