X3T9.2/87-197



                                                             November 25, 1987

To:       X3T9.2 Membership

From:     John B. Lohmeyer, NCR Principal member in X3T9.2

Subject:  SCSI Arbitration Timing Problem

Several years ago, the arbitration timing for SCSI was revised due to a 
problem known as the "Wire-OR Glitch".  An informal model for arbitration 
timing was used as the basis for this revision, but the model was never 
published (or even written down). I have attempted to capture the essential 
features of this model in a spreadsheet model, which is available to 
interested parties.  That (revised) model is summarized below:

Maximum Cable Length (MCL):                         100 feet
Propagation Speed (PS):                             0.5 feet/ns
One-Way Propagation Time (OWPT):  MCL/PS            200 ns
Clock Period (CP):                                  100 ns (min)  200 ns (max)
Driver/Receiver Propagation Delay (PD):             100 ns (each)
Maximum Wire-OR Glitch (MWOG):    2*OWPT            400 ns (max)
Glitch filter rejects (GR):  INT((MWOG+CP-1)/CP)*CP 400 ns (min)  800 ns (max)

BUS SETTLE DELAY (BSD):   2*OWPT                    400 ns (min)
BUS FREE DELAY (BFD):     BCD                       800 ns (min)
BUS SET DELAY (BSET):     CPmax+(CPmax/CPmin)*BFD  1800 ns (max)
BUS CLEAR DELAY (BCD):    (GRmax-GRmin)+CPmin+2*PD  800 ns (max)
ARBITRATION DELAY (AD):   2*OWPT+BSET+2*PD         2400 ns (min)
DATA RELEASE DELAY (DRD): 2*PD+CPmax                400 ns (max)
Minimum Time to Arbitrate: CP+GR+BFD+AD            3700 ns       7200 ns

Please note that the model assumes the arbitration logic has a clock available 
to it which may range between 5 and 10 MHz.  A 2-to-1 range was deemed 
acceptable for most protocol chips -- larger ranges require that the chip be 
informed of the approximate clock frequency.  Also note that the timings were 
developed assuming a maximum cable length of 100 feet, somewhat longer than 
the specified 25 meters.

I gave copies of my model to several people for review.  A colleague at NCR 
noticed a minor flaw in the specified ARBITRATION DELAY.  The standard does 
not appear to allow enough time for worst case propagation time through 
drivers and receivers on a maximum cable.  He suggested that the term 2*PD be 
added to the formula for ARBITRATION DELAY.  This would add 200 ns to the 
ARBITRATION DELAY increasing it to 2400 ns.

While this problem has existed for some time, I don't believe it has caused 
any problems in actual practice because very few people are using maximum 
cable lengths and most protocol chips are conservative in generating the 
ARBITRATION DELAY.  Nonetheless, we should fix the problem in SCSI-2.