2 Gbit RX with HP Teaming and Cisco 6500 series switch

[Gera2]

Hello

We have a HP DP backup server DL 585 G1 with 2 integrated NC 7782 NICs connected to Cicso 6500 series switch with Port Aggregation Protocol (PAgP) enabled on
A network bottleneck of 1 Gbit during many backup sessions is identified in this system. I am trying to team 2 onboard adapters to get 2 Gbit RX.
In HP NCU I set Team Type Selection as SLB and Transmit Load Balancing Method as Automatic. Now Information tab shows both links as 1000/Full Tx/Rx. All other settings left by default. But stressing server with, for example, iperf, or many backup sessions, Team Utilization Chart shows that team bandwidth refuses to go over 1 Gbit. Load Balancing is working, but every link canâ t be loaded over around 480 Mbps, in sum no more than 985 Mbps.

Is it possible to resolve this problem? A report from info tool attached.

Thanks,
Gera, BDC

[rick jones]

Unless the switch implements the potentially bad "round-robin" packet scheduling policy, packets of any single "flow" will only travel across one link in an aggregate. This is to avoid reordering the traffic and causing TCP performance to tank. At best RR and reordering means more ACKs from TCP (and thus more CPU util) at worst it means even lower throughput as TCP is tricked into believing there is congestion when there isn't.

If you have only one TCP connection, perhaps the switch is shifting traffic from one link to another, which is why each looks only half-loaded.

So, the first thing to do is make sure you have multiple TCP connections going.

After that, you have to examing the packet scheduling algorithm in use to make sure that the switch's definition of a "flow" is one that will tell one of your connections from another.

If it is just using destination MAC (ethernet) address then all traffic, regardless of number of TCP connections, to the server will use only one link.

Similarly if it is using just destination IP address, then all traffic, regardless of number of TCP connections, to the server will use only one link of the aggregate.

You need the switch to include the TCP port numbers (source and destination) in its algorithm to get traffic from the switch to the server spread across both links in the aggregate.

(Didn't see the attachment, browser didn't know how to display it)

If you need a single connection to go faster than 1 Gbit/s you _could_ try experimenting with round-robin (and risk things being worse). However the proper "fix" is to upgrade to 10G.