[comp.os.vms] Thoughts On Single Versus Dual Processors, And Impact of VMS 4.x/5.x...

CLAYTON@xrt.upenn.EDU ("Clayton, Paul D.") (02/08/88)

A question from Rand P. Hall on single processor versus dual processor has
been asked, and to a certain degree answered here at TSO. We have a 8350
which was purchased for one purpose only. To run jobs written using the FOCUS
language. I requested an 8530/8550 cpu to perform this function but the powers 
at large made a dollars decision and the 8350 walked in the door. I was not 
ALL that unhappy as I now have a platform for verification of VMS 5.0 and the 
new SMP abilities built into it. If things worked well with v5.0, then I was 
recommending upgrading my 8700's to 8800's.

The facts from the 8350, and VMS 4.X, are that despite having dual processors 
you have to realize that only ONE, the master, is able to perform ANY I/O. The 
secondary processor is exclusively compute mode. There were no changes, and 
none are needed for anything, to run on a 8350 using VMS V4.X. The process 
will switch between the different processors according to how much compute and 
I/O it is doing. VMS has handled this feature of ASMP very nicely from my 
point of view.

The SMP abilities of VMS 5.X will eliminate the 'master/slave' relationship 
between the two processors. Both will be free to perform I/O AND be a 'compute 
mode' processor at the same time. Note that a restriction of master/slave 
remains in existance as far as the backplane of the processor is concerned. 
Only one interface/nexus can have control of the bus at any one given point in
time. There are a few limited exceptions to this rule, but these are based in 
hardware configurations and not software, by and large.

Therefore the biggest impact is to have more I/O bandwidth of the bus or 
backplane used when V5.X arrives. 

There are tricky areas that SMP will allow according to recent articles. Those
dealing with programs that are designed/coded in order to have different 
sections of the code executing on different processors AT THE SAME TIME. 
Needles to say this raises all sorts of questions about sequencing of events 
that have never been raised before when computers where based on the Von 
Nuehman principles. This is neat stuff if the need is there. 

There is something else to consider, which my powers at large refused to 
listen to, when dealing with single versus multi processors on the same bus. 
The 8350 is TWO 1.5 MIPS processors, while for the sake of comparison a 8530 
is a SINGLE 3.9 MIPS processor. DO NOT THINK THAT THE 8350 IS THE EQUILALENT 
OF THE 8530, IE. 2 * 1.5 IS NOT GREATER THEN OR EQUAL TO 3.9. By comparing the
number of instructions executed per quantum, the following comes to light.
	 30,000 instructions per quantum at 1.5 MIPS
	120,000 instructions per quantum at 6 MIPS
I have not found suitable mathmatics to describe the model of how VMS works
on a single or dual processor, but can offer the following example from a real 
life case study we did.

Using FOCUS to generate a report from data that was contained in several disk 
files, each of which was over 1.2 MILLION BLOCKS, yields the following
comparison.

Dedicated 8350 system, elapsed time 13:38:11.93, CPU time 11:55:38.17
	Both processors enabled.
Dedicated 8530 system, elapsed time 3:52:07.97, CPU time 2:32:05.64

My only remaining comment concerning FOCUS is that, as far as system resource 
utilization is concerned, it makes ALL-IN-1 look FANTASTIC.

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