Take the 2-minute tour ×
Server Fault is a question and answer site for professional system and network administrators. It's 100% free, no registration required.

I have 4-ports Intel 82571EB Gigabit adapter. It uses e1000e driver. The test machine is running Ubuntu 9.10, clients 9.10/10.04. I tried link aggregation (mode 4) and adaptive load balancing (mode 6). The setup is as follows:

server == 4-port NIC == 4 cat6 cables == Linksys SRW2024 switch == 4 or more clients.

I am testing with iperf and TCP. For mode 4 I setup a LAG in the switch. For mode 6 - no. My tests are:

4 or more clients, all of which have 1 Gbps connections are iperf clients. They run simultaneously. The server already listens for these connections (I had iperf -s running there). The results are the same as the server is running 1 Gbps only and not 4 Gbps combined. If I use one client I get 1 Gbps. If I have 4 or 8 clients the bandwidth per client goes down to 250 or 125 Mbps (these numbers are for illustration purposes only).

Why I can't get 4 Gbps to and out of the server? How I can fix it?

share|improve this question
1  
Look at the arp tables in all of the devices. –  dbasnett Jul 11 '11 at 14:22
    
Seems to work at geekomatic.ch/2010/08/09/1281357060000.html –  Paul Jul 11 '11 at 14:35
    
In the link the server and the switch were both running LACP. Not clear if that is what the OP is doing. –  dbasnett Jul 11 '11 at 14:59
    
@dbasnett - your first comment was the key. I needed to do ifdown on the respective interfaces to clean it up for sure. It took a bit of playing on the server side. Please add it as an answer so I could mark it as correct one. Thanks! –  grs Jul 12 '11 at 20:31
add comment

2 Answers

Load balancing across links in a port aggregation group is not done strictly per frame. The switch and the server both use a methodology taking some combination of source mac address, destination mac address, source and/or destination tcp or udp port, etc. into account to prevent out of order packet delivery for a single session.

You might find you'll see the performance you expect if your run multiple iperf sessions on different listening ports with each client connection to a different session...depending on the load balancing methods being employed.

share|improve this answer
add comment

I remember reading somewhere about bonding that it does not scales linearly, so with 4x 1Gbps NICs you should see something around 2.5-2.8 Gbps. If i find the source I'll update my answer.

share|improve this answer
    
Not usually. I've got systems with 4x NICs that can push 3.2Gbps+ throughput (which is the expected throughput with ~20% overhead). –  Chris S Jul 28 '11 at 2:53
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.