[tulip] Problems with full-speed routing on a tulip quad inte rface

[Narayanaswamy, Naga (Naga)]

What is the packet size that you are trying? Is it 64 byte packets?

I don't know if one should be able to get full throughput
via a software stack, but I am also interested in knowing answers.

Is there any fast forwarding cache in Linux without using full IP stack?
I am skeptical about achieving max throughput of 100 M full duplex,
also doing IP routing in software. Of course, one would need a 
custom board with a network processor to do a fast forwarding in hardware.


> -----Original Message-----
> From: Luke Stratman [mailto:lastra@wm.edu]
> Sent: Monday, November 04, 2002 2:15 PM
> To: tulip@scyld.com
> Subject: [tulip] Problems with full-speed routing on a tulip quad
> interface
> 
> 
> Hello all:
> 
> I'm in the process of configuring several linux boxes to act as high
> performance routers in a network testbed environment; the 
> hardware specs
> for each of them are as follows:
> 
>  - CPU:  Athlon XP 2000+
>  - RAM:  256 MB PC 2100
>  - Motherboard:  Asus A7V-266E
>  - NIC:  Quad interface card (not sure of the manufacturer) with a DEC
> 21143 
>    chip
> 
> I'm currently using Redhat 7.3 with the 2.4.20-rc1 kernel on these
> routers, but the problems/results that I will detail below are more or
> less consistent with those obtained on the 2.4.17 kernel installed
> previously.
> 
> Now, here's the problem:  I'm unable to obtain full speed routing when
> all four of the interfaces are active.  For illustrating purposes,
> here's the network topology for one of the routers:
> 
> 
>                      +--------------+
>                      |   Subnet 1   |
>                      +--------------+
>                             |
>                             |
> +--------------+     +--------------+     +--------------+
> |   Subnet 2   |-----|   Router 1   |-----|    Subnet 3  |
> +--------------+     +--------------+     +--------------+
>                             |
>                             |
>                      +--------------+
>                      |   Subnet 4   |
>                      +--------------+
> 
> 
> When I attempt to blast data at full speed from subnet 1 to 
> subnet 2 and
> subnet 3 to subnet 4 simultaneously, the max throughput that I can
> achieve is about 73 Mbps.  Initially, I was concerned about the
> interrupt count (when I first began investigating this, the interfaces
> were throwing about 50K interrupts each during the bandwidth 
> tests), but
> with the most recent kernel compiled with 
> CONFIG_HW_NET_FLOWCONTROL, the
> interfaces seem to only be throwing about 20K interrupts apiece during
> the test.  So, from this it would seem that interrupt mitigation is
> working properly, yet the box still cannot route at full speed.  In
> addition to the stock tulip driver included in 2.4.20, I've tried the
> NAPI driver from robur.slu.se, but to no effect.  I've also tried
> turning ipchains off, which yielded a minimal (~.1 Mbps) increase in
> throughput.  The box can route at full speed when only two of the
> interfaces are active, but when all four are trying to 
> get/send data at
> wire speed, then I see the drop in throughput.
> 
> So, my question is this:  are there any specific options that could be
> set on the interfaces, or patches that could be applied to 
> the driver or
> kernel to improve performance?  Or is it simply unreasonable to expect
> full speed routing in this scenario given the hardware configuration?
> If you need any further information from me, please don't hesitate to
> ask, and thanks in advance for any suggestions you might be able to
> give.
> 
> Thanks,
> -Luke Stratman
> 
> 
> 
> _______________________________________________
> tulip mailing list, tulip@scyld.com
> To change to digest mode or unsubscribe visit
> http://www.scyld.com/mailman/listinfo/tulip
>