Scott:
>NT is *very* CPU bound because it is a bulky, general purpose, >preemptive OS. NT *has* to have support for multi-processors because >it will eat them up to try and provide any kind of
>performance. Adding I2O to a CPU bound OS is useless. NT will not >see the benefits until they clean up their OS and move more code to >unprotected ring-0 operation.
Close. Actually the simple differentiator is that an I/O bound system will gain more performance out of a faster I/O subsystem if all things are equal.
Since I20 will enable a CPU to offload some processing power to the disk subsystem then I contend that an NT-based system will recover MORE CPU cycles than a Netware server. (I have not been inside the NT or Netware disk driver software, however so unless your engineers have decompiled them or benched them then we will wait and see)
>NetWare, on the other hand, is an extremely efficient I/O engine and >is usually I/O bound! NetWare, at it's core, is designed to pump >data between high-speed networks and high-capacity disk. A single >processor BorderManager is currently showing a minimum of two to >three times the performance of a *four* processor NT box! And the NT >processors are max'd out at 90+% utilization ...
In theory this sounds reasonable, however in practice this has not been the case. I have personally designed firmware and interfaces between IBM's SSA disk and the HP3000 family of servers. (The 3000 and the HP9000 are the same hardware, only the HP9000 runs HPUX, and the HP3000 runs MPE).
MPE can be equated to netware in it's ability to get about 3x more work out (very generous, but go with me here) of the same hardware because of the operating system and it's IO efficiency. What we found is that the SSA adapter gave 2 to 4x better performance on UNIX and 0x to 3x better on MPE. Furthermore, we found many MPE systems to get a statistically insignificant performance increase.
On analysis, the reason turned out to be that most MPE systems weren't as I/O bound as we thought. Their disk drivers were so efficient that they just didn't see the benefit that we thought they would.
This is also not just my research. EMC disk subsystems exhibit the same performance drop between these two operating systems. Granted every architecture and O/S is different, but don't make the generalization that I20 will benefit one OS over another! |
