How NIC Array Software Packages Work
Load balancing works essentially the same way using either of the software packages we tested. Clients are assigned a NIC for inbound traffic; outbound traffic is sent via any available NIC. Even TCP sessions are unaffected by the different source MAC (Media Access Control) addresses. The network sees only one IP address for the server, eliminating the usual Windows NT problems with multihoming (such as NetBIOS names with multiple IP addresses). When a client sends an ARP (Address Resolution Protocol) request for that IP address, the server responds with one of the NICs' MAC addresses, depending on each NIC's current load.

As you might guess, this all works best in a switched or shared-media environment. If there is a router between the clients and the server, the router will respond in proxy, giving out the same MAC address to all clients. Load balancing on inbound traffic will not work in this case, but fault tolerance will.

The software can detect that a NIC is down in one of two ways: via a notice from the NDIS4 stack, or by dropped packets (NDIS3 cannot update NIC status). Both products automatically send packets out from each NIC and make sure that the information comes back in other NICs. In this way, the software can detect a downed NIC, even if the NDIS interface doesn't. In a failure situation, the server will send out an unsolicited ARP to all clients that were attached to that NIC. This forces the clients to start sending their inbound traffic to a different NIC. Communication should continue with very little interruption.

How We Tested NIC Arrays
To quantify performance, we used Ganymede Software's Chariot (long connection, file send) on a variety of Pentium and Pentium II clients running Windows NT 4.0 Workstation. All machines were equipped with PCI 100-Mbps Ethernet cards and attached to a Cisco Systems Catalyst 2926 switch. We tested scalability against a Compaq Computer Corp. ProLiant 6500 with four Pentium Pro processors and a Dell Computer Corp. PowerEdge with two Pentium II processors. For each of our two servers, we gathered performance data using one, two and three 100-Mbps adapters.

For a baseline comparison, we bound a separate IP address to each NIC and ran the Chariot benchmark. The Compaq topped out at 245 Mbps and the Dell at 205 Mbps. Neither of the NIC array products hit the mark on the Dell, but Balance Suite topped it on the Compaq, demonstrating how both of these products rely on CPU performance.