Cisco, IBM to enhance cluster load balancing

By Jim Duffy
Network World, 09/11/00

RESEARCH TRIANGLE PARK, N.C. - Cisco and IBM will soon unveil software designed to enhance traffic load balancing across clustered servers.

The software, called Generic Routing Encapsulation (GRE), is intended to overcome the limitations of Cisco and IBM load balancing techniques for equalizing client requests and server responses among IBM S/390 systems in a Sysplex cluster. The Cisco/IBM GRE implementation is based on technology defined in Internet Engineering Task Force RFC 1701 and 1702 documents, which means the software can work in multivendor networks, and is a result of the two vendors' $2 billion networking alliance (See story).

Currently, IBM's Network Dispatcher and Cisco's Multi-Node Load Balancer (MNLB) products have two restrictions. The first is the load balancer has to be on an adjacent subnet to the servers, or it has to translate network addresses between inbound and outbound traffic. The second is there has to be a one-to-one relationship between the application server and the physical connection, such as IBM's OSA Express adapter, which lets users directly attach LANs to the mainframe at gigabit speeds.

If the load balancer is in an adjacent subnet, it is an IP hop away from each application server. Client requests can therefore be forwarded at Layer 2 with the client's IP address in the header. If the target servers are not on that network, address translation must be utilized to provide a new Layer 3 source-destination address.

The significance of this restriction is that translated addresses must be retranslated, forcing return traffic from the server application to flow through the same translation device. Because server-generated response traffic is typically much larger than the client request, this could present a bottleneck, IBM and Cisco say.

The second restriction is significant in terms of cost. Currently, the IBM and Cisco load balancers require users to install an OSA Express per application server. But one S/390 in a Sysplex cluster can be virtually divided into 15 servers, or logical partitions (LPAR). And when you can have up to 32 S/390s in a single Sysplex cluster, 480 is a lot of OSA Express adapters to purchase just for load balancing between LPARs.

GRE encapsulation lets the 15 LPARs share the OSA Express adapter's MAC address while hiding the unique IP addresses of each application within an LPAR. Thus, applications within each LPAR appear to be in an adjacent subnet, or one hop away from the load balancer.

A large aerospace company in Bellevue, Wash., is looking to GRE to help balance the load on a Web application running across four LPARs in a Sysplex.

"Our intent is to run two OSAs per physical [S/390] machine and then share them between the LPARs within that machine," says the company's senior network architect, who spoke on the condition of anonymity. "[Currently,] if you're sharing OSA with multiple LPARs, it doesn't know how to distribute the connections."

GRE will alleviate that by enveloping TCP/IP packets for transport through MNLB and OSA Express, and then stripping off that envelope and sending the packet to the proper TCP/IP stack within the Sysplex, he says.

GRE is an enhancement to IBM's Communications Server for OS/390 software. It will be available next week as a free maintenance upgrade to users of OS/390 Release 8.