The setup is as follow : 3 Linux servers with Intel CX4 10 GigE controllers and an X-Serve with a Myricom 10 GigE CX4 controller are connected to a ProCurve 4208vl switch, with a myriad of other machines connected through good ol' 1000 base-T.
The interfaces are actually set up as 10 Gig, according to both the switch monitoring interface and the servers (ethtool, etc). However a single connection between two 10 GigE equipped machines through the switch is limited to exactly 1Gb.
If I connect two of the 10 GigE machines directly with a CX4 cable, netperf reports the link bandwidth as 9000 Mb/s. NFS achieves about 550 MB/s transfers. But when I'm using the switch, the connection tops at 950 Mb/s through netperf and 110 MB/s with NFS.
When I open several connections from 3 of the machines to the 4th, I get 350 MB/s of NFS transfer speed. So each individual 10 GigE ports actually can reach much more than 1 Gb, but individual connections are strictly limited to 1 Gb.
Conclusion : the 10 GigE connection through the switch behaves exactly like a trunk of 10 1 Gb connections. That doesn't make any sense to me, unless HP planned these ports only for cascading switches or strictly for many-clients-to-single-server connection. Unfortunately this is NOT the envisioned setup, we need big throughput from machine to machine.
Is this a not-so-known (or carefully hidden...) limitation of this type of switch? Should I suggest seppuku to the HP representative? Does anyone have any idea on how to enable a proper behaviour ? I upgraded for an hefty price from bonded 1Gb links to 10 GigE and see exactly ZERO gain! That's absolutely unacceptable.