By Jack Ganssle
For novel ideas about building embedded systems (both hardware and firmware), join the 40,000+ engineers who subscribe to The Embedded Muse, a free biweekly newsletter. The Muse has no hype and no vendor PR. Click here to subscribe.
By Jack Ganssle
A Boss's Quick Start to Firmware Engineering, Part 2Published in ESD August, 2004
(Find part 1 of this article here.)
Firmware engineers often tell me they want to do the right things, to employ careful design practices and the like, but are defeated by their bosses' ignorance of software issues. So cut out last month's and this article and slip it under your boss's door.
A poll on embedded.com (http://embedded.com/pollArchive/?surveyno=12900001) suggests 85% of companies won't spend more than $1k on any but the most essential tools. Considering the $100k+ loaded cost of a single engineer, it's nuts to not spend a few grand on a tool that offers even a small productivity boost.
Like what? Lint, for one. Lint is a program that examines the source code and identifies suspicious areas. It's like the compiler's syntax checker, but one on steroids. Only a Lint is smart enough to watch variable and function usage across multiple files. Compilers can't do that. Aggressive Lint usage picks out many problems before debugging starts, for a fraction of the cost. Lint all source files before doing code inspections.
Gimpel (www.gimpel.com) sells one for $239. It's up to you to buy it, and to ensure your engineers use it on all new code. Lint is annoying at first, often initially zeroing in on constructs that are indeed fine. Don't let that quirk turn your people off. Tame it, and then reap great reductions in debugging times.
Debugging eats 50% of most projects' schedules. The average developer has a 5 to 10% error rate. Anything that trims that even a smidgen saves big bucks.
Make sure the developers aren't cheating their tools. Warning levels on compilers, for instance, should be set to the lowest possible level so all warnings are displayed. And then insist the team writes warning-free code. It's astonishing how we ship firmware that spews warnings when compiled. The compiler, which understands the language's syntax far better than any of your people, is in effect shouting "Look here. Here! This is scary!" How can anyone ignore such a compelling danger sign?
Write warning-free code so that maintenance people in months or decades won't be baffled by the messages. "Is it supposed to do this? Or did I reinstall the compiler incorrectly? Which of these is important?" This means changing the way they write C. Use explicit casting. Parenthesis when there's any doubt. These are all good programming practices anyway, with zero cost in engineering, execution speed, or code size. What's the downside?
Editors, compilers, linkers, and debuggers are essential and non-negotiable tools as it's impossible to do any development without these. Consider others. Complexity analyzers can yield tremendous insight into functions, identifying "bad code" early, before the team wastes their time and spirits trying to beat the cruddy code into submission. Bug tracking software helps identify problem areas - see a list of resources at http://www.aptest.com/resources.html.
Most firmware developers are desperate for better debugging tools. Unhappily, the grand old days of in-circuit emulators are over. These tools provided deep insight into the intrinsically hard-to-probe embedded system. Their replacement, the BDM, offers far less capability. Have mercy on your folks and insist the hardware team dedicate a couple of spare parallel output bits just to the software people. They'll use these along with instrumented code for a myriad of debugging tasks, especially for hard-to-measure performance issues.
Your developers - not tools, not widgets, not components - are your prime resource. As one wag noted, "my inventory walks out the door each night."
I've recommended several books in these two articles. Please, though, read Peopleware by DeMarco and Lister (ISBN 0932633439, 1999 Dorset House Publishing, NY NY). It's a slender volume that you'll plow through in just a couple of enjoyable hours. Pursuing the elusive underpinnings of software productivity, for 10 years the authors conducted a "coding war" between volunteering companies.
The results? Well, at first the data was a scrambled mess. Nothing correlated. Teams that excelled on the projects (by any measure: speed, bug count, matching specs) were neither more highly paid nor more experienced than the losers. Crunching every parameter revealed the answer: developers imprisoned in noisy cubicles, those who had no defense against frequent interruptions, did poorly.
How poorly? The numbers are breathtaking. The best quartile was 300% more productive than the lowest 25%. Yet privacy was the only difference between the groups.
Think about it - would you like 3x faster development?
It takes your developers 15 minutes, on average, to move from active perception of the office busyness to being totally and productively engaged in the cyberworld of coding. Yet a mere 11 minutes passes between interruptions for the average developer. Ever wonder why firmware costs so much? Email, the phone, people looking for coffee filters and sometimes you, boss, all clamor for attention
Sadly, most developers live in cubicles today, which are, as Dilbert so astutely noted, "anti-productivity pods". Next time you hire someone peer into his cube occasionally. At first he's anxious to work hard, focus, and crank out a great product. He'll try to tune out the poor sod in the next cube who's jabbering on the phone with his lawyer about the divorce. But we're all human; after a week or so he's leaning back from the keyboard, ears raised to get the latest developments. A productive environment? Nope.
I advise you to put your developers in private offices, with doors and off-switches on the phones. You probably won't do that. Every interior designers promise cubes offer more "flexibility". But even cubicles have options.
Encourage your people to identify their most productive hours, that time of day when their brains are engaged and working at max efficiency. Me, I'm a morning person. Others have different habits. But find those productive hours and help them shield themselves from interruptions for about three hours a day. In that short time, with the 3x productivity boost, they'll get an entire day's work done. The other five hours can be used for meetings, email, phone contacts, supporting other projects, etc.
Give your folks a curtain to pull across the cube's opening. Obviously a curtain rod would decapitate employees, generally a bad idea despite the legions of unemployed engineers clamoring for work. Use a Velcro strip to secure the curtain in place. Put a sign on the curtain labeled "enter and die"; the sign and curtain go up during the employee's 3 superprogramming hours per day. Train the team to respect their colleagues' privacy during these quiet hours. At first they'll be frantic: "but I've GOT to know the input parameters to this function or I'm stuck!" With time they'll learn when Joe, Mary or Bob will be busy and plan ahead. Similarly, if www.chris-lott.org/resources/cmetrics/ normal">you really need a project update and Shirley has her curtain up, back slowly and quietly away. Wait till their hours of silence are over.
Have them turn off their phone during this time. If Mary's spouse needs her to pick up milk on the way home, well, that's perfect voicemail fodder. If the kids are in the hospital, then the phone attendant can break in on her quiet time.
The study took place before email was common. You know, that cute little bleep that alerts you to the same tired old joke that's been circulating around the 'net for the last three months! while diverting attention from the problem at hand. Every few seconds, it seems. Tell your people to disable email while cloistered.
When I talk to developers about the interruption curse they complain that the boss is the worst offender. Resist the temptation to interrupt. Remember just how productive that person is at the moment, and wait till the curtain comes down.
(If you're afraid the employee is hiding behind the curtain surfing the net or playing Doom, well, there are far more severe problems than just productivity issues. Without trust - mutual trust - any engineering department is in trouble).
Where should you use your best people? It's natural to put the superprogrammers on the biggest and most complex projects. Resist that urge - it's wrong.
Capers Jones showed that the best people excel on small (one man-month) projects, typically being 6 times more productive than the worst members of the team. That advantage diminishes as the system grows. On an 8 man-month effort the ratio shrinks to under 3 to 1. At 64 man-months it's about 1.5 to 1, and much beyond that the best do as badly as the worst. Or the worst as well as the best. Whatever.
That observation tells us something important about how we partition big projects. Find ways to break big systems down into many small, mostly independent parts. Or at least strip out as much as possible from the huge carcass of code you're planning to generate, putting the removed sections into their own tasks or even separate processors. Give these smaller sections to the superprogrammers. They'll crank out solutions fast.
An example: suppose an I/O device, say an optical encoder, is tied to your system. Remove it. Add a CPU, a cheap PIC, ATMEL, Z8 or similar sub-$1 part, just to manage that one device. Have it return its data in engineering units: "the shaft angle is 27 degrees". Even a slowly rotating encoder would generate thousands of interrupts a second, a burden to even the fastest CPU that's also tasked with a many other activities. Yet even a tiny microcontroller can easily handle the data if there's nothing else going on. One smart developer can crank out perfect I/O code in little time.
(An important rule of thumb states that 90% loaded systems double development time, compared to one of 70% or less; 95% loading triples development time.)
While cleverly partitioning the project for the sake of accelerating the development schedule, think like the customer does, not as the firmware folks do. The customer only sees features; never objects, ISRs or functions. Features are what sell the product.
That means break the development effort down into feature-chunks. The first feature of all, of course, is a simple skeleton that sets up the peripherals and gets to main(). That and a few critical ISRs, perhaps an RTOS and the like form the backbone upon which everything else is built.
Beyond the backbone are the things the customer will see. In a digital camera there's a handler for the CCD, an LCD subsystem, some sort of Flash filesystem. Cool tricks like image enhancement, digital zoom, and much more will be the sizzle that excites marketing. None of those, of course, has much to do with the basic camera functionality.
Create a list of the features and prioritize. What's most important? Least? Then! and this is the trick! implement the most important features first.
Does that sound trite? It is, yet every time I look at a product in trouble no one has taken this step. Developers have virtually every feature half-implemented. The ship date arrives and nothing works. Worse, there's no clear recovery strategy since so much effort has been expended on things that are not terribly important.
So in a panic management starts tossing out features. One 2002 study showed that 74% of projects wind up with 30% or more of the features being eliminated. Not only is that a terrible waste - these are partially implemented features - but the product goes to market late, with a subset of its functionality. If the system were built as I'm recommending, even schedule slippages would, at worst, result in scrubbing a few requirements that had as yet not consumed engineering time. Failure, sure, but failure in a rather successful way.
Finally, did you know great code, the really good stuff, that which has the highest reliability, costs the same as cruddy software? This goes against common sense. Of course, all things being equal, highly safety critical code is much more expensive that consumer-quality junk.
But what if we don't hold all things equal? O. Benediktsson (Safety Critical Software and Development Productivity, conference proceedings, Second World Conference on Software Quality, Sept 2000) showed that using higher and higher levels of disciplined software process lets one build higher-rel software at a constant cost. If your projects march from low reliability along an upwards line to truly safety-critical code, and if your outfit follows, in his study, increasing levels of the Capability Maturity Model, the cost remains constant.
Makes one think. And hopefully, it makes one reign in the hackers who are more focused on cranking code than specifying, designing, and carefully implementing a world-class product.