IEEE 802.3ad NIC teaming a good idea? (no speed increase)

[Paul W]

I've run into a snag.. I was hoping to use NIC teaming (link aggregation) to actually improve the throughput.

I was hoping to team 2 1Gb nics to improve speed, thus reaching close to 2Gbit throughput for my iSCSI channels.

I have Quad Intel NICs and a HP Procurve 1800 24G, and 802.3ad LACP is supported all the way. When I configure everything it all is reported to be good & working. Until I actually do some measurements: it's no faster than using a 1Gb NIC.

Then I read this: http://www.ieee802.org/3/hssg/public/ap ... 1_0407.pdf

Especially page 7 and the last page make me think I'm barking up the wrong tree. It is mentioned that to reach better speed multiple connections should be used, is that actually the case using Starwind or iSCSI for that matter?

Is teaming only good for redundancy, but won't it improve throughput? IS there anyone here that is getting more than 120MB/s through aggregated 1Gbit links?

[georgep]

For ISCSI I am getting very good results with MPIO RR.

[Paul W]

I'm currently testing with MPIO RoundRobin from server1 to server2 over 2 1Gbit ports per server.
The best speeds I see using Atto is 115MB/s writes and 113MB/s reads (at 1024k blocks) which is what I also get just using 1 Nic port.

Are your results better? What are they?

[kmax]

Does task manager show both iscsi interfaces utilized at around 50% when you do the test?

[Anatoly (staff)]

Actually the best results should show up at 32 and 64 block size

[Paul W]

Thanks guys,

I'm actually building that config as we speak; all links are teamed (802.3ad). Ive got two storage servers and two hyper-V servers. Using MPIO over the teamed links.
Using HP procurve switches that understand LACP.

[Anatoly (staff)]

I'm suggesting you to use IOmeter instead of ATTO. Setup 64 Outstanding IOps, 100% sequential, 100% read, 64 block size, disk should be unformatted.

[anton (staff)]

For sync channel use teamed links for now (upcoming versions of StarWind should use custom MPIO and get rid of MS iSCSI initiator as a sync transport as it's slow and unreliable in some cases) and use MPIO configured for Round Robin for client connections. Using NIC teaming to accelerate iSCSI traffic does not work well as iSCSI PDUs don't chunk very well.

Thanks guys,

I'm actually building that config as we speak; all links are teamed (802.3ad). Ive got two storage servers and two hyper-V servers. Using MPIO over the teamed links.
Using HP procurve switches that understand LACP.