I've been searching around the web for this problem i've been having. It's similar in a way to this question: How exactly & specifically does layer 3 LACP destination address hashing work?
My setup is as follows: I have a central switch, a Procurve 2510G-24, image version Y.11.16. It's the center of a star topology, there are four switches connected to it via a single gigabit link. Those switches service the users.
On the central switch, I have a server with two gigabit interfaces that I want to bond together in order to achieve higher throughput, and two other servers that have single gigabit connections to the switch.
The topology looks as follows:
sw1 sw2 sw3 sw4 | | | | --------------------- | sw0 | --------------------- || | | srv1 srv2 srv3
The servers were running FreeBSD 8.1. On srv1 I set up a lagg interface using the lacp protocol, and on the switch I set up a trunk for the two ports using lacp as well. The switch showed that the server was a lacp partner, I could ping the server from another computer, and the server could ping other computers. If I unplugged one of the cables, the connection would keep working, so everything looked fine.
Until I tested throughput. There was only one link used between srv1 and sw0. All testing was conducted with iperf, and load distribution was checked with systat -ifstat.
I was looking to test the load balancing for both receive and send operations, as I want this server to be a file server. There were therefore two scenarios:
- iperf -s on srv1 and iperf -c on the other servers
- iperf -s on the other servers and iperf -c on srv1 connected to all the other servers.
Every time only one link was used. If one cable was unplugged, the connections would keep going. However, once the cable was plugged back in, the load was not distributed.
Each and every server is able to fill the gigabit link. In one-to-one test scenarios, iperf was reporting around 940Mbps. The CPU usage was around 20%, which means that the servers could withstand a doubling of the throughput.
srv1 is a dell poweredge sc1425 with onboard intel 82541GI nics (em driver on freebsd). After troubleshooting a previous problem with vlan tagging on top of a lagg interface, it turned out that the em could not support this. So I figured that maybe something else is wrong with the em drivers and / or lagg stack, so I started up backtrack 4r2 on this same server.
So srv1 now uses linux kernel 220.127.116.11. I set up a bonding interface bond0. The kernel module was loaded with option mode=4 in order to get lacp. The switch was happy with the link, I could ping to and from the server. I could even put vlans on top of the bonding interface.
However, only half the problem was solved:
- if I used srv1 as a client to the other servers, iperf was reporting around 940Mbps for each connection, and bwm-ng showed, of course, a nice distribution of the load between the two nics;
- if I run the iperf server on srv1 and tried to connect with the other servers, there was no load balancing.
I thought that maybe I was out of luck and the hashes for the two mac addresses of the clients were the same, so I brought in two new servers and tested with the four of them at the same time, and still nothing changed. I tried disabling and reenabling one of the links, and all that happened was the traffic switched from one link to the other and back to the first again. I also tried setting the trunk to "plain trunk mode" on the switch, and experimented with other bonding modes (roundrobin, xor, alb, tlb) but I never saw any traffic distribution.
One interesting thing, though:
one of the four switches is a Cisco 2950, image version 12.1(22)EA7. It has 48 10/100 ports and 2 gigabit uplinks. I have a server (call it srv4) with a 4 channel trunk connected to it (4x100), FreeBSD 8.0 release. The switch is connected to sw0 via gigabit. If I set up an iperf server on one of the servers connected to sw0 and a client on srv4, ALL 4 links are used, and iperf reports around 330Mbps. systat -ifstat shows all four interfaces are used.
The cisco port-channel uses src-mac to balance the load. The HP should use both the source and destination according to the manual, so it should work as well. Could this mean there is some bug in the HP firmware? Am I doing something wrong?