SUMMARY: High Bandwidth (14 to 20Mbyte / sec) Disk (RAID?) systems

From: Bob Berger (
Date: Wed Apr 01 1992 - 19:42:49 CST

Well it sounds like using multiple IPI's and a VMEbus card may be the
highest performance solution. We will be looking into this in the future.

Thanks to all who sent me info. Its such a pleasure to get so much support
and high quality information from folks on the Net!

I have only put people's name who offered product info from their own company.

Original Message:

I am looking for disk subsystems that can support sustained digital
I/O of somewhere from a little more than 14 Mbytes / second to a
little less than 20 Mbytes / second. Of course, if the I/O bandwidth
is even higher, I won't complain! Capacity is not all too important as
long as it is at least 2 Gbyte or so and could be expanded to larger

It seems that one of the variations of the new SCSI II standard
combined with a RAID subsystem could do this kind of thing.

I haven't seen any RAID products that offer the high bandwidth, just
the high capacity and fault tolerence.

If you know of any products or research that come close to these
requirements, please send me the info via email. I will summarize to
the net.

------------- Summaries ---------------

Storage Concepts offers a disk array that's capable of 25MB/sec. Core
International offers a SCSI-2 array, I don't know how fast it is.
Some striped IPI's on VME may be your ticket. (
offered potential product help on this one)
In our supercomputer center we have a Maximum Strategies disk array
which claims about 440 Mb/s. It hangs directly off a Hippi switch.
    I don't know that any RAID vendor makes an array with such a high
bandwidth. You may have to go to some type of parallel transfer drive,
such as those manufactured by Ibis Systems, Camarillo, CA, (805)
In terms of existing products, you could do either of the following:

A. 2xIPI-120T (RAID-0 striped to aggregate data rate of 14.8 MB/sec) with
a disk array of up to 16xSeagate Sabre-7/2HP (RAID-1 shadowed to total data
base of 19.6GB).

B. 1xIPI-153T with CDC 99020 RAID-3 Disk Array Controller (12.5MB/sec) and
a disk array of up to 32 Sabre-7/2HP's (organized 3 data + 1 parity for up
to 24 data & 8 parity, i.e. 58.8GB). You could stripe a pair of these
for 25 MB/sec.

Both systems can hot swap drives, rebuild backup without shutting
system down, and dual port to another host. You can trade off the high
availability for additional capacity.

Within about 4-6 months, you will be able to use our new IPI-300T with
the newer Seagate Elite-3 5.25" drives and get 25 MB/sec of sustained
data rate with two drives! The IPI-303T will also allow that data rate
with a single CDC disk array.

Let me know if you need more details.

Carl A. Pick, Chairman			Office Phone: 414-644-8700
GENROCO, Inc.                  		Office Fax: 414-644-6667
205 Kettle Moraine Drive North          Voicemail: 414-791-5530	
Slinger, WI 53086   U.S.A.              INTERNET:
From: grc! (Carl Pick)
We also have similar product for Sun. IPI-120T = S212 & IPI-153T = S315.
	FAST WIDE SCSI-2 (when available) *might* be able to sustain 14-20
	MB/s -- remember that just because FAST SCSI-2 is transferring
	synchronously at 10 MB/s there are a lot of other variables that
	can limit the actual sustained rate at which you can get data 
	to/from memory from/to disk.

My first thought would be to look into Multi-head parallel IPI-2 disk technology. This is the kind of thing that Silicon Graphics uses to get the sustained transfer rate needed to do real-time graphics animation, etc. Seagate is real big on Multi-head parallel drives and make them all the way up to 8 and 9-HP which is spec'd at 24.3 and 27.3 MB/s respectively. These two particular drives are mainly made for Cray Supercomputers and, although it is actually an 8" Sabre series HDA, the actual drive is a 9-inch form-factor which is needed to house all of the extra electronics needed for the 8 and 9 parallel read channels. They do make true 8-inch form-factor versions of multi-head parallel versions of the drive but I am not sure how many channels they go up to while still maintaining true 8-inch form-factor. --------- There are 2 basic solutions to this problem:-

1) By a parallel transfer disk (PTD). These are disks which have multiple read/write channels which permit the drive to retrieve/store data on all of its platters simultaneously. The draw back is that they are very expensive ($15000+ last time I looked). Also it is not a scalable solution unless you also use 2) as well.

2) Multiple disks in a disk array. You are correct in saying that most RAID implementations are geared towards high capacity and fault tolerance.

What is required for your application is a means of writing to "N" disks in parallel from a single high bandwidth source. The problem therefore is how to permit data to flow to all disks in parallel.

Inmos is a parallel processing company that designs and manufactures Transputers (16 & 32 bit RISC link microprocessors with on chip memory and 4 20Mbits per second serial links). To permit customers to quickly build either prototype or end user systems, we provide Transputer Modules (TRAMs) that range in size from 1/2 credit card size with a 30Mhz Transputer (with onchip floating point) and 4MBytes of DRAM to 3 credit card size with Image Processing). One particular TRAM is the B422 SCSI TRAM (credit card size) converts Inmos serial links to SCSI. By using 4 of these TRAMs connected to a TRAM with a graphics frame buffer (B419) we can sustain between 5 and 6 Mbytes per second from 4 low cost SCSI disks. Typical sustained performance of 3.5inch SCSI disks are in the range 1-2MBytes per second.

One could design a custom frame buffer which had 4 transputers each responsible for its own quadrant of the screen which would provide 16 links for connection to B422's. This would give a total bandwith of 16 * 1.5Mbytes per second (24 MBytes).

B422's come with the F002C software support package which permits a C program running on a connected transputer to issue SCSI commands to any device connected to the B422.

Motherboards for TRAMs are available for PC, MCA, VME. A VME board (B016) is also available that provides a mechanism for transferring high bandwidth VME data to and from the 4 links of an on board T801 transputer which has non multiplexed address and data bus for maximum performance. Again you could use multiple B016's to increase the data throughput until you reach the endstop's of the VME Bus bandwidth.

By the way, Inmos also designs and manufactures Colour Video Controllers which provide VTG, CLUT and H/W cursor in a sigle CQFP package (1,2,4,8,15/16, 24 bits per pixel)

If this sounds at all intesting, let me know and I will arrange for the local representative to contact you (we have an office in San Jose). -- Paul Fellows iq Systems | EMail(UK) ukc!inmos!paulf Inmos Limited, 1000 Aztec West, | or Almondsbury, Bristol, BS12 4SQ, UK. | Internet: ---------------------------------------| UUCP:(US) uunet!!paulf ------------ How about a single drive IPI solution. We support single drive solutions in the 1-3GB range at the following sustained transfer rates(MB/s): 3,4.3,6,9.6,24,27. (*sector overhead is normally 10-25%)

These are very mature technologies that have been available for a few years now. The key is multi-head parallel transfer rate. With ultra-high trasfer rates, the bottleneck can be the VME bus itself(i.e. SUN). Therefore, we will soon have an Sbus IPI controller that will support all of the above transfer rates. Currently, we have customers that use the ultra-high transfer rates successfully on SGI machines.

If you think that we may have a fit please call me at 708-325-4151. I will be happy to profile a complete solution, references, and direct you to a sales representative for pricing. *********************************************************** Phil Ruff phone: (708) 325-4151 Systems Engineer fax: (708) 325-4230 National Peripherals Inc. uunet!npi!ruff -------------

I can say for certain that no RAID 5 design will give you that kind of sustained throughput, because every data block from your file is stored on one block on on one drive. In addition, for large files each drive in the array will have some of the blocks in one location, followed by parity data, followed by some more blocks from the file, but not contiguous with the previous blocks. While all the drives of the array can be accessed at once to assemble the file, the best aggregrate transfer rate you could hope for (assuming the file size is larger than the buffer of the array controller) is the transfer speed of one disk.

The other common RAID alternative is RAID 3. Here each byte is broken into bits across each channel containing disks of the array (plus some parity bits for redundancy). Thus the data is accessed in parallel across all the drives at the same time, and the transfer speed is the transfer speed of one drive times the number of channels containing your data (but not parity data). NCR in Witchita has a RAID 3 design which they claim will do 20 megabytes using a 16-bit wide SCSI-2 host interface. I also think that Dell, ALR, and Compaq Computers have RAID 3 designs as well. I'm sure there are plenty other RAID 3 designs, though. ------------- Try calling Ron Ursich. The company is called 1776


I have not used their product, but have looked over their literature and have talked with them.

It may be just what you are looking for.

Their RAID software will allow many different configuration. I don't know if it will meet your bandwidth specs, so I suggest calling them. ------------ You may be pushing current SCSI-2 implementations, I think. You might check the Ciprico SCSI-2 raid controller. Or get a HIPPI card and use something like the Maximum Strategy or IBM HIPPI disk array at 50 MB/sec sustained. -- Bob Berger - SONY Advanced Video Technology Center 685 River Oaks Parkway San Jose, CA 95134 408-944-4964 FAX: 408-954-1027 INTERNET: UUCP: [uunet,mips]!sonyusa!sfcsun!berger

This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:40 CDT