However, we should point out that Silverback's test bed was what the industry terms a "racing configuration": It wasn't a real-world setup.
For example, the test servers weren't accessing disk storage. Silverback set up the test to access RAM (random access memory), "so that we were measuring the performance of the networking, not the disk," says Mike Strickland, the company's director of product marketing. He adds that the benchmark "is actually quite representative of what a storage product would see for cached reads and writes."
Silverback tested its iSNAP 2100 chip using Iometer, a data throughput measurement tool developed by Intel Corp. (Nasdaq: INTC), with 512-byte block sizes. It set up four Windows 2000 servers running the beta version of Microsoft's iSCSI initiator, connected via Gigabit Ethernet to a Linux-based PC with a 2.4GHz Intel Xeon processor that acted as the iSCSI target. All servers used ServerWorks Corp. Gigabit Ethernet chipsets.
The company claims it was unable to measure CPU utilization because it hasn't been able to get Iometer to work with its Linux initiator. In the coming weeks, Silverback says, it plans to be able to measure the CPU utilization of a system running its chip; Strickland says he expects it to be somewhere around "10 percent or less" using 8K block sizes.
The iSNAP 2100 is a single-chip implementation that has dual Gigabit Media Independent Interfaces (GMIIs) and a PCI-X port. It includes external memory, which is "typically no more than 24 Mbytes" of external DRAM. So far, the company has developed only a Linux driver but claims it can port to any operating system.