🚀 go-pugleaf

On Thu, 25 Sep 1997 15:46:25 +1000, Brian Miller <bmiller@telstra.com.au> wrote:
>Alan Palmer wrote:
>> 
>> I have a ISA-based 50 MHz i486 box connected to a 166 MHz UDB (Multia)
>> via 2 metres of 10Base2 coax. Nothing else is connected to this network
>> (apart from cable terminators).
...[snip]
>> If I transfer large files (> 300 Kbytes) using ftp from either box, I
>> get a throughput of about 100 kbyte/s, plus or minus 20%.
...[snip]
>
>I have two Linux hosts at home on a 10Base2 Lan.  They both have old
>8 bit WD8003E cards in them.  One is a 6x86 P200, the other is a
>Pentium 90.
>
>I just upgraded one of the hosts a few weeks ago, and FTP was one
>thing I looked at.  I now get _up_to_ 610KB/s with a FTP GET in
>Binary to /dev/null.

The speed of ISA ether cards is very dependent on the type of CPU, since
ISA cards generate a lot of interrupts and have fairly slow I/O. So we
see above the pentiums do much better than a 486DX2. (Bus masters, such as
the Tulip, depend less on the CPU speed.) These comments are also true
for PLIP using the // port. 486's get up to 20kb/s while pentiums get
over 50kb/s.  Also, my guess is that 10base2 is a good bit slower
than 10baseT.

Measuring the speed of ether needs a bit of planning. You don't want to
measure disk speed or the speed of an overloaded kernel. There's a utility
called tcpblast that measures the h/w speed of tcp/ip directly.
It's best used with random tcp packets, e.g.:
  tcpblast -rt host.domain.edu.au

The only real performance tuning is done by changing the MTU size
with ifconfig. Generally it's best between 1500 and 2000 (max).
There are other tuning parameters for NFS, etc.

-- 
=================================================================
Rowan Hughes, Maths Dept, James Cook University. x-csrdh@lion.jcu.edu.au