## Shocking Tell-All Interview on Software Assurance

August 29, 2010

I was recently interviewed by Flight International magazine, one of the oldest aviation news magazines.  Their reporter, Stephen Trimble, was writing on the Air Force’s Chief Scientist’s recent report stating that new software verification and validation techniques are desperately needed.

Here’s an online copy of the article.

## Copilot: A Hard Real-Time Runtime Monitor

August 22, 2010

I’m the principal investigator on a NASA-sponsored research project investigating new approaches for monitoring the correctness of safety-critical guidance, navigation, and control software at run-time.  We just got a paper accepted at the Runtime Verification Conference on some of our recent work developing a language for writing monitors.  The language, Copilot, is a domain-specific language (DSL) embedded in Haskell that uses the powerful Atom DSL as a back-end.  Perhaps the best tag-line for Copilot is, “Know how to write Haskell lists?  Good; then you’re ready to write embedded software.”

Stay tuned for a software release and updates on a flight-test of our software on a NASA test UAV…  In the meantime, check out the paper!

## Twinkle Twinkle Little Haskell

May 31, 2010

Update Sept 28,2010: the Makefile mentioned below worked fine, except for something having to do with timing.  I was too lazy to track the problem down, but fortunately, I found an scons script (using the scons build system) that I modified to run on Mac OSX, and it works perfectly.  The original script is here—thanks Homin Lee!  The post has been modified appropriately.

Update Oct 1, 2010: Homin Lee has updated the script to work on Mac OSX, so you can just grab the original script now.

It’s been a few months almost a year(!) since John Van Enk showed us how to twinkle (blink) an LED on his Arduino microcontroller using Atom/Haskell.  Since that time, Atom (a Haskell embedded domain-specific language for generating constant time/space C programs) has undergone multiple revisions, and the standard Arduino tool-chain has been updated, so I thought it’d be worthwhile to “re-solve” the problem with something more streamlined that should work today for all your Haskell -> Arduino programming needs.  With the changes to Atom, we can blink a LED with just a couple lines of core logic (as you’d expect given the simplicity of the problem).

For this post, I’m using

If you’ve played with the Arduino, you’ve noticed how nice the integrated IDE/tool-chain is.  Ok, the editor leaves everything to be desired, but otherwise, things just work.  The language is basically C with a few macros and Atmel AVR-specific libraries (the family to which Arduino hardware belongs).

However, if you venture off the beaten path at all—say, trying to compile your own C program outside the IDE—things get messy quickly.  Fortunately, with the scons script, things are a piece of cake.

What we’ll do is write a Haskell program AtomLED.hs and use that to generate AtomLED.c.  From that, the scons script will take care of the rest.

## The Core Logic

Here’s the core logic we use for blinking the LED from Atom:

ph :: Int
ph = 40000 -- Sufficiently large number of ticks (the Duemilanove is 16MHz)

blink :: Atom ()
blink = do
on <- bool "on" True -- Declare a Boolean variable on, initialized to True.

-- At period ph and phase 0, do ...
period ph $phase 0$ atom "blinkOn" $do call "avr_blink" -- Call a locally-defined C function, blink(). on <== not_ (value on) -- Flip the Boolean. period ph$ phase (quot ph 8) $atom "blinkOff"$ do
call "avr_blink"
on <== not_ (value on)

And that’s it!  The blink function has two rules, “blinkOn” and “blinkOff”.  Both rules execute every 40,000 ticks.  (A “tick” in our case is just a local variable that’s incremented, but it could be run off the hardware clock.  Nevertheless, we still know we’re getting nearly constant-time due to the code Atom generates.)  The first rule starts at tick 0, and executes at ticks 40,000, 80,000, etc., while the second starts at tick 40,000/8 = 5000 and executes at ticks 5000, 45,000, 85,000, etc.  In each rule, after calling the avr_blink() C function (we’ll define), it modulates a Boolean upon which blink() depends. Thus, the LED is on 1/8 of the time and off 7/8 of the time. (If we wanted the LED to be on the same amount of time as it is off, we could have done the whole thing with one rule.)

## The Details

Really the only other thing we need to do is add a bit of C code around the core logic.  Here’s the listing for the C code stuck at the beginning, written as strings:
 [ (varInit Int16 "ledPin" "13") -- We're using pin 13 on the Arduino. , "void avr_blink(void);" ] 
and here’s some code for afterward:
 [ "void setup() {" , " // initialize the digital pin as an output:" , " pinMode(ledPin, OUTPUT);" , "}" , "" , "// set the LED on or off" , "void avr_blink() { digitalWrite(ledPin, state.AtomLED.on); }" , "" , "void loop() {" , " " ++ atomName ++ "();" , "}" ] 

The IDE tool-chain expects there to be a setup() and loop() function defined, and it then pretty-prints a main() function from which both are called. The code never returns from loop().

To blink the LED, we call digitalWrite() from avr_blink(). digitalWrite() is provided by the Arduino language.  (In John’s post, he manipulated the port registers directly, which is faster and doesn’t rely on the Arduino libraries, but it’s also lower-level and less portable between Atmel controllers.)  Atom-defined variables are stored in a struct, so state.AtomLED.on references the Atom Boolean variable defined earlier.

## Make it Work!

Now just drop the scons script into the directory (the directory must have the same name as the Haskell file, dropping the extension), and run
 > runhaskell AtomLED.hs > scons > scons upload 
And your Haskell should be twinkling your LED. runhaskell AtomLED.hs invokes the Atom compile function to generate a C file and headers; scons invokes the build script to build an ELF image to upload, and scons upload again invokes the compiler to upload to your board.

This should work for any Atom-generated program you want to run on your Arduino (modulo deviations from the configuration I mentioned initially). Also, note the conventions and parameters to set in the scons script.

Post if you have any problems, and I might be able to help. Also, I’d love to package the boilerplate up into a “backend” for Atom, but if you have time, please beat me to it.  Thanks.

Code:

## New Group: Functional Programming for Embedded Systems

May 30, 2010

Are you interested in how functional programming can be leveraged to make embedded-systems programming easier and more reliable?  You are not alone.  For example, check out what’s been happening in just the past couple of years.

Now Tom Hawkins (designer of Atom) has started a Google group, fp-embedded, to discuss these issues.  Please join and post your projects & questions!

## An Apologia for Formal Methods

March 14, 2010

In the January 2010 copy of IEEE Computer, David Parnas published an article, “Really Rethinking ‘Formal Methods’” (sorry, you’ll need an IEEE subscription or purchase the article to access it), with the following abstract:

We must question the assumptions underlying the well-known current formal software development methods to see why they have not been widely adopted and what should be changed.

I found some of the opinions therein to be antiquated, so I wrote a letter to the editor (free content!), which appears in the March 2010 edition.  IEEE also published a response from David Parnas, which you can also access at the letter link above.

I’ll refrain from visiting this debate here, but please have a look at the letters, enjoy the controversy, and do not hesitate to leave a comment!

## 10 to the -9

January 24, 2010

$10^{-9}$, or one-in-a-billion, is the famed number given for the maximum probability of a catastrophic failure, per hour of operation, in life-critical systems like commercial aircraft.  The number is part of the folklore of the safety-critical systems literature; where does it come from?

First, it’s worth noting just how small that number is.  As pointed out by Driscoll et al. in the paper, Byzantine Fault Tolerance, from Theory to Reality, the probability of winning the U.K. lottery is 1 in 10s of millions, and the probability of being struck by lightening (in the U.S.) is $1.6 \times 10^{-6},$ more than a 1,000 times more likely than $10^{-9}.$

So where did $10^{-9}$ come from?  A nice explanation comes from a recent paper by John Rushby:

If we consider the example of an airplane type with 100 members, each flying $3000$ hours per year over an operational life of 33 years, then we have a total exposure of about 107 flight hours. If hazard analysis reveals ten potentially catastrophic failures in each of ten subsystems, then the “budget” for each, if none are expected to occur in the life of the fleet, is a failure probability of about $10^{-9}$ per hour [1, page 37]. This serves to explain the well-known $10^{-9}$ requirement, which is stated as follows: “when using quantitative analyses. . . numerical probabilities. . . on the order of $10^{-9}$ per flight-hour. . . based on a flight of mean duration for the airplane type may be used. . . as aids to engineering judgment. . . to. . . help determine compliance” (with the requirement for extremely improbable failure conditions) [2, paragraph 10.b].

[1] E. Lloyd and W. Tye, Systematic Safety: Safety Assessment of Aircraft Systems. London, England: Civil Aviation Authority, 1982, reprinted 1992.

[2] System Design and Analysis, Federal Aviation Administration, Jun. 21, 1988, advisory Circular 25.1309-1A.

(By the way, it’s worth reading the rest of the paper—it’s the first attempt I know of to formally connect the notions of (software) formal verification and reliability.)

So there a probabilistic argument being made, but let’s spell it out in a little more detail.  If there are 10 potential failures in 10 subsystems, then there are $10 \times 10 = 100$ potential failures.  Thus, there are $2^{100}$ possible configurations of failure/non-failure in the subsystems.  Only one of these configurations is acceptable—the one in which there are no faults.

If the probability of failure is $x,$ then the probability of non-failure is $1 - x.$  So if the probability of failure for each subsystem is $10^{-9},$ then the probability of being in the one non-failure configuration is

$\displaystyle(1 - 10^{-9})^{100}$

We want that probability of non-failure to be greater than the required probability of non-failure, given the total number of flight hours.  Thus,

$\displaystyle (1 - 10^{-9})^{100} > 1 - 10^{-7}$

which indeed holds:

$\displaystyle (1 - 10^{-9})^{100} - (1 - 10^{-7})$

is around $4.95 \times 10^{-15}.$

Can we generalize the inequality?  The hint for how to do so is that the number of subsystems ($100$) is no more than the overall failure rate divided by the subsystem rate:

$\displaystyle \frac{10^{-7}}{10^{-9}}$

This suggests the general form is something like

Subsystem reliability inequality: $\displaystyle (1 - C^{-n})^{C^{n-m}} \geq 1 - C^{-m}$

where $C,$ $n,$ and $m$ are real numbers, $C \geq 1,$ $n \geq 0,$ and $n \geq m.$

Let’s prove the inequality holds.  Joe Hurd figured out the proof, sketched below (but I take responsibility for any mistakes in it’s presentation).  For convenience, we’ll prove the inequality holds specifically when $C = e,$ but the proof can be generalized.

First, if $n = 0,$ the inequality holds immediately. Next, we’ll show that

$\displaystyle (1 - e^{-n})^{e^{n-m}}$

is monotonically non-decreasing with respect to $n$ by showing that the derivative of its logarithm is greater or equal to zero for all $n > 0.$  So the derivative of its logarithm is

$\displaystyle \frac{d}{dn} \; e^{n-m}\ln(1-e^{-n}) = e^{n-m}\ln(1-e^{-n})+\frac{e^{-m}}{1-e^{-n}}$

We show

$\displaystyle e^{n-m}\ln(1-e{-n})+\frac{e^{-m}}{1-e^{-n}} \geq 0$

iff

$\displaystyle e^{-m}\left(e^{n}\ln(1-e^{-n}) + \frac{1}{1-e^{-n}}\right) \geq 0$

and since $e^{-m} \geq 0,$

$\displaystyle e^{n}\ln(1-e^{-n}) + \frac{1}{1-e^{-n}} \geq 0$

iff

$\displaystyle e^{n}\ln(1-e^{-n}) \geq - \frac{1}{1-e^{-n}}$

Let $x = e^{-n}$, so the range of $x$ is $0 < x < 1.$
$\displaystyle\ln(1-x) \geq - \frac{x}{1-x}$

Now we show that in the range of $x$, the left-hand side is bounded below by the right-hand side of the inequality.
$\displaystyle \lim_{x \to 0} \; \ln(1-x) = 0$

and
$\displaystyle - \frac{x}{1-x} = 0$

Now taking their derivatives
$\displaystyle \frac{d}{dx} \; \ln(1-x) = \frac{1}{x-1}$

and
$\displaystyle \frac{d}{dx} \; - \frac{x}{1-x} = - \frac{1}{(x-1)^2}$

Because $\displaystyle x-1 \geq - (x-1)^2$ in the range of $x$, our proof holds.

The purpose of this post was to clarify the folklore of ultra-reliable systems.  The subsystem reliability inequality presented allows for easy generalization to other reliable systems.

Thanks again for the help, Joe! Read the rest of this entry »

## Writer’s Unblock

September 30, 2009

I’ve recently got a few technical papers out the door involving Haskell, physical-layer protocols, SMT, security modeling, and run-time verification of embedded systems (phew!).  One of the benefits of industrial research is getting your hands involved in a lot of different research projects.

• This paper is about using Haskell to model physical-layer protocols and using QuickCheck to test them.  Physical-layer protocols are used to transmit bits from one clock-domain to another and are used in ethernet, credit card swipers, CD players, and so on.  The gist of the paper is that even though Haskell is pure & lazy, it works great for modeling and testing real-time protocols and even for computing reliability statistics.  I presented it at the Haskell Symposium in September ’09, which was associated with ICFP.  (The talk video is online!)  The paper is a short experience report—indeed, it is the only experience report that was accepted at the symposium.  The Haskell Symposium was an entertaining and friendly environment for presenting.
• This paper actually precedes the Haskell paper, but it extends the results by describing how to formally verify physical-layer protocols using SMT solvers and k-induction (we use SRI’s SAL tool in this work).  The paper is a journal article accepted at Formal Aspects of Computing.  You’ll find at least two things interesting about this article: (1) For all the excitement about SMT, there don’t seem to be a lot of great examples demonstrating its efficacy—the problems described in this paper were (laboriously!) verified using theorem-provers by others previously, and our approach using SMT is much more automated.  (2) We provide a nice general model of cross clock-domain circuits and particularly metastability.

So if you can verify physical-layer protocols, why model them in Haskell and QuickCheck them (as we did above)?  There are at least two reasons.  First, if you’re using SMT, then your timing constraints need to be linear inequalities to be decidable.  For systems that with nonlinear constraints, QuickCheck might be your only recourse.  Second, QuickCheck gives you concrete counterexamples and test-cases that you can use to test implementations (SMT solvers often return symbolic counterexamples).

• This paper describes a simple model for analyzing information flow in a system (where a “system” could be a program, a network, an OS, etc.).  The main portion of the paper describes heuristics based on graph algorithms for deciding what sort of information flow policies you might want to enforce in your system.  In general, there’s been a lot of work on analyzing access control policies but not so much work in figuring out what kind of policy you should have in the first place (if you know of such work, please tell me!).  The paper isn’t deep, and it’s also preliminary insofar as I don’t describe building a complex system using the techniques.  Still, there’s a small (Haskell) script available that implements the algorithms described; I’d love to see these analyses find their way into a tool to help system designers build secure systems.
• Finally, this report describes the field of run-time monitoring (or run-time verification) as it applies to safety-critical real-time embedded software.  Run-time monitoring compliments formal verification since when a system is too complicated to verify a priori, it can be monitored at run-time to ensure it conforms to its specification.  Not a lot of work has been done on monitoring software that’s hard real-time, distributed, or fault-tolerant—which ironically could benefit the most from run-time monitoring.  The report should serve as a nice, gentle introduction.  The report should be published soon as a NASA Contractor Report—the work was done under a NASA-sponsored project for which I’m the PI.

Don’t hesitate to give me feedback on any of these papers.  Ok, time to fill up the queue again…

## Finding Boole

August 10, 2009

Here’s a simple challenge-problem for model-checking Boolean functions: Suppose you want to compute some Boolean function $spec :: B^k \rightarrow B$, where $B^k$ represents 0 or more Boolean arguments.

Let $f_0$, $f_1$, $\ldots$, $f_j$ range over 2-ary Boolean functions, (of type $(Bool, Bool) \rightarrow Bool$), and suppose that $f$ is a fixed composition of $f_0$, $f_1$, $\ldots$, $f_j$. (By the way, I’m going to talk about functions, but you can think of these as combinatorial circuits, if you prefer.)

Our question is, “Do there exist instantiations of $f_0$, $f_1$, $\ldots$, $f_j$ such that for all inputs, $f = spec$?

What is interesting to me is that our question is quantified and of the form, “exists a forall b …”, and it is “higher-order” insofar as we want to find whether there exist satisfying functions. That said, the property is easy to encode as a model-checking problem. Here, I’ll encode it into SRI’s Symbolic Analysis Laboratory (SAL) using its BDD engine.  (The SAL file in its entirety is here.)

To code the problem in SAL, we’ll first define for convenience a shorthand for the built-in type, BOOLEAN:

B: TYPE = BOOLEAN;

And we’ll define an enumerated data type representing the 16 possible 2-ary Boolean functions:

B2ARY: TYPE = { False, Nor, NorNot, NotA, AndNot, NotB, Xor, Nand
, And, Eqv, B, NandNot, A, OrNot, Or, True};

Now we need an application function that takes an element f from B2ARY and two Boolean arguments, and depending on f, applies the appropriate 2-ary Boolean function:

app(f: B2ARY, a: B, b: B): B =
IF    f = False   THEN FALSE
ELSIF f = Nor     THEN NOT (a OR b)
ELSIF f = NorNot  THEN NOT a AND b
ELSIF f = NotA    THEN NOT a
ELSIF f = AndNot  THEN a AND NOT b
ELSIF f = NotB    THEN NOT b
ELSIF f = Xor     THEN a XOR b
ELSIF f = Nand    THEN NOT (a AND b)
ELSIF f = And     THEN a AND b
ELSIF f = Eqv     THEN NOT (a XOR b)
ELSIF f = B       THEN b
ELSIF f = NandNot THEN NOT a OR b
ELSIF f = A       THEN a
ELSIF f = OrNot   THEN a OR NOT b
ELSIF f = Or      THEN a OR b
ELSE                   TRUE
ENDIF;

Let’s give a concrete definition to f and say that it is the composition of five 2-ary Boolean functions, f0 through f4. In the language of SAL:

f(b0: B, b1: B, b2: B, b3: B, b4: B, b5: B):
[[B2ARY, B2ARY, B2ARY, B2ARY, B2ARY] -> B] =
LAMBDA (f0: B2ARY, f1: B2ARY, f2: B2ARY, f3: B2ARY, f4: B2ARY):
app(f0, app(f1, app(f2, b0,
app(f3, app(f4, b1, b2),
b3)),
b4),
b5);

Now let’s define the spec function that f should implement:

spec(b0: B, b1: B, b2: B, b3: B, b4: B, b5: B): B =
(b0 AND b1) OR (b2 AND b3) OR (b4 AND b5);

Now, we’ll define a module m; modules are SAL’s building blocks for defining state machines. However, in our case, we won’t define an actual state machine since we’re only modeling function composition (or combinatorial circuits). The module has variables corresponding the function inputs, function identifiers, and a Boolean stating whether f is equivalent to its specification (we’ll label the classes of variables INPUT, LOCAL, and OUTPUT, to distinguish them, but for our purposes, the distinction doesn’t matter).

m: MODULE =
BEGIN
INPUT b0, b1, b2, b3, b4, b5 : B
LOCAL f0, f1, f2, f3, f4 : B2ARY
OUTPUT equiv : B

DEFINITION
equiv = FORALL (b0: B, b1: B, b2: B, b3: B, b4: B, b5: B):
spec(b0, b1, b2, b3, b4, b5)
= f(b0, b1, b2, b3, b4, b5)(f0, f1, f2, f3, f4);
END;

Notice we’ve universally quantified the free variables in spec and the definition of f.

Finally, all we have to do is state the following theorem:

instance : THEOREM m |- NOT equiv;

Asking whether equiv is false in module m. Issuing

$sal-smc FindingBoole.sal instance asks SAL’s BDD-based model-checker to solve theorem instance. In a couple of seconds, SAL says the theorem is proved. So spec can’t be implemented by f, for any instantiation of f0 through f4! OK, what about spec(b0: B, b1: B, b2: B, b3: B, b4: B, b5: B): B = TRUE; Issuing $ sal-smc FindingBoole.sal instance

we get a counterexample this time:

f0 = True
f1 = NandNot
f2 = NorNot
f3 = And
f4 = Xor

which is an assignment to the function symbols. Obviously, to compute the constant TRUE, only the outermost function, f0, matters, and as we see, it is defined to be TRUE.

By the way, the purpose of defining the enumerated type B2ARY should be clear now—if we hadn’t, SAL would just return a mess in which the value of each function f0 through f4 is enumerated:

f0(false, false) = true
f0(true, false) = true
f0(false, true) = true
f0(true, true) = true
f1(false, false) = true
f1(true, false) = true
f1(false, true) = false
f1(true, true) = true
f2(false, false) = false
f2(true, false) = true
f2(false, true) = false
f2(true, true) = false
f3(false, false) = false
f3(true, false) = false
f3(false, true) = false
f3(true, true) = true
f4(false, false) = false
f4(true, false) = true
f4(false, true) = true
f4(true, true) = false

OK, let’s conclude with one more spec:

spec(b0: B, b1: B, b2: B, b3: B, b4: B, b5: B): B =
(NOT (b0 AND ((b1 OR b2) XOR b3)) AND b4) XOR b5;

This is implementable by f, and SAL returns

f0 = Eqv
f1 = OrNot
f2 = And
f3 = Eqv
f4 = Nor

Although these assignments compute the same function, they differ from those in our specification. Just to double-check, we can ask SAL if they’re equivalent:

spec1(b0: B, b1: B, b2: B, b3: B, b4: B, b5: B): B =
((b0 AND ((NOT (b1 OR b2))  b3)) OR NOT b4)  b5;

specifies the assignments returned, and

eq: THEOREM m |- spec(b0, b1, b2, b3, b4, b5) =  spec1(b0, b1, b2, b3, b4, b5);

asks if the two specifications are equivalent. They are.

## “Schrodinger’s Probability” for Error-Checking Codes

May 15, 2009

In a previous post, I discussed the notion of Schrödinger CRCs, first described by Kevin Driscoll et al. in their paper Byzantine Fault Tolerance, from Theory to Reality. The basic idea is that error-detecting codes do not necessarily prevent two receivers from obtaining messages that are semantically different (i.e., different data) but syntactically valid (i.e., the CRC matches the respective data words received). The upshot is that even with CRCs, you can suffer Byzantine faults, with some probability.

… So what is that probability of a Schrödinger’s CRC? That’s the topic of this post—which cleans up a few of the ideas I presented earlier. I published a short paper on the topic, which I presented at Dependable Sensors and Networks, 2010, while Kevin Driscoll was in the audience!  If you’d prefer to read the PDF or get the slides, they’re here.  The simulation code (Haskell) is here.

View this document on Scribd

## An Atomic Fibonacci Server: Exploring the Atom (Haskell) DSL

May 5, 2009

This post is consistent with Atom 0.0.1 and not the latest version, Atom 0.0.5 (the author went off and implemented changes I and others suggested :)).  I’ll update the post… soon.

Tom Hawkins has open-sourced Atom, a domain-specific language (DSL) for writing embedded real-time software. Atom is actually an “embedded DSL” (I prefer the term “lightweight DSL”) in the functional language Haskell. It’s a lightweight DSL (LwDSL) because you write legal Haskell and let the Haskell compiler do all the heavy lifting. The DSL is a set of special functions and data types and a “compile function” that generates embedded (i.e., no dynamic memory) C code.  You don’t have to write your own compiler from scratch.

John Van Enk has already posted a couple of blog entries on using Atom; first on adding slightly to the LwDSL (one major advantage of a LwDSL is that it’s easy to extend the language—you don’t have to re-engineer a standalone compiler) and then on using Atom to blink some LEDs on the Arduino.  Keep checking his blog for more updates!

Here, I write a little device and driver program in Atom: the driver sends an index i, and the device returns the ith Fibonacci number. The little bit of challenge in doing this is that the device and driver may run at different rates, so their communication is asynchronous.  How does this work in a language like Atom?

## Writing in the Atom DSL

Let’s think about the Fibonacci device (we’ll call it fibDev) first.  The device fibDev will do three things:

1. Wait for a new index i from the driver.
2. Produce a result, fib(i).
3. Give the result to the driver.

Let’s think about step (2) first.  Think for a second how we’d write this (efficiently) in Haskell:

fib :: Int -> Int
fib n = fst $fibHlp n where fibHlp n = case n of 0 -> (1, 1) _ -> let (a,b) = fibHlp (n-1) in (b,a+b) The Atom implementation will use the same algorithm, but it’ll look different. Atom is a synchronous language, so you specify rules that fire on clock ticks. Here’s what the core of the algorithm looks like in Atom (I haven’t shown the variable declarations, but look you can look at the full source): atom "computeFib"$ do
cond $value runFib cond$ value i >. 0
decr i
snd <== (value fst) + (value snd)
fst <== value snd

Atom is written in a monadic style.  Here, we have two conditions, both of which must be true for the rule to “fire”.  The first condition is that runFib is true (telling the device it’s in its computation step), and the second condition is that the index is greater than 0 (we stop computing at zero).  If the conditions are true, then the value of i is decremented, and we update the values of the fst and snd variables, corresponding the first and second elements, respectively, of the pair in the Haskell specification. Again, this is legal Haskell; the Atom library defines the special operators (e.g., >.).  One great thing about writing embedded code in Atom is that variable updates are synchronous.  For example, in the code above, fst is updated to the previous value of snd. That’s the core of the Fibonacci device.

The rest of the architecture handles the message passing (in the C code we’ll generate, messages are passed via global variables) and synchronization between the driver and device, as summarized below:

System Architecture

We do not assume that fibDvr and fibDev execute at the same rate, so we handle message passing with a series of handshakes.  First, fibDvr sends a new value x and notifies fibDev that the value is ready (by issuing newInd).  fibDev acknowledges that x has been received with valRcvd.  At this point, fibDvr knows to wait for fibDev to compute fib(x).  Once it receives the notice ansReady, it reads off the answer, ans.

All we have to do now is implement the handshakes.  For example, let’s look at step (3) of the device, sending the final answer to the driver.  It’s behavior should be clear from the architectural description.

atom "sendVal" $do cond$ value i ==. 0
cond $value runFib runFib <== false ans <== value fst ansReady <== true valRcvd <== false And here’s step (1) for fibDev, waiting for a new index from the driver: atom "getIndex"$ do
cond $not_ (value runFib) cond$ value newInd
i        <== value x
runFib   <== true
fst      <== 1
snd      <== 1
ansReady <== false
valRcvd  <== true

These three rules for fibDev define the body of fibDev‘s “do” block.

fibDev :: Atom ()
fibDev = period 3 $do ... We tell atom that the period is 3, meaning execute each of our three rules every three clock ticks (based on the underlying clock). Now that we’re comfortable with the language, let’s look at the entire definition of fibDvr in one go. Recall the job of fibDvr is to send a value then wait for an answer. Our driver will increment values by 5, starting at 0. It’ll stop sending new values if the index is bigger than 50. fibDvr :: Atom () fibDvr = period 20$ do
x        <- word64 "x" 0 -- new index to send
oldInd   <- word64 "oldInd" 0 -- previous index sent
-- external signals --
valRcvd  <- bool' "valRcvd" -- has the device received the new index?
ans      <- word64' "ans" -- the newly-computed fib(x)
ansReady <- bool' "ansReady" -- is an answer waiting?
----------------------
valD     <- word64 "valD" 1 -- local copy of fib(x)
newInd   <- bool "newInd" True -- a new index is ready
waiting  <- bool "waiting" True -- waiting for a new computation

atom "wait" $do cond$ value valRcvd
cond $not_$ value waiting
newInd  <== false
waiting <== true

atom "getAns" $do cond$ value ansReady
cond $value waiting cond$ value x <. 50
valD    <== value ans
x       <== value x + 5
waiting <== false
newInd  <== true
oldInd  <== value x

Note that we’ve specified the period of the driver to be 20, meaning that its two rules get executed every 20 ticks.  So the driver is much slower than the device, but if our handshakes are correct, the two devices communicate correctly for any rates of execution of the two components.  (Proving it for all-time is a classic model checking problem.)

## Compiling to C

We include a little Haskell function that we can call to “compile” fibDev and fibDvr into embedded C files.  (The compile function is part of Atom, and it takes a name for the generated C file and Atom specifications to compile.)

compileFib :: IO ()
compileFib = do
compile "fibDev" $fibDev compile "fibDvr"$ fibDvr

We can call this from an interpreter for Haskell; it takes about a second to compile. Doing so almost produces the source files fibDvr.c and fibDev.c. We do a few things manually:

• Write two header files, fibDvr.h and fibDev.h and import them. This is the code we want to talk to each other through global variables.  We’ll also include stdio.h so we can printf our results.
• Because Atom automatically (atomatically? :)) generates variable and function names in the generated code, we declare some of the identifiers in fibDev.c to be static so they aren’t globally visible.
• We #define the variable names from the Atom-generated identifiers back to the expected identifiers for the variables that are shared.
• And we add a little main function to execute the code: let’s execute the driver and device for 500 clock ticks:
int main() {
while(__clock < 500) {
fibDvr();
fibDev();
}
return 0;
}

Of course, Atom could be extended to handle these things itself—John Van Enk has already started doing some of it.  In all, our 80-some lines of Atom compile to over 200 lines of embedded C.  So let’s test it!

> gcc -Wall -o fibDvr fibDev.c fibDvr.c > ./fibDvr

generates the following output:

i: 0, fib(i): 1
i: 0, fib(i): 1
i: 0, fib(i): 1
i: 5, fib(i): 8
i: 5, fib(i): 8
i: 10, fib(i): 89
i: 10, fib(i): 89
i: 10, fib(i): 89
i: 15, fib(i): 987
i: 15, fib(i): 987
i: 15, fib(i): 987
i: 15, fib(i): 987
i: 20, fib(i): 10946
i: 20, fib(i): 10946
i: 20, fib(i): 10946
i: 20, fib(i): 10946
i: 25, fib(i): 121393
i: 25, fib(i): 121393
i: 25, fib(i): 121393
i: 25, fib(i): 121393
i: 25, fib(i): 121393
i: 30, fib(i): 1346269
i: 30, fib(i): 1346269
i: 30, fib(i): 1346269
i: 30, fib(i): 1346269

Wait, why are we getting the same answers multiple times? Recall that Atom is a synchronous language, so functions are executed based on time (measured in underlying clock ticks), not events. But most times, the guards don’t hold, so state isn’t updated. That’s what we see here.

Oh, we should check our specification. We can do that using our original Haskell specification:

> map fib [0,5..30]
[1,8,89,987,10946,121393,1346269]

Looks good!

Let me know if this helps you understand Atom, or if you have thoughts on how Atom compares to other languages.

Finally, here are the sources:

## N-Version Programming… For the nth Time

April 27, 2009

Software contains faults.  The question is how to cost-effectively reduce the number of faults.  One approach that gained traction and then fell out of favor was N-version programming.  The basic idea is simple: have developer teams implement a specification independent from one another.  Then we can execute the programs concurrently and compare their results.  If we have, say, three separate programs, we vote their results, and if one result disagrees with the others, we presume that program contained a software bug.

N-version programming rests on the assumption that software bugs in independently-implemented programs are random, statistically-uncorrelated events.  Otherwise, multiple versions are not effective at detecting errors if the different versions are likely to suffer the same errors.

John Knight and Nancy Leveson famously debunked this assumption on which N-version programming rested in the “Knight-Leveson experiment” they published in 1986.  In 1990, Knight and Leveson published a brief summary of the original experiment, as well as responses to subsequent criticisms made about it, in their paper, A Reply to the Criticisms of the Knight & Leveson Experiment.

The problem with N-version programming is subtle: it’s not that it provides zero improvement in reliability but that it provides significantly less improvement than is needed to make it cost-effective compared to other kinds of fault-tolerance (like architecture-level fault-tolerance).  The problem is that even small probabilities of correlated faults lead to significant reductions in potential reliability improvements.

Lui Sha has a more recent (2001) IEEE Software article discussing N-version programming, taking into account that the software development cycle is finite: is it better to spend all your time and money on one reliable implementation or on three implementations that’ll be voted at runtime?  His answer is almost always the former (even if we assume uncorrelated faults!).

But rather than N-versions of the same program, what about different programs compared at runtime?  That’s the basic idea of runtime monitoring.  In runtime monitoring, one program is the implementation and another is the specification; the implementation is checked against the specification at runtime.  This is easier than checking before runtime (in which case you’d have to mathematically prove every possible execution satisfies the specification).  As Sha points out in his article, the specification can be slow and simple.  He gives the example of using the very simple Bubblesort as the runtime specification of the more complex Quicksort: if the Quicksort does its job correctly (in O(n log n), assuming a good pivot element), then checking its output (i.e., a hopefully properly sorted list) with Bubblesort will only take linear time (despite Bubble sort taking O(n2) in general).

The simple idea of simple monitors fascinates me.  Of course, Bubblesort is not a full specification, though.  Although Sha doesn’t suggest it, we’d probably like our monitor to compare the lengths of the input and output lists to ensure that the Quicksort implementation didn’t remove elements.  And there’s still the possibility that the Quicksort implementation modifies elements, which is also unchecked by a Bubblesort monitor.

But instead of just checking the output, we could sort the same input with both Quickcheck and Bubblesort and compare the results.  This is a “stronger” check insofar as different sorts would have to have exactly the same faults (e.g., not sorting, removing elements, changing elements) for an error not to be caught.  The principal drawback is the latency of the slower Bubblesort check as compared to Quicksort.  But sometimes, it may be ok to signal an error (shortly) after a result is provided.

Just like for N-version programming, we would like the faults in our monitor to be statistically uncorrelated with those in the monitored software.  I am left wondering about the following questions:

• Is there research comparing the kinds of programming errors made in radically different paradigms, such as a Haskell and C?  Are there any faults we can claim are statistically uncorrelated?
• Runtime monitoring itself is predicated on the belief that the implementations of different programs will fail in statistically independent ways, just like N-version programming is.  While more plausible, does this assumption hold?

## Programming Languages for Unpiloted Air Vehicles

April 20, 2009

I recently presented a position paper at a workshop addressing software challenges for unpiloted air vehicles (UAVs).  The paper is coauthored with Don Stewart and John Van Enk.  From a software perspective, UAVs (and aircraft, in general) are fascinating.  Modern aircraft are essentially computers that fly, with a pilot for backup.  UAVs are computers that fly, without the backup.

Some day in the not-so-distant future, we may have UPS and FedEx cargo planes that are completely autonomous (it’ll be a while before people are comfortable with a pilot-less airplane).  These planes will be in the commercial airspace and landing at commercial airports.  Ultimately, UAVs must be transparent: an observer should not be able to discern whether an airplane is human or computer controlled by its behavior.

You won’t be surprised to know a lot of software is required to make this happen.  To put things in perspective, Boeing’s 777 is said to contain over 2 million lines of newly-developed code; the Joint Strike Fighter aircraft is said to have over 5 million lines.  Next-generation UAVs, with pilot AI, UAV-to-UAV and UAV-to-ground communications, and arbitrary other functionality, will have millions more lines of code. And the software must be correct.

In our paper, we argue that the only way to get a hold on the complexity of UAV software is through the use of domain-specific languages (DSLs). A good DSL relieves the application programmer from carrying out boilerplate activities and providers her with specific tools for her domain. We go on and advocate the need for lightweight DSLs (LwDSLs), also known as embedded DSLs. A LwDSL is one that is hosted in a mature, high-level language (like Haskell); it can be thought of as domain-specific libraries and domain-specific syntax.  The big benefit of a LwDSL is that a new compiler and tools don’t need to be reinvented. Indeed, as we report in the paper, companies realizing the power of LwDSLs are quietly gaining a competitive advantage.

Safety-critical systems, like those on UAVs, combine multiple software subsystems.  If each subsystem is programmed in its own LwDSL hosted in the same language, then it is easy to compose testing and validation across subsystem boundaries. (Only testing within each design-domain just won’t fly, pun intended.)

The “LwDSL approach” won’t magically make the problems of verifying life-critical software, but “raising the abstraction level” must be our mantra moving forward.

## Byzantine Cyclic Redundancy Checks (CRC)

April 18, 2009

I’m working on a NASA-sponsored project to monitor safety-critical embedded systems at runtime, and that’s started me thinking about cyclic redundancy checks (CRCs) again.  Error-checking codes are fundamental in fault-tolerant systems.  The basic idea is simple: a transmitter wants to send a data word, so it computes a CRC over the word.  It sends both the data word and the CRC to the receiver, which computes the same CRC over the received word.  If its computed CRC and the received one differ, then there was a transmission error (there are simplifications to this approach, but that’s the basic idea).

CRCs have been around since the 60s, and despite the simple mathematical theory on which they’re built (polynomial division in the Galois Field 2, containing two elements, “0″ and “1″), I was surprised to see that even today, their fault-tolerance properties are in some cases unknown or misunderstood.  Phil Koopman at CMU has written a few nice papers over the past couple of years explaining some common misconceptions and analyzing commonly-used CRCs.

Particularly, there seems to be an over-confidence in their ability to detect errors.  One fascinating result is the so-called “Schrödinger’s CRC,” so-dubbed in a paper entitled Byzantine Fault Tolerance, from Theory to Reality, by Kevin Driscoll et al. A Schrödinger’s CRC occurs when a transmitter broadcasts a data word and associated CRC to two receivers.   and at least one of the data words is corrupted in transit and so is the corresponding CRC so that the faulty word and faulty CRC match!  How does this happen?  Let’s look at a concrete example:

             11-Bit Message           USB-5
Receiver A   1 0 1 1 0 1 1 0 0 1 1    0 1 0 0 1
Transmitter  1 ½ 1 1 0 1 1 ½ 0 1 1    ½ 1 0 0 1
Receiver B   1 1 1 1 0 1 1 1 0 1 1    1 1 0 0 1

We illustrate a transmitter broadcasting an 11-bit message to two receivers, A and B. We use USB-5 CRC, generally used to check USB token packets  (by the way, for 11-bit messages, USB-5 has a Hamming Distance of three, meaning the CRC will catch any corruption of fewer than three bits in the combined 11-bit message and CRC).  Now, suppose the transmitter has suffered some fault such as a “stuck-at-1/2” fault so that periodically, the transmitter fails to drive the signal on the bus sufficiently high or low. A receiver may interpret an intermediate signal as either a 0 or 1. In the ﬁgure, we show the transmitter sending three stuck-at-1/2 signals, one in the 11-bit message, and two in CRC.  The upshot is an example in which a CRC does not prevent a Byzantine fault—the two receivers obtain different messages, each of which passes its CRC.

One question is how likely this scenario is.  Paulitsch et al. write that The probability of a Schrödinger’s CRC is hard to evaluate.  A worst-case estimate of its occurrence due to a single device is the device failure rate.”  It’d be interesting to know if there’s any data on this probability.