Douglas Gilbert wrote:
> This url is to a white paper by IBM about "packetized" scsi.
>
> http://www.storage.ibm.com/hardsoft/diskdrdl/library/whitepap/tech/hdwpacket.htm
>
> It seems to address the current speed problems of the slow async (8 bit)
> setup protocols in parallel scsi in a backward compatible way. IBM
> are already marketing what they classify as U160+ disks which include
> this feature. It will most likely be standardized when U320 and
> U640 arrive.
>
> Does anybody make a U160+ controller?
>
> Doug Gilbert
Packetized parallel SCSI is in SPI-3 which has been letter ballot approved by T10
(www.t10.org), but I know of no controller on the market that supports it yet. The
IBM URL is a white paper. I don't think there are any packetized drives from any
manufacturers including IBM in distribution. BTW oddly you won't find the term
packetized SCSI in SPI-3 anywhere it is referred to as Information Units.
I'm a big fan of packetized SCSI. The most interesting benchmark case I think
that the IBM white paper doesn't mention is a 512 byte streaming test or IOPS
test. I measured current LVD SCSI controllers (80 MB/S) using IOMETER. (I wish
Intel would release a Linux version of IOMETER. It's a great measurement tool. I've
sent them several e-mails asking them to port it to Linux at [EMAIL PROTECTED] -
maybe if others would to it would help. Also all disk access is raw there is no
file system and little OS component in these numbers. ) I found with the set-up I
had ( PIII 500 and 8 Seagate LVD drives) all leading controllers (Adaptec,
AdvanSys, Initio, Symbios, Qlogic) max out at about 20,000 IO/S.
This translates to 50 us per IO. 512 bytes of data in wide SCSI at 80 MB/S takes
6.4 us to transfer. A bus analyzer profile indicated that there was on average 15
us of bus free time per I/O. This works out to the data transfer time being only
12% of the total I/O time and 18% of the SCSI busy time. At a data rate of 160 MB/S
these percentages go down by half making the data transfer only 9% of the SCSI busy
time - that's a lot of overhead! By transferring task attributes, command, and
response bytes at the data rate instead of async (3.5 MB/S) I estimated the
SCSI busy time could be reduced from 28 us to 10 us. This makes a total IO at 160
MB/S of 28 us which would be 35,000 IO/S.
This brings parallel SCSI overhead IOPS in line with Fibre Channel. So I think it's
a given it will be adopted and will keep parallel SCSI alive at least in the
back-end in RAID cabinets.
Bob Frey
[EMAIL PROTECTED]
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [EMAIL PROTECTED]
