iSCSI, Switches and NICs Oh My!

A recent post by Marc Farley over at Equallogic brings to light an important fact, not all switches are created equal. In the fiber world, administrators are well aware of the cost of fiber switches, maybe painfully aware. Because iSCSI delivers the prospect of a lower TCO, some folks rush to the conclusion that any gigabit ethernet switch will do as they are not forced into a limited selection of vendors like on the fiber side, where Cisco, Brocade and McData seem to rule.

Unfortunately, many of them will come to the harsh realization that low end switches are not ready for production environments or environments which support a lot of IO. Couple the performance implications of switches with that of subpar NIC hardware and you have a recipe for disaster.

For example, the recommendations for iSCSI on an Equallogic device are to utilize flow control, jumbo packets and disable unicast storm control. In order for the switch settings to be effective, the NIC has to support the same options. Enabling jumbo packets on a switch does nothing if the NIC cannot support the setting. If the NIC does not support flow control, then again the flow algorithms will not be effective in controlling traffic from end-to-end. Additionally it is recommended that the NIC support a 2048 receive buffer, especially for applications like SQL, to ensure packets are not lost. If the NIC buffer is too small, packets may be lost during rapid network transitions.