Abstract

IBM commissioned The Tolly Group to evaluate the EtherStreamer MC 32 Adapter, with the main focus on measuring the application impact of full duplex Ethernet and switching on server performance. The project examined how IBM’s full duplex EtherStreamer adapter, used with switched Ethernet, affected aggregate throughput in IBM OS/2 LAN Server 3.0 and Novell NetWare 3.11 client/server environments, while also highlighting the value of an “upgrade in place” approach that reused existing cabling and standard client adapters.  

The February 1994 report tested a server equipped with IBM’s full duplex EtherStreamer MC32 micro-channel adapter and connected it to clients through standard 10Base-T infrastructure. In the baseline half duplex configuration, eight client stations connected through IBM 8222 hubs. With the server’s EtherStreamer running in half duplex mode, Tolly measured aggregate throughput of a little over 8Mbit/s. The chart on page 1 shows 8.63Mbit/s for NetWare and 8.68Mbit/s for LAN Server. Tolly notes that in this mode, NetWare’s read and write throughput were more balanced, while LAN Server showed faster reads but somewhat slower writes.  

The next test added switching and full duplex operation. The server was reconfigured as full duplex and attached at 20Mbit/s to a Kalpana EtherSwitch EPS-1500, while the client groups remained on half duplex hubs connected to 10Mbit/s switch ports. In this setup, the maximum theoretical aggregate throughput was 20Mbit/s for the server connection. Tolly measured 15.22Mbit/s aggregate throughput for NetWare and 17.03Mbit/s for LAN Server. The report notes that the full duplex and switching combination dramatically increased server throughput compared with the half duplex baseline. LAN Server, for example, added almost another 2Mbit/s of aggregate throughput relative to NetWare in this configuration.  

The final phase tested multi-interface scaling by adding a second full duplex EtherStreamer to the server and connecting it to its own switch port. Using twelve client stations across four segments, the lab saw aggregate throughput rise to 25.29Mbit/s for NetWare and 25.17Mbit/s for LAN Server, as shown in the chart on page 1 and the full duplex multi-interface test bed on page 4. The report suggests that even greater throughput would have been possible with additional client stations. Overall, the evaluation presents the EtherStreamer MC 32 Adapter as a low-cost, high-performance way to improve Ethernet server performance through the combined use of full duplex NICs and switching, while preserving existing cabling and half duplex client infrastructure.  

Note: Kalpana was acquired by Cisco Systems and their switches became the basis for the Cisco Catalyst line of switches.

This report was originally published under the InterLAB brand but was rebranded when InterLAB became The Tolly Group in 1994.

[OCR needed.]