can't get full duplex speed 1

[jimfixit]

I am working with a tool called iperf,

http://dast.nlanr.net/Projects/Iperf/

which is excellent by the way. I'm measuring performance between two workstations with a 6500 between. The 6500 has blades w/ 48 ports per, 10/100/1000.

The ports for each station show 1000 full duplex, but the best speed I can get is avg. 500 Mbps. If I force the ports to 100 full, I get avg 92 Mbps. It's like I can't get full duplex to run even though that's what the switch shows me connected at. I see no errors on the port stats.

Any ideas why I can't get the full speed?

 

[chipk]

Are you sure the bottleneck is not with the transferring or receiving station? Lots of things affect transfer speeds. You could be overrunning a queue somewhere, or you could be trashing the hard disks, or any number of other things. I wouldn't immediately suspect the network.

 

[chipk]

that should read: "thrashing" the hard disks, not "trashing".

 

[voltron1011]

I don't understand what you mean by getting them to run full duplex. If the port speed is set above 10mbps then you are in full-duplex... If there was some reason why something couldn't do full-duplex (like old cat-3 cabling), then the switch would go down to auto-10 half-duplex automatically. So, since you're equiptment is all auto negotiating to auto-1000 full-duplex, there is no issues with the switch or nics on the work stations.

 

[jneiberger]

Voltron1011, you are misunderstanding a couple of things.

First, duplex is not dependent on speed. You can have 100 Mbps connections running at half duplex. Depending on the hardware you can also have gigabit ethernet running at half duplex.

Second, the autonegotiation process does no sort of link quality testing. Even if you're running Cat 3 cable, if the endpoints are capable of 100/full, that's what they're going to negotiate to. It's up to you to determine whether the resultant performance is acceptable or not and manually make a change, if necessary.

Back to the original post, there are lots of things that limit how fast transfers can be made. Someone already mentioned hard drive speeds. Also keep in mind that there is a LOT of hardware out there that is allegedly gigabit ethernet, but can't actually forward frames at those speeds. Lots of older Cisco stuff is like that. On some Cisco gear you can't ever get over about 800 Mbps because the interface hardware simply can't forward the frames any faster than that.

Another key is your TCP settings. TCP doesn't scale well to really fast speeds. At gigabit speeds, the recipient is constantly sending a stream of acknowledgments to the sender and this really slows things down. Do a Google search on tweaking TCP to increase your file transfer performance on high speed links. You'll find some really good information that might be useful.

Good luck!

 

[jimfixit]

Thank you. I've wondered about tuning TCPIP but shy away because most often it talks about tuning for a particular usage, say large databases. If a person typically uses a variety of stuff, large databases, email, small word or excel files, it's likely to make one thing better and another even worse....but will take a look see.

And yes, I loved ChipK's note about trashing disks..for all I know that could be closer to the truth!!

I did look at backplane speeds of these stacks and, it would seem some are so heavily loaded it's possible they can't transfer at the full speed.

 

[jneiberger]

It's not just backplane speeds that you need to worry about. Often they have capacity, but the interface hardware can't handle full rate. You should be getting better than 500 Mbps on your link assuming that your endpoints can actually handle sending and receiving at that rate.

 

[voltron1011]

jneiberger, I guess my post was a bit off... However, my comment were based on what the OP gave as his setup (6500 with 48port 10/100/1000 blade) and two PC's running Gig-nics. You are correct that switches/nics can run 100-half or even 1000-half (like if you introduced a hub into the scenario), however in the scenario given above, it can be assumed that everything should (and was) running 1000-full.

 

[voltron1011]

Also, 1000mbs doesn't work without cat-5e cable (with all 8 wires).