Slashdot Mirror
Ultra-Stable Software Design in C++?
null_functor asks: "I need to create an ultra-stable, crash-free application in C++. Sadly, the programming language cannot be changed due to reasons of efficiency and availability of core libraries. The application can be naturally divided into several modules, such as GUI, core data structures, a persistent object storage mechanism, a distributed communication module and several core algorithms. Basically, it allows users to crunch a god-awful amount of data over several computing nodes. The application is meant to primarily run on Linux, but should be portable to Windows without much difficulty." While there's more to this, what strategies should a developer take to insure that the resulting program is as crash-free as possible?
"I'm thinking of decoupling the modules physically so that, even if one crashes/becomes unstable (say, the distributed communication module encounters a segmentation fault, has a memory leak or a deadlock), the others remain alive, detect the error, and silently re-start the offending 'module'. Sure, there is no guarantee that the bug won't resurface in the module's new incarnation, but (I'm guessing!) it at least reduces the number of absolute system failures.
How can I actually implement such a decoupling? What tools (System V IPC/custom socket-based message-queue system/DCE/CORBA? my knowledge of options is embarrassingly trivial :-( ) would you suggest should be used? Ideally, I'd want the function call abstraction to be available just like in, say, Java RMI.
And while we are at it, are there any software _design patterns_ that specifically tackle the stability issue?"
How can I actually implement such a decoupling? What tools (System V IPC/custom socket-based message-queue system/DCE/CORBA? my knowledge of options is embarrassingly trivial :-( ) would you suggest should be used? Ideally, I'd want the function call abstraction to be available just like in, say, Java RMI.
And while we are at it, are there any software _design patterns_ that specifically tackle the stability issue?"
I'd hate to say it, but you might want to SERIOUSLY consider managed code. You could build some of the parts in C++ if need to be, but doing it purely in C++ seems like a bad idea to me. You're asking for a silver bullet that just doesn't exist...but managed code is getting faster and can be pretty stable.
I think perhaps what you REALLY mean here by stability is Fault Tolerance. It's impossible to write code that has zero defects, outside of any trivial examples. Real Code Has Real Defects. Now, as you talk about modular design and being able to restart modules, you're talking about, not stability, but fault tolerance; the ability of the application to recognize and recover from faults. For instance, you can't necessarily guarantee that the module on machine A running task B won't die, hell the computer could accidently fry, but if your application was Fault Tolerant then the application would kick off another process somewhere else on computer C to rerun job B. Stable systems aren't built necessarily by trying to write defect-free code, but by recognizing that defects will occur and architecting the system in such a way that it can recover from them. Here you need to be concerned about things like transactions, data roll-back, consistency, techniques (active vs. passive, warm vs. cold). The key thing is before you even write a LINE of this C++ code, make sure that you have a complete, comprehensive ARCHITECTURE for your application that will gracefully handle faults.
The reasons? A unit test suite that implements several million test cases (mostly pseudo-random probes -- the actual test code is about 1/3 the size of the functional code). In fact, the "defects" that hit production were more "oversights"; stuff that didn't get accounted for and hence didn't get implemented.
Just as importantly; every dynamically allocated object just got assigned to a "smart pointer" (see Boost's boost::shared_ptr implementation).
Quite frankly, compared to any Java implementation I've seen, I can't say that "Garbage Collection" would give me anything I didn't get from smart pointers -- and I had sub-millisecond determinism, and objects that destructed precisely when the last reference to them was discarded. The only drawback: loops of self-referencing objects, which are very simple to avoid, and dead trivial if you use Boost's Weak Pointer implementation.
We didn't have access to Boost (which I Highly Recommend using, instead of our reference counted pointer) when we first started the project, so we implemented our own Smart Pointers and Unit Testing frameworks.
I've since worked on "Traditional" C++ applications, and it is literally "night and day" different; trying to do raw dynamic memory allocation without reference counting smart pointers is just insane (for anything beyond the most trivial algorithm). And developing with Unit Testing feels like being beaten with a bat, with a sack tied around your head...
-- -pjk Perry Kundert perry@kundert.ca http://kundert.2y.net
Yes, some of those do conflict. How to keep things simple AND have fault-tolerence, for example. That's where a good design comes in handy, because you can get a better feel for where you should make the trade-off between certainty of working, certainty of working later on and getting some sleep this side of 2008. It's all a matter of weighing the options and investing time in the place most likely to benefit.
(Because everything is a trade-off, anything listed above may not apply. But then, it may not need to. If you've tested a component thoroughly along all boundaries, a good sample of valid conditions and a good sample of erronious conditions, AND everything has been kept as simple as possible so that really wierd cases are unlikely to crop up, then you may decide you can simplify or eliminate fault-tolerent components. There is no point in catching errors that won't occur. In fact, that adds complexity and violates the Keep It Simple rule.)
Oh, and as this is a networked system, testing should include testing network I/O. Use packet generators if necessary, to see how the system handles erronious packets or massive packet floods. You don't want "perfect" responses (unless you can define what "perfect" means), you want reliable responses. If X occur
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Jesus christ people he is asking you how he should go about building an ultra-stable application in C++. He told you he *has* to build it in C++ because there are critical libraries and other components that aren't availible in C++. Telling him he shouldn't build it in C++ anyway just isn't helpfull.
I hate to break it to people but there *are* libraries, especially for types of scientific computing, that are only (reasonably) availible in C++ or sometimes FORTRAN. Not only would abandoning these libraries mean he would completely have to reinvent the wheel but also might cause serious compatibility problems not to mention a much greater ongoing maintenence responsibility (he can't just check his program to make sure things still work when someone fixes a library bug).
Moreover, the idea that because he is considering using CORBA, IPC or whatever else speed can't matter enough to require C/C++ is dead wrong. It is true that whatever *parts* of the process are done using these components may not require huge amounts of speed but this doesn't mean one of these components isn't doing something very processor heavy.
In particular what he says sounds like the situation in some areas of scientific computing. If one is writing a program to do some sort of simulation or similar math intensive operations speed can be *very* important in the critical parts of the code but (in some cases) transfering information to the GUI or other components need not be particularly speedy (increasing by an order of magnitude may make a small difference in overall runtime). Imagine a program that does some kind of weather, or nuclear detonation simulation. The cross-processor communication and the core simulation kernel need to be very fast but the GUI and data input components need not be particularly fast. Also it is my understanding that often the critical libraries in this area are often only availible (at least freely) with C/C++ or fortran bindings.
Anyway I think it is important to distingush several different goals, ultra-stability, minimal downtime, and minimal data/computation loss. For instance a climate simulation that may run on a supercomputer for months it is very important to have minimal data/computation loss (i.e. if something goes bad you don't lose months of very valuable supercomputer time) but you need not have ulta-stability or minimal downtime. As long as when any node crashes the simulation can easily be restarted without loss of data there is no problem. On the other hand if you are running a website like slashdot it is minimal downtime that is important it doesn't really matter if some of the web server processes are rebooted once in awhile. If, on the other hand, you are writing code to monitor a nuclear power plant it is ultra-stability that is important (though I can't at the moment think of something that requires distributed processing and ultra-stability but I'm probably just missing something).
So I think the answer depends on what sort of stability you want. If it is important that no individual *node* crashes (though the GUI/other non-core components can crash) then you should pursue the seperation you described above. I have to admit I'm not an expert here but the client-server model (like mysql, X etc.) seems to work well in this context. However, this depends alot on what sort of data you need to transfer. If you just need to send the core setup commands and get back mostly unstructured info (say a grid of tempratures or other simple datasets) then I would suggest sticking with one of the simpler abstractions and don't get lost in CORBA. On the other hand if you need to send back and forth real objects with significant structure then creating your own serialization system/bindings is just asking for bugs.
On the other hand if what you want is minimal data/computation loss, downtime, or any other property where it is the overall system you care about not a crash at any particular node then I suggest concentrating less on dividing any one node into comp
If you liked this thought maybe you would find my blog nice too:
Avoid the latest "big thing" for the core of your project. It's usually specialized, non-portable, etc. The standard template library for C++ (for example) is here to stay, with tested algorithms that are safer and faster than you can usually write (because they are optimized for the platform you compile on). For the GUI, on the other hand, you may be better off with a GUI-based language/tool. That's less likely to be portable, but that's the way GUIs work.
Next, spend some time upfront on your design, with things like use cases, sequence diagrams, and other visualization tools to help you understand just what you want to happen in best case situations as well as failures. The level of detail/formality required is a moving target, so update as needed. You should have a solid error detection/correction plan so that you can design each component to follow it. Also design for test and with logging - it will help you while debugging, while testing, and while fixing the bug the customer is seeing.
Make sure management will allow sufficient time for testing. A lot more lip service goes into support for testing than actual schedule and money. Your test plan should be as bulletproof as your design.
That's my 2 cents. And a random book recommendation: books like Scott Meyers' "Effective " provide info on effective/error reducing ways to use the language/libraries, but won't help you get started with the architecture.
Perl? Fuck. He wants a stable app with good code. Sheesh.
The secret of stable system design is designing from failure. Designing and implementing defensively. If you want to design an ultrastable system you start with the failure analysis for every component, following with failure analysis of modules and the entire thing as it grows.
This in the world of C++ (and C for that matter) quite often means checking paranoiacally everything everywhere for NULLs before doing anything about it.
Designing and writing from failure means that every system or library call should be assumed to fail first and all failures handled cleanly. This may be quite painfull because it usually requires the development of special tools like wrappers around malloc, file calls, etc that return error conditions which are nearly impossible to achieve on a live system.
Only after all codepaths for "bad" results have been handled, the actual "normal" codepaths should be written. This unfortunately is not the way code is written in 99% of the shops out there. Most design and implement from success first and add failure handling later.
Just ask in your shop: "Where is our memalloc wrapper that simulates a failed memory allocation? I need to link versus it to do some testing to see how our app handles NULLs in a few places". The usual answer you will get is "Ugh? WTF you are talking about Dude... We do not smoke that stuff here... Just go and write the code you have been assigned to write..."
And the results are quite bloody obvious.
Baker's Law: Misery no longer loves company. Nowadays it insists on it
http://www.sigsegv.cx/
The suggestions I have seen here so far seem to boil down to "Don't do it that way". Sometimes that's not possible. If it truly has to be C++, and it truly has to be as fast as possible and as bug free as possible, there are a few guidelines that can help:
1. Unless the GUI will be I/O bound, and that's unlikely, try to write it in a safer language that has better GUI support.
2. Make all your classes small and simple, and create test harnesses that are as complete as possible. Try to make the classes simple enough that they can be individually tested in such a way that all code paths are exercised.
3. Check your arguments. This includes checking for invalid combinations, and arguments that are invalid given the state of the object.
4. Don't use new or pointers directly. If there may be multiple references to an object, then reference count it and create handle classes that hold the references so all instantiation is controlled, and all destruction is implicit. Make these handles STL compatible, and never pass around pointers to them.
5. Try to design the application to fail fast and recover from failure. For example, maintain the state of work being done in discrete transactions that can be aborted if a failure is detected. This can be on disk or in memory depending on your performance needs. This could be combined with the ability to restart the app in a new process and have it pick up where the last one left off.
6. Have the app keep track of its memory usage, and be prepared to recover from memory leaks, possibly by restarting as in item 5.
7. If the compiler you're using supports structured exceptions, then use them. They can degrade performance a bit, but they can also enable you to recover from NULL pointer exceptions.
8. If you have multiple threads, then to avoid both the performance hit from context switches and the chance of deadlocks, don't let them access the same data directly. Instead, have them communicate through lock free queue structures. That way, all your main threads can pretty much spin freely. Spawn worker threads for any I/O or other operations that can block. A context switch can take as much time as thousands of instructions. You want to use as much of every time slice as possible.
9. Keep the number of main threads down to the number of CPU's or less. That way, except for the times when the CPU is being used by the OS or other processes, (should be relatively rare) each non blocked thread gets its own CPU.
10. Have an experienced QA team, that understands their job goes beyond unit testing.
Now here's a few that are always important, but for what you want to do, they become critical.
11. Have the design laid out at least roughly before you start.
12. If at all possible, don't let requirements change in midstream.
13. Overestimate the time it will take very generously. You will probably still be crunched.
pornking
Question 1: what strategies should a developer take to insure that the resulting program is as crash-free as possible?
Answer:
a. Use OO techniques and maintain all objects in your system extremely simple; furthermore, maintain all methods in your system extremely short, well-contained, well-defined.
b. Don't use C++ arrays, ever. Especially not for strings. Use and abuse the STL. is just plain beautiful IMH?O.
c. Check extensively the behaviour of your constructors and destructors.
d. Make a object-lifecycle diagram of each class you program. In the diagram, relate it to the neighboring classes (parents, children, siblings, classes involved in design patterns with, classes aggregated, classes value-aggregated, classes where this is aggregated or value-aggregated)
e. Use, carefully, and always when possible, smart pointers. Remember std::auto_ptr is your best friend -- its limitations are a defining part of its strength. Remember boost::shared_ptr is also a good friend, but its cousin boost::intrusive_ptr is even more friendly -- but use one of those (and their other cousins scoped_{ptr,array}, shared_array, weak_ptr) only in the (rare) cases where auto_ptr does not apply.
f. As a corollary to (e) above, use boost. This is really an extension of (b), too.
Question 2: How can I actually implement such a decoupling?
Answer:
I would use a simple, socket-base, take-my-data, gimme-my-results scheme. It would be network-distributable, easy to detect if some service is or isn't alive via timeouts... If you want something more sofisticated/RMI-like, SOAP (with binary XML or compressed) may be an option. The simpler the better IMHO.
Question 3: are there any software _design patterns_ that specifically tackle the stability issue?
Answer:
All of them? IMHO, DPs can represent huge tool to increase the stability of a system. Take a look athere [WARNING: PDF] (and in the bibliography) for some ideas.
I know many of my posts were self-marketing lately, but if you need someone to work with you, I'll be happy to send you my resume... write me at hmassa (at) gmail.
It's better to be the foot on the boot than the face on the pavement. ~~ tkx Kadin2048