[comp.os.vms] Thoughts On Sequencing Of RA Type Devices On HSC Requestor Cards...

CLAYTON@XRT.UPENN.EDU ("Clayton, Paul D.") (09/20/87)

Information From TSO Financial - The Saga Continues...
Chapter 23 - September 19, 1987

John Macallister has asked the following question regarding the mixing of
RA81 and RA82 type devices on the same requestors in an HSC.

	We're installing an HSC50 and have a mixture of RA81/RA82 disks. To
	balance the load it would be necessary to mix RA81's and RA82's on
	the same controller. Any known problems with this given that the
	disks have slightly different access times?

While it 'sounds' right to mix the RA81 and RA82 devices on the various 
requestor cards in an HSC I STRONGLY recommend AGAINST it. If this is suggested
by your local field service office, suggest to them they read the internal DEC
mail from DEC West Disk gurus. The HSC code SHOULD be at rev. level 3.5, or 
3.51 if you have shadowed devices, and there is a SPECIFIC order to the 
requestors now. I talked about this order in a past article of mine and some 
problems (horrors) I have been having. 

The order is as follows.

	CPU
	CI Link Module
	Requestor Card(s) For Tape Devices
	.
	.
	Requestor Card(s) For RA81 Devices
	.
	.
	Requestor Card(s) For RA82 Devices
	.
	.

Now there is nothing you can do about combining RA81 and RA82 devices on the 
same requestor card when crossing from one type to the other. If the above
sequence is not followed, the result, at least in my case, was somewhere 
between 1 and 100 'SDI Collision' errors per day. DEC resequenced my requestors
and devices and there has not been a 'SDI Collision' error for weeks.

The other area that you will be finding out is that the HSC50 just isn't fast
enough setting up the commands for the faster drives. The PDP11 inside the HSC 
is performing I/O decomposition, sequencing and disk management. We have RA81,
SI83C and SI93C drives and in comparing the 'throughput' between an HSC50 and 70
the 70 will perform 10 to 20% more I/O's then the 50. It also needs to be 
pointed out that the actual transfer of data to/from the disk/CI is NOT done
by the PDP11, but is performed by the bit-slice processors on the requestor
cards themselves. In summation, with the faster drives, more I/O's can be 
completed in the same period, and this creates a new bottleneck, that being
the HSC. If Shadow Set software is in use the bottleneck can be worse because
extra comparisons are done in generating the 'sequencing' of I/O command 
packets to the disks so that you receive the data from the disk that has it
closest at hand.

It has never been said, that I know of, that VAXClusters are easy to 
understand. This area proves it once again. Hope the information helps.

Paul D. Clayton - Manager Of Systems
TSO Financial - Horsham, Pa. USA
Address - CLAYTON%XRT@CIS.UPENN.EDU