> could you post the results for running iozone on a raw device too?
> on my ds5000/200, raw device writes are 50% slower since the integration
> of thorpej_scsipi vs. 1.5 (with the MI SCSI in both cases).

I have no good idea to run iozone on raw device.  The previous raw
figures are obtained by dd bs=32k on rsd0.

> also, could you give some more info about your setup, such as how much
> memory in your machine and the parameters you use to run iozone?

Sure.  20MHz R3000A <- corrected, 40MB MAXINE.  40MB write/read to
flood available memory.  The NCR is clocked at 25MHz has IOASIC
DMA assist.

I happen to hook PMAZ-A and PMAZB-A on the same MAXINE.  The
three TURBOchannel NCRes have internals different from each other.

Now, here is the result of RZ28, measured with 40MB of iozone.

(iozone)	K48		X1		 CHS0
PMAZ-A
	W sec	41.98		72.87		69.24
	R sec	38.93		46.16		43.16
	W B/s	  999,119	 575,587	  605,763
	R B/s	1,077,396	 908,644	  971,803

PMAZB-A
	W sec	37.31		74.11		68.37
	R sec	30.13		37.36		33.72
	W B/s	1,124,176	  565,956	  613,471
	R B/s	1,392,069	1,122,672	1,243,862

Raw write/read performance of RZ28, measured with dd bs=32k. These
look consistent between K48 and CHS0.

(dd)	PMAZ-A 		PMAZB-A
W B/s	1,104,000	1,332,000
R B/s	1,170,000	2,607,000

I ran iozone measurements on DEC3000/600.  RZ28 disk with 175MHz
21064 processor, 8bit fast SCSI transfer, 1.5X kernel, 64MB iozone
to flood memory.

(iozone)	3000/600, PMAZC-A equivalent, 1.5X
W B/s		4,142,000
R B/s		4,722,000

The figure is quite impressive and implicative.
- Same OS, same filesysystem, same SCSI automata, same SCSI driver.
- Same SCSI adaptor with parametric differences; 8bit with 12.5MHz
  TURBOchannel interconnect while 8bit-fast with 25MHz.
- Mostly same machine architecture; enhancements for larger memory
  bandwidth.
- Differs in processor, R3000 v.s. 21064, and cache capacities.

-nisimura