Jim -
You are on very weak technical ground. The SCSI drives, bus control chips and so on are about as much connected to performance as the gasoline is to performance in a car. Still a Ferrari is faster than a Ford LTD even though both use premium unleaded -
As a rather basic example, many UNIX systems use a CAM-based async driver - this allows a single driver to handle pretty much the whole range of possible SCSI devices, but has several disadvantages, among them twice as many interrupts (and context switches) as a driver designed to enhance transaction performance. A driver specifically tailored to optimize transaction performance, with a sufficiently strong subsystem, can execute a transfer with NO interrupts or context switches.
The differences in code path are stunning. A scatter-gather transfer of a megabyte of data from application space to disk using a CAM driver executes more than a MILLION instructions and performs 80 context switches, each of which flushes the I and D cache and causes a processor stall. The same operation on an enhanced driver requires 150,000 instructions and 40 context switches. An optimized driver can execute in less than 30,000 instructions with No context switches. These translate to aggregate system I/O levels of 2000, 8000 and 40,000 I/O per second... clearly, CPU speed will not overcome a 20X architectural advantage.
BTW, the NT drivers in current use with the standard release are equivalent in performance to the CAM model... |
