The Risks Digest

The RISKS Digest

Forum on Risks to the Public in Computers and Related Systems

ACM Committee on Computers and Public Policy, Peter G. Neumann, moderator

Volume 4 Issue 58

Sunday, 8 March 1987

Contents

o The Sperry Plan, FAA Certification, and N-Version Programming (Nancy Leveson)
LONG MESSAGE
o Info on RISKS (comp.risks)

The Sperry Plan, FAA Certification, and N-Version Programming

Nancy Leveson <nancy@ICSD.UCI.EDU>
07 Mar 87 10:34:16 PST (Sat)
Since my original message to Risks about the Sperry plan, I have visited
the FAA Certification Office in Seattle and, for other reasons, the Boeing
Commercial Aircraft Co. in Seattle.  The Boeing employees I spoke to told
me that they have rejected the Sperry plan for their software.  However,
it is planned to be used for a Category III autopilot for the MD-11.  This
autopilot is safety-critical for 30-45 seconds during autolanding.  There is
also great dependence on n-version programming for the fly-by-wire Airbus
320, but I have no details about the A320 software development such 
as the testing procedures used (except for a blurb in Aviation Week that 
states that n-version programming provides "optimum" protection against 
software errors in that aircraft).

To remind Risks readers, Sperry wants to use N-version software in place of
"white box" (structural) testing.  Black box (system) testing will still 
be performed.  System testing would be done "back-to-back."  Back-to-back
testing means that multiple versions of the software are executed on the
same test data.  If they all agree on the answer, it is assumed that they
are all correct.  

Kevin Driscoll writes in Risks 4.55 that this plan is not really so scary.
In order to follow the discussion, it is necessary to have some background
on aircraft software certification.

There is a document used by the FAA (and written primarily by the
manufacturers) called RTCA/DO-178A: Software Considerations in Airborne 
Systems and Equipment Certification.  It lays out the requirements for 
software development, testing, configuration control, and documentation.  
The requirements are pretty basic -- about what I would recommend for a 
good inventory program.  In general the requirements are: 

  (1) developing software requirements and verifying them against system 
      requirements [no requirement for any formality in the process];
  (2) using a design discipline or method [not specified as to which 
      one -- just says you need to use one] that makes software traceable, 
      testable, maintainable, and understandable;
  (3) doing a design review against requirements;
  (4) using an implementation technique that is understandable, testable,
      and traceable to the design;
  (5) doing requirements-based and structure-based tests including module 
      testing, module integration testing, and hardware/software integration 
      testing including a requirements coverage analysis and a software 
      structural coverage analysis [this is what Sperry wants to eliminate
      except for system test]
  (6) providing software configuration management;
  (7) providing a quality assurance plan.

The document specifies software function criticality categories of: 
   Level 1 (flight-critical: failure prevents continued safe flight), 
   level 2 (flight essential: functions reduce capability of aircraft 
            or ability of crew to cope with adverse conditions), and 
   level 3 (non-essential: failures could not significantly degrade 
            aircraft capability or crew ability).  

The difference in criticality level seems to determine what information
is provided to the FAA for certification and, in some cases, which of the
above requirements are enforced.  For level 1 software, for
example, the manufacturer must provide detailed information about the
verification that was done.  For level 2, in general only a summary
description of the process along with a statement of compliance must
be submitted.  For level 3, no assurance is required.  In terms of
certification effectiveness, independent evaluation is possible only with
information.  So providing just a statement of compliance seems to me
to imply that no external, independent evaluation is possible.  There is
no way to check that they actually did comply and that the verification
that was done was adequate and correct.  I certainly do not want to
imply that the manufacturers and subcontractors will not try to do the
best job possible -- after all, they have the liability and they are
decent human beings who care about human life.  The problem is that 
without external review we are depending on the competence 
of the people at these companies, and I am not as sanguine about the 
general state of software engineering knowledge and practice in industry 
as I am about the good intention of humans.  

So far, though, things are not really TOO awful, but wait ...
The problem seems to arise from one sentence (which was added between 
version 178 and 178A and seems to be the major change) that states 
"Using appropriate design and/or implementation techniques and
considerations, it may be possible to use a software level lower than
the functional categorization."  This is the kicker.  Sperry is arguing
that although the software autopilot is Level 1, they are using n-version
programming and therefore it can be treated as Level 2.  There is
also a phrase "the software level implies effort that ... may vary
within criticality level."  So they can modify any of the requirements
also, it appears, within level (given that the FAA agrees).  
BOTTOM LINE:  even those very basic requirements that are specified above
can be eliminated fairly easily.  Personally, I would require MORE than is
stated in DO-178A for both Level 1 *AND* Level 2 software development and 
verification.

As examples of what is possible, the DoD, besides requiring good software 
engineering practice, requires a safety and hazard analysis of the software.  
The Air Force and Navy also require an IV&V by a qualified company (Logicon 
does a lot of this) for all nuclear systems [called Nuclear Safety Cross 
Check Analysis by the Air Force and Software Nuclear Safety Analysis by 
the Navy].   These IV&V efforts are MUCH more rigorous than anything 
the FAA appears to be doing.  Note that the DoD requires proof of the 
safety of the software itself and not just proof that the developers have 
satisfied minimal development practices.

The most amazing part of the RTCA document is the fact that using a particular
method, such as n-version programming, can somehow magically change the
criticality level of the software (from flight-critical to flight-essential
or non-essential).  Since the function of the software does not change
with the development method, this appears ridiculous.  I can only assume
that they are arguing that the reliability will be so high that failures
will never occur and therefore the criticality of the function is irrelevant;
this is the only interpretation that makes sense to me.  But there is no
current software engineering technique that can guarantee such ultra-high
reliability! (including N-version programming). And since they dismiss in 
the document the use of any measurement techniques (they state that currently 
available methods do not yield results in which confidence can be placed to 
the degree required) and don't even mention any formal verification methods, 
there is NO demonstration required that they have reached perfection (or any 
particular level of reliability) using the particular design or implementation 
technique.  

In the Sperry case, their argument for N-version programming appears to
rest on a simplistic model presented by Larry Yount at an AIAA Conference
in Long Beach.  This model assumes statistical independence between failures 
of the n versions.  This assumption has been shown to not hold in
controlled experiments and, in fact, is not believed by most researchers 
in the field.  At a workshop this summer, Larry put up a chart that showed 
his model predicted 20,000 times improvement in reliability based on the 
use of n-version programming.  Since actual experiments have found at best 
only 7-10 times improvement, his figures appear to be patently ridiculous. 

Kevin Driscoll (Risks 4.55) states:

   > In its letter to Sperry, the FAA says that this method "appears to be
   > satisfactory" with the following constraints:
   > a.  Level 1 must be used for paragraphs 6.2.2 (Requirements Development
   > and Verification) and 6.2.3 (Design).
   > b.  Formal configuration control must used and, if common errors are
   > found, structural testing may be required for some or all of the
   > modules.

Common errors have been found in EVERY experiment done so far in n-version
programming (at least, in all that have checked for them which is about 4
or 5).  The problem is that with only three versions of the software and the
use of back-to-back testing, the only common errors that can be detected are
those within only two modules.  Any common errors found in all three of the
modules cannot be detected (unless some outside method of correctness
determination is used).  In my experiment with John Knight at the University
of Virginia, we found common failures in up to eight independently developed
programs.  Also, any errors that can be traced back to the specification 
will, of course, have a tendency to manifest themselves in common between 
the versions.

   > c.  Formal review and comparison of source code must be used to verify
   > dissimilarity.  Where this is not feasible, Level 1 structural test
   > and analysis must be used.

How does one verify dissimilarity?  In fact, how does one even define it?
Obviously the programs must be similar in that they are computing the
same function.  The only dissimilarity we really want is a dissimilarity in
failure behavior.  Syntactic dissimilarity is irrelevant.  Again, John Knight
and I found programs that used completely different algorithms to compute a
function yet failed on all the same input data.  The problem is that certain
input cases are inherently more difficult to handle.  For example, when
computing the angle determined by three points, programs tended to fail on
inputs where the points were colinear or coincident.  The errors were not
the same nor were the algorithms, but they failed on the same input data.
So looking to see that different algorithms are used is not adequate.  This
is the problem in talking about a concept like "dissimilarity" or "diversity"
without ever formally defining it;  there is no way to know whether you
have it nor any way to measure it.  It is similar to the problem with using 
the term "artificial intelligence" when the term "intelligence" remains
undefined.  One can merely claim that their program is intelligent and it
is difficult to dispute it (or to prove it either).  How does one prove or
disprove that dissimilarity or diversity exists?

   > It seems to me that c. is the same as doing structural analysis. 
   > Therefore, this method is not any less rigorous than "full" DO-178A
   > Level 1.

I can see no relationship between verifying dissimilarity between two
or three programs and structural analysis of the correctness of a single 
program;  especially given that I know how to do the second but not the 
first.  I am not quite sure what Kevin means by "less rigorous."  Certainly, 
we have much more experience with structural testing than with n-version 
programming.  There is no evidence anywhere that structural testing is 
equivalent to n-version programming (e.g., that they detect the same errors)
nor that one can replace the other.  Although somewhat beside the point, 
I would argue that even the *FULL* DO-178A is not nearly rigorous enough for 
safety-critical software.

   > d. Functional tests of the system must be performed.  It must be shown
   > that the system will not have false alarms.

   > However, how one complies with c. and d. I do not know.

THAT IS THE WHOLE POINT!  Sperry is suggesting replacing something we
know how to do with something nobody knows how to do and has never been
shown to work with the degree of effectiveness required.  I would certainly
feel happier if the Sperry plan were tried first on real software 
that was not Level 1 or Level 2 (by real software, I do not mean just
university or industrial experiments where the software is never used in a
real production environment).  I have few qualms about N-version programming
being used in conjunction with normal software development techniques
even on safety-critical software.  But I have grave reservations about 
eliminating any testing or other standard procedures on the basis of
using it.  The problem, of course, is that developing multiple versions
is expensive.  So I assume Sperry is trying to cut down on testing in
order to save money.  Unfortunately, I do not know how to develop
safety-critical software cheaply.  For the most part, greater reliability
and safety requires more money.  Just using some sleight of hand to relabel
the software as Level 2 or Level 3 instead of Level 1 does not make it
any less safety-critical.  And voting together relatively untested and
unverified single versions has not been shown (in the experiments that
have tried it) to guarantee high reliability or safety.  In fact, the
little experimental evidence available has shown that as the number of
errors in the individual versions increases, the amount of reliability 
gain to be expected fron using n-version programming decreases.

I am still worried despite Kevin's attempt at reassurance.

    Nancy Leveson,          University of California, Irvine

Please report problems with the web pages to the maintainer

Top