I used linux bonding driver to bond 2 NICs in mode 6 in client PC. ethtool ethX shows speed =1000 ethtool bond0 shows speed =2000
but when i use iperf bandwidth of eth0 = 934Mbps bandwidth of eth1 = 637Mbps bandwidth of bond0 = 934Mbps
Shouldnt the bandwidth of bond0 be around 2000Mbps?
In short: no, bonding does not work in this manner.
Long story: the linux bonding, with its various bonding scheme, is very configurable. It has no less than 7 different bonding types, each with its strong and weak points. I strongly suggest you to read the documentation you can found here. However, the take away is the (except for the round-robin mode, which I detail below) no bonding scheme is capable of increase the throughput of a single session, rather they speedup multiple concurrent session. So your iperf output is perfectly normal, as it open a single session which can not be accelerated by the bonding driver.
The only expcetion to this rule is the round-robin mode, which transmits packet in a, well, round-robin fashion: the first one goes out from the first inteface, the second one from the second interface, and so on. This bonding mode can accelerate a single session by the virtue of concurrent sent packets. However it has many pitfalls, ranging from incompatible switches, out-of-order packet delivery (with relative retransmission), bad scaling over 2 interfaces, etc.
This is only a very concise summary. If you are interesting in how bonding works, you should really give a serious look at the documentation I provided above.
I am currently experimenting with round robin - I noticed that when bonding 2 NICs it does provide a 1.6-1.7 Gb/s when using iperf (keep in mind that two machines I use to test the speed have 2 NICs each in mode 0). I did a test today with 3 NICs and got ¬900 Mb/s - the reason for this is that round robin works best with an even number of NICs. I would only use it as a backbone for server backup (on the cheap) with an NFS share...
