A Tale Of Two Gateways: Bridging The Old And New
	How We Tested SNA Gateways We tested both SNA gateways on an IBM PC Server 520, dual-SMP Pentium 100-capable, with a 2-GB SCSI hard drive and two IBM Auto LANStreamer PCI Token-Ring cards. One card was attached to a LAN segment containing our AS/400, a PowerPC Model 510-2143. The AS/400 was also configured with IBM's fastest Token-Ring adapter. The second gateway server NIC was connected to another LAN segment, where our client PCs resided. To benchmark these servers, we configured eight PCs (four 486-66s and four Pentium 100s) to begin sessions at the same time and simultaneously download a 100-MB file from our AS/400 host. Client software c onsisted of NetSoft's Elite/400 for the Windows 3.1 test, and IBM's Client Access/400 for Windows95/NT. This test roughly simulates the workload encounte red by the gateway when faced with 500 very active users performing heads-down data entry and client/server applications. The test was designed to exercise the gateways and the host communications links without placing a heavy load on the host CPU. Using the same hardware and the same gateway, but different client software and OSes gave us wildly different results. But the overall theme to performance is that wider selections of vendor packages seem to perform better with Novell for SAA than with Microsoft SNA Server. IBM's Client Access/400 for Win95/NT, when tested on NT 4.0 Workstation, did not seem to follow this trend, as it showed little preference for a particular gateway product. Instead of building a generic APPC transport to work with a variety of protocols, IBM has customized se parate transports to optimize performance. The strategy seems to be valid: Overall transfer times between the two products were close, with NetWare for SAA slightly edging out NT SNA 3.0. In examining the performance numbers, you might believe that CPU load is directly related to the number of packets being sent. Using IBM's Performance Tools/400, we validated this finding. Although the Token-Ring I/O processor handles the majority of work involved, the main CPU still must get rows from the table, do some housekeeping in parsing the rows and pass them on to the NIC. The higher the throughput of the gateway, the more host CPU cycles consumed. We could find no discernable difference between the two products in host CPU usage when the data rate was considered. An exception to this rule was IBM ANYNET, which was used for comparison purposes: Although it was far slower in throughput than the gateways in our test, it averaged at least 10 percent higher CPU usage.

For more on IBM's ANYNET, Which provides support for TCP/IP on SNA hosts. ISDN PC Cards Adapters: Tiny Bundles of Remote Connectivity Pack a Punch by Jeffery L. Carrell Updated June 27, 1997