LACP & port trunk question

[njia_1]

Hi

Is LACP only supported between switches ? What I am tring to do is, I installed 4 Intel pro server adapters on one of RHEL 3 server and setup them working as a team using Adaptive Load balancing mode.

I connect all the 4 NICs to a HP 2824 switch (port 1,2,3 and4) then I login the switch and setup port 1-4 using LACP to increase the bandwidth.

I assumed this is how LACP works and I am not sure

After I done this, I lost the connection to that RHEL box ? Should I setup the Link aggregration before I connecting that RHEL box ? or LACP is just supported between switches.

[GGA]

hello

u can use lacp only between switches.. it is the same as trunking

regards gga

[rick jones]

IIRC, LACP can run between any two endpoints that wish to aggregate links. They do not both have to be switches. For example, the Auto Port Aggregation software for HP-UX supports LACP.

There are other mechanisms available - Cisco Fast Etherchannel is one. Some host platforms offer "poor man's" versions that only do outbound - round-robbining connections across the links, presuming that just balancing output is sufficient.

In just about all these cases, one has to remember that while the _aggregate_ bandwidth will be increased, the performance of any single flow (aka connection) will still be limited by the bandwidth of a single link.

My guess is that the Adaptive Load balancing software in Linux is something other than LACP. A bit of web searching would probably uncover more there. In linux land, I believe that the "bonding" drivers are the ones that speak the link aggregation protocols.

[njia_1]

Hi Rick

thanks for your reply, yes I already did some research myself, ALB does not need any support from switch and it is something different to LACP.

I found another project on sourceforge.net it support a wide range features like LACP and ALB.

Just regarding your response, so one connection (fox example, ftp) can only use one physical link, is that what you are saying ? I would thought that the aggregated link is some kind of a logical NIC to the system. Just like RAID.

For example, if I create a 4 100M port link aggregation, but one ftp connection still can only use up to 100M of the total bandwidth. Is this right ?

thanks
Nan

[rick jones · Accepted solution]

That is correct - while there may be 4 100 BT links in your aggregate, a single FTP connection will only go over one of them. The reason is simply that if the packets of a single flow were put across the four links, there is a good chance that packet reordering would happen. While TCP will "deal" with packet reordering, the effect on performance can be substantial. Once packet reordering is "bad" enough, it leads to spurrious retransmissions and shrinkage of the congestion window.

So, very few, if any link aggregatation mechanisms will do "round robin" of packets of a given flow. There may be some that do, but enabling that has to be done with great care.

Why it isn't like RAID0 centers on things like the writes to the different discs not being related to one another like the bytes in a TCP stream - SCSI doesn't care that Disc A finished its write after Disc B - they are separate discs afterall.

[njia_1]

thanks Rick