Performance of 1 gig and 10 gig ethernet cards

[spidey07]

I have never seen such a good blend of server/bus discussion coupled with networking. If you want to learn or possiblely hardcore look at bus timings please take a look. I'd be happy to discuss the reasons on why they achieved their results from a networking perspective. The main reason why full 1 and 10 gig ethernet throuput is not achieved is a combination of interframe gap (required by Ethernet), processor and bus. Very good read.

1 gig and 10 gig NICs from a research perspective

 

[Peter]

Excellent read, yes. Thanks for the pointer!

 

[spidey07]

It's to the point now that "the network" is not a bottle neck. The host just can't push as much data as a single high speed network interface can handle.

It's similar to when 100 Base-T came out (PCI was still very new then) and there wasn't much reason for an ISA 100 Base-T card....bus couldn't keep up. Heck even on PCI the PC couldn't keep up or fill a 100 meg connection.

 

[Peter]

Back then the bottleneck wasn't so much the PCI bus. The memory interface and CPU couldn't handle the traffic even though PCI could have.

Today it's basically the same thing with 10GE - the bandwidth on the currently available peripheral busses could do it, but the x86 PC system core can't /quite/ keep up.

 

[Gannon]

Solid state disk drives could not come any faster considering the magnitude of speed increases in network bandwidth.

 

[nightowl]

I'll have to read this. Going forward, I have been hearing that TCP will eventually break (be a limiting factor) for faster mediums of transmission. Enhancements will need to be made to TCP to allow faster network transmission on hosts.

 

[spidey07]

Originally posted by: nightowl
I'll have to read this. Going forward, I have been hearing that TCP will eventually break (be a limiting factor) for faster mediums of transmission. Enhancements will need to be made to TCP to allow faster network transmission on hosts.

That is absolutely true. TCP version 4 (IPv4) shows it's age with the algorithyms used to "ramp up in speed" a session with the 10+ gigabit speeds in networking today.

this paper uses experimental tcp/ip stacks and other enhancements to isolate the card/hardware/subsystem from the limitations of TCP.

But you still will see very little movement with IPv6. Maybe in 5-10 years, but today nobody wants to change.