View Single Post
Old 24 August 2010, 14:26   #40
Olaf Barthel
Registered User
Join Date: Aug 2010
Location: Lehrte, Germany
Posts: 154
Originally Posted by Zetr0 View Post

my thanks for your answers and some impressive metrics =)

would you have any tests results for those on lesser 68k based machines like the humble 68k moto or 020 / 030? (would you expect them to be about the same?)
I only tested this on my development machines, and it's been a while since I did those tests (Roadshow 4.205 was in development in 2003). I'm not sure I can muster a plain 68k machine to test with. I have an A600HD in storage (acquired for research purposes: I wanted to figure out why the scsi.device driver in the A600 OS 3.1 ROM was version 39 and not version 40), but if I remember correctly, the driver required for the only PCMCIA networking hardware I have does not support the 68000 processor.

The overall performance depends upon several factors: memory performance, networking card transmit buffer size, CPU performance. If you don't hit the high notes on memory performance and networking card transmit buffer size then your performance figures will be more on the disappointing side, I'm afraid.

I can see there is a limmitations of the adapters and this (pending on the stack used) swings quite wildly... especially with your latest "Roadshow 4.205" core achieving almost 37% more on an Adriane 1 equipped 040@40mhz [977KB] compared to "Miami Deluxe 1.0c SANA-II" but with an 060@50 [633KB]

Thats quite impressive - not to mention 977KB is has probably maxed out the Adriane, what do you think this stack could achieve on a PCI based 10/100 card?
As I wrote in a previous posting to illustrate how these figures came about, the transfer figures are not entirely realistic as they merely exercised sustained data transmission without doing anything useful with the data. The 977 KByte figure, although it was repeatable, almost seems to be physically impossible, given the limitations of the Zorro II bus.

On a PCI based system with a faster bus this TCP/IP stack probably won't go that much faster. The reason is in the SANA-II driver design which does not mesh at all with how PCI networking hardware does DMA. Hence, you will end up copying to/from the networking hardware buffer to local Amiga memory a lot, and this is where the performance hits a wall.

(forgive my ignorance its been a while, but this would likely be limmited to the bus-board speed? - Z2 being about 3.58MB a second and Z3 being about 9MB - of course the CPU will make a difference.)

Some great stuff Olaf =)
The bus speed is one of the limiting factors. But it also matters how the networking hardware offers up the data it receives. If you have a small receive buffer, like apparently all of the PCMCIA cards do, then the consequences for the I/O performance should be obvious. But it's not just the buffer size alone. If I remember correctly, the "Ariadne I" and the "Ariadne II" both have 32 KBytes of on-board transmit memory, but the "Ariadne II" performs slower than the "Ariadne I" under otherwise identical circumstances.
Olaf Barthel is offline  
Page generated in 0.05222 seconds with 9 queries