Fast Ethernet Wire-speed?

[cautery]

Can anyone tell me:

1)what the theoretical wire-speed is for 100Mbit Ethernet?
2)approximate real-word wire speed for 100Mbit FULL-Duplex transfers through a Netgear RT314 with FA310TX NICs on either end over Win 2000 Pro?

I'm getting about 9.89 average....

Thanks,
Clay

 

[Dually]

1) 13megabytes/sec
2) 26megabytes/sec (duplex is 200mbit)

 

[cautery]

Thanks for the response Dually...

Followup... 13 and 26 sound kinda high.... Are you just multiplying 1024 bits/kb x 1024kb/MB x 100 and then dividing by 8? And doubling that for Duplex?

I understand this mathematically, but this doesn't seem to account for the known overhead <all the control info in each packet> of TCP/IP.

Seems that this would take a pretty big chunk out of the max.

Thanks again,
Clay

 

[spidey07]

It really depends on frame size. But the maximum is in the range of 95-99.7 megabits per second. This is not theorectical. This is just how fast it really is.

I'm assuming full duplex here. Dually may be mistaken, you do not add the TX and RX together...it is simply 100 megabit full duplex.

 

[cautery]

That's what I thought I remembered from the Beowulf project pages.... about 95-98 is what I remember.... max wire speed. So the other 0.3-5% is eaten up by overhead?

Wonder what else I can tweak to get my efficiency up?

I'm only about 79% right now....

I think you are right.. 100Mbit is the nominal rating at Full-Duplex.

Thanks,

 

[spidey07]

The .3-5% variance comes from the preamble and interframe gap. Every ethernet frame has a series of ones and zeros that is used for clock syncronization. Also every frame has to have a gap of 9.6 microseconds.

This is why with max frame size of 1518 the interframe gap and preamble are of little consequence with regards to overhead. With 64 byte frames they start to have an impact.

Now when you mention overhead I normally think of addressing overhead which can be as much as 64 bytes.

 

[cautery]

How do I tweak the fram size?

 

[spidey07]

you don't. Small requests such as &quot;HTTP: GET&quot; use very small frames. But the replies will use larger frame sizes.

 

[Shadow07]

Spidey's correct. You can probably expect to get about 80Mb/sec on a full duplex 100Mb connection. But, you can get slower speeds when you are using a GUI hog program to transfer files, like Windows Explorer. I have seen transmissions speeds of about 70-80Mb/s using FTP on a local LAN.

 

[CTR]

How 'bout your cabling? A bad crimp or a broken conductor can cause your throughput to suffer.

 

[cautery]

CTR - Good suggestion.... But I get +/- 1% of the same results no matter which two workstations or which machine I initiate from, so unless they are all bad, it's probably not the cables....

... besides, &quot;I&quot; made all of the custom cables myself using CAT-5E cable and CAT-5E modular connectors (and I'm a freak about making quality cables!)

Good thinking though....

It's getting interesting... What I really need to do is go back and re-run the tests with freshly booted machines (though the minor deviations between combinations would seem to negate the need for this).

Thanks... y'all keep the ideas coming!

 

[Dually]

Wrong, 100mbit is half duplex and 200mbit is full duplex, hench the duplex word folks, come on really.

 

[spidey07]

Dually, what is the bandwidth of a T1 or OC3?

These are full duplex technologies, but we don't say a T1 is 3.0 megabits or an OC3 is 310 megabits. (1.5 and 155 respectively)

100 megabit ethernet still has a bandwidth of 100 megabits. Just because a link is full-duplex does not mean you double your bandwidth.

 

[CTR]

Uh-oh...the term &quot;full-duplex&quot; implies simultaneous two-way communication, so if you said &quot;full-duplex at 200Mbps&quot; you would be talking about 200Mbps TX and 200Mbps RX.

 

[Dually]

um a T1 is 1.5mbits. It isn't like ethernet and they don't work like that. Here is a link to a ethernet card spec page that proves i am right. http://www.sybercom.com/Nc32W100.html

 

[Shadow07]

Dude, are you really going to believe some marketing person over some one who is actually technical and knows what the hell they are talking about?

 

[spidey07]

Dually, you've been reading too much marketing stuff.

Answer me this...What is the serilization delay of a single bit on fast ethernet? What is the serialization delay of a single bit on fast ethernet-full duplex?

They are the same. If full duplex truly is 200 megabits then the serial delay would be half that of 100 megabits. But it is not.

The bandwidth is 100 megabits full duplex. My main point is that you don't have a 3.0 megabit T1 do you? Then how can you have a 200 megabit ethernet connection? I know a lot of manufacturers like to say &quot;Wow! Look you have 200 megabits per second of bandwidth&quot;. That is a misconception and is only used to fool people.

 

[Shadow07]

Full duplex only means that you can send/recieve AT THE SAME TIME. It does not mean that you have 200Mb both ways. It just means that you can push 200Mb/s through the pipe at the same time. 200Mb/s limit. Nothing more.

So, lets put your theory to test again. So you are saying that I can get 400Mb/s throught he same pipe, since it is full duplex? Right? ERRRRRRRRR! I don't think so. If you know anything about Ethernet and how duplexing works, you will know that it only means that you can send and recieve data at the same time. Instead of half-duplex which means you can only send data until the device has recieved data, and vice versa.

 

[cautery]

Man! Did this discussion heat up or what!

I'm not going to spend a couple of hours to go find it, but I can assure you that Full Duplex does NOT mean 200Mbit in each direction. I KNOW this because I spent days reading/studying up on Donald Becker's work of creating the Linux Tulip driver and how it was tweaked specifically for the FA310TX (actually for the tulip chip). (CEDIS/NASA Beowulf project)

I posted this question because I could not remember what this same research info had said the wire-speed was for 100Mbit Ethernet.

I can see everyone is passionate about their positions, but could we please re-focus on what is causing my 100Mbit Full Duplex connections to languish down at 79% (or 34.5% if you choose 200Mbit in both directions) of wire-speed? No matter your position, this is much too inefficient, and I'd like to clean it up.

Let's attack it from a &quot;What can I do?&quot; standpoint.

Thanks,

 

[CTR]

1. Anybody know if the netgear being discussed uses store-and-forward or cut-through switching? 2. What OS have you been testing with?

 

[spidey07]

Also provide PC hardware.

Overall 80% is pretty good. Keep in mind a switch does add some latency (CTRs just trying to figure out how much)

 

[Dually]

It is you guys that are silly, the same argument could be said for ddram and such. It isn't really 200mhz or 266mhz or what have you. Of course you can't get 200mbit but you badwidth is doubled.

 

[Mday]

<< &quot;I&quot; made all of the custom cables myself using CAT-5E cable and CAT-5E modular connectors (and I'm a freak about making quality cables!) >>



lol clay ;-)

 

[Mday]

<< It is you guys that are silly, the same argument could be said for ddram and such. It isn't really 200mhz or 266mhz or what have you. Of course you can't get 200mbit but you badwidth is doubled. >>



DDR has nothing to do with &quot;full duplexing&quot; it's that &quot;transactions&quot; occur at the rise and fall of a clock cycle...

full duplex is just sending and receiving at the same time, which &quot;essentially&quot; reduces time spent since instead of waiting to receive while sending, you can receive while sending. does this double your bandwidth? NO!...

 

[Dually]

yes it does if you listen to what you jsut said. It not the same but its a good comparision and ddr also can be argued is realy doubling jsut like duplexing can be.