Code Quality In Open and Closed Source Kernels 252
Diomidis Spinellis writes "Earlier today I presented at the 30th International Conference on Software Engineering a research paper comparing the code quality of Linux, Windows (its research kernel distribution), OpenSolaris, and FreeBSD. For the comparison I parsed multiple configurations of these systems (more than ten million lines) and stored the results in four databases, where I could run SQL queries on them. This amounted to 8GB of data, 160 million records. (I've made the databases and the SQL queries available online.) The areas I examined were file organization, code structure, code style, preprocessing, and data organization. To my surprise there was no clear winner or loser, but there were interesting differences in specific areas. As the summary concludes: '..the structure and internal quality attributes of a working, non-trivial software artifact will represent first and foremost the engineering requirements of its construction, with the influence of process being marginal, if any.'"
Is it just me? (Score:5, Interesting)
Of course, I could try to RTFA, but hey, this is Slashdot, after all...
Not that surprising (Score:5, Interesting)
Interesting, but not shocking for those who have worked with disciplined commercial teams. I wonder what the results would be in less critical areas than the kernel, say certain types of applications.
Re:Is it just me? (Score:4, Interesting)
The 99% Solution (Score:5, Interesting)
This would explain some things like lower LOC count - after all, if you just have a bunch of global functions there's no need for a lot of API wrapping, you just call away.
I do hate to lean on LOC as any kind of metric but - even besides that, the far lower count of Windows made me wonder how much there, is there. Is the Windows kernel so much tighter or is it just doing less? That one metric would seem to make further conclusions hard to reach since it's such a different style.
Also, on a side note I would say another conclusion you could reach is that open source would tend to be more readable, with the WRK having a 33.30% adherence to code style and the others being 77-83%. That meshes with my experience working on corporate code, where over time coding styles change on more of a whim whereas in an open source project, it's more important to keep a common look to the code for maintainability. (That's important for corporate code too - it's just that there's usually no-one assigned to care about that).
Re:statistical wash-out? (Score:3, Interesting)
It sounds more like they're saying "If someone built it, and someone else is using it, and it's important, then the code quality is going to be pretty good. If it matters, it's going to get attention and be improved."
Of course, I can think of a bunch of counter-examples in Windows where something was important *to me* and mattered *to me* and no one at Microsoft saw fit to do anything about it for decades.
Re:Closed Source? (Score:5, Interesting)
The WRK is under the Microsoft Windows Research Kernel Source Code License [microsoft.com]. I'm not sure that this license conforms with anyones definition of open source, but it's reasonably free for reasearch.
But PP addresses a crucial point, if something really is closed source there is no reviewable way to compare and present this code. So if the WRK would be total crap they could always say: yes that's only the WRK, not the real kernel.
Only statements about open source code are directly verifiable/falsifiable. One of the reasons, why the FOSS approach is superior from a scientific as well as technical point of view.
So.... (Score:3, Interesting)
Closed Source Developer: I will try to do my best job as I possibly can so I can keep my job and make money because that is what I value.
Open Source Developer: I will try to do my best job as I possibly can so I can help the comunity and feel better about myself/get myself noticed in the comunity/Something cool to put on my resume... because that is what I value.
People who choose to license their software OpenSource vs. Closed Source says nothing about their programming ability. There are a bunch of really crappy GNU projects out there as well as a bunch of crappy closed source projects... Yea there is the argument of millions of eyes fixing problems but really when you get millions of people looking at the same thing you will get good and bad ideas so the more good ideas you get the more bad ideas you get and the more people involved the harder it gets to weed out good ones and bad ones. Closed source is effected often by a narrow level of control where bad ideas can be mandated.... All in all everything really ballances out and the effects of the license are negledgeable.
An interesting point.. (Score:5, Interesting)
From my personal experiences, it typically seems code is written to solve a specific need. Said another way, in the pursuit of solving a given problem, whatever engineering is required to solve the problem must be accomplished - if existing solutions to problems can be recognized, they can be used (for example, Gang of Four/GOF patterns), otherwise, the problem must have a new solution engineered.
Seeing as how there are teams successfully developing projects (with both good, and bad code quality) using traditional OO/UML modeling, the software development life-cycle, capability maturity model, scrum, agile, XP/pair programming, and a myriad of other methods, it would seem to be that what the author is saying is, it didn't necessarily matter which method was used, it was how the solution was actually built (the.. robustness of the engineering) that mattered.
Further clarification on the difference between engineering and "process" would strengthen this paper.
I went to a Microsoft user group event some time ago - and the presenter described what they believed the process of development of code quality looked like. They suggested the progression of code quality was something like:
crap -> slightly less crappy -> decent quality -> elegant code.
Sometimes, your first solution at a given problem is elegant.. sometimes, it's just crap.
Anyways, just my two cents. Maybe two cents too many..
SixD
Re:Not that surprising (Score:3, Interesting)
Generally speaking, commercial desktop apps are still way ahead of their open counterparts, with the exception of code development tools and anything that directly implements a standard (browsers, mail clients, etc.)
One reason for this is that code quality as measured in this study may not directly relate to application quality as measured by the typical user. Photoshop is "good" not least because of its well-understood interface and the fact that everyone uses it, regardless of how admirable the code is.
What I took from it was: (Score:2, Interesting)
How ever I was left wondering how it was possible to compare fairly? He already stated:
"Excluded from the kernel code are the device drivers, and the plug-and-play, power management, and virtual DOS subsystems. The missing parts explain the large size difference between the WRK and the other three kernels."
and reading I see even more of the drivers aren't there:
"The NT Hardware Abstraction Layer, file systems, network stacks, and device drivers are implemented separately from NTOS and loaded into kernel mode as dynamic libraries. Sources for these dynamic components are not included in the WRK. "
http://www.microsoft.com/resources/sharedsource/licensing/researchkernel.mspx [microsoft.com]
So it's not like for like. Maybe you would draw different conclusions if it was, maybe the Linux style issue is because of all the drivers the WRK lacks. So even though I think his conclusion sounds probable, I don't feel I can state it as so with any confidence.
Stupid metrics (Score:4, Interesting)
The metrics used in this paper are lame. They're things like "number of #define statements outside header files" and such.
Modern code quality evaluation involves running code through something like Purify, which actually has some understanding of C and its bugs. There are many such tools. [wikipedia.org] This paper is way behind current analysis technology.
"Code quality" is bunk (Score:5, Interesting)
The worst looking piece of spaghetti code could have fewer bugs, be more efficient, and be easier to maintain than the most modular object oriented code.
What is the "real" measure of quality or productivity? Is it LOC? No. Is it overall structure? no. Is it the number of "globals?" maybe not.
The only real measure of code is the pure and simple darwinian test of survival. If it lasts and works, its good code. If it is constantly being rewritten or is tossed, it is bad code.
I currently HATE (with a passion) the current interpretation of the bridge design pattern so popular these days. Yea, it means well, but it fails in implementation by making implementation harder and increasing the LOC benchmark. The core idea is correct, but it has been taken to absurd levels.
I have code that is over 15 years old, almost untouched, and still being used in programs today. Is it pretty? Not always. Is it "object oriented" conceptually, yes, but not necessarily. Think the "fopen,"fread," file operations. Conceptually, the FILE pointer is an object, but it is a pure C convention.
In summation:
Code that works -- good.
Code that does not -- bad.
Re:Is it just me? (Score:5, Interesting)
Re:Is it just me? (Score:4, Interesting)
Yup, and the author of the paper is Diomidis Spinellis, who wrote the excellent book Code Reading [spinellis.gr]. This is a great study of code analysis and familiarization techniques. He also wrote a fine article on C preprocessors... in Dr. Dobb's Journal, I think.
Re:"Code quality" is bunk (Score:4, Interesting)
This example [ioccc.org] is code that works and also has some nice quality attributes: 96% of the program lines (631 out of the 658) are comment text rendering the program readable and understandable. With the exception of the two include file names (needed for a warning-free compile) the program passes the standard Unix spell checker without any errors.
This example [ioccc.org] is also code that works, and is quite compact for what it achieves.
I don't consider any of the two examples quality code. And sprucing bad code with object orientation, design patterns, and a layered architecture will not magically increase its quality. On the other hand, you can often (but now always) recognize bad quality code by looking at figures one can obtain automatically. If the code is full of global variables, gotos, huge functions, copy-pasted elements, meaningless identifier names, and automatically generated template comments, you can be pretty sure that its quality is abysmal.
Re:"Code quality" is bunk (Score:3, Interesting)
"sort" of working is not "working."
exists a 6000 line SQL statement that no one understands
This is "bad" code because it needs to be fixed and no one can do it.
Surely you wouldn't define good architecture as "a building that remains standing,"
I'm pretty sure that is one of the prime criterion for a good building.
Your post ignores the "works" aspect of the rule. "Works" is subtly different than "functions." "Works" implies more than merely functioning.
Re:Not that surprising (Score:3, Interesting)
NetBSD is the one to compare here (Score:3, Interesting)
Isn't NetBSD the system filled with academics who insist upon clean, manageable, and portable code above all other standards? Too bad the NetBSD kernel didn't get judged here, I suspect it would have taken the cake.
I still recall this exhaustive report [bulk.fefe.de] comparing several kernels' performance back in 2003 in which NetBSD pretty much beat the pants off of everybody else (note the two updates with separate graphs). The initial poor performance was due to an old revision, and upon seeing that there were some places in which the newer revision wasn't so hot, the developers fixed them and in only two weeks, NetBSD beat out FreeBSD on every scalability test. Their pragmatism and insistence on quality code finally paid off.
Ever since seeing those charts, I've been waiting for Debian/NetBSD [debian.org] to come out...
Re:Is it just me? (Score:5, Interesting)
% style conforming lines: FBSD:77.27 LIN:77.96 SOLARIS:84.32 WIN:33.30
% style conforming typedef identifiers: FBSD:57.1 LIN:59.2 SOLARIS:86.9 WIN:100.0
% style conforming aggregate tags: FBSD:0.0 LIN:0.0 SOLARIS:20.7 WIN:98.2
(I'm far too lazy to clean up the rest)
% of variable declarations with global scope 0.36 0.19 1.02 1.86
% of variable operands with global scope 3.3 0.5 1.3 2.3
% of identifiers with wrongly global scope 0.28 0.17 1.51 3.53
Re:Closed Source? (Score:3, Interesting)
Re:So.... (Score:3, Interesting)
Simplistic approach (Score:1, Interesting)
E.g., measuring the average length of the files or the average number of files in a directory reflects an underlying assumption: software quality is better when granularity is higher. But it is not true.
The reality is that you need an *appropriate* level of granularity in your software, and this also depends on your overall design. It's a little bit nonsense to try measuring such properties of the code as if they were somehow absolute and context-free !
Such criterions and the related underlying assumptions seem very debatable: thus its no wonder that no significant difference can be extracted out of these measurements.
Re:Is it just me? (Score:4, Interesting)
Re:Not that surprising (Score:5, Interesting)
Preprocessing: here we go again (Score:3, Interesting)
That subjective conclusion is precise effect reading too much into the metrics.
Sun or Microsoft programmers need to support 2 and 2 platforms respectively. (Sun: SPARC and AMD64; M$: IA32 and AMD64). All portability are of boolean complexity.
But FreeBSD and Linux run on dozen of platforms. I do not know how it is in BSD land, but in Linux first and foremost requirement for platform support, is that it has no negative side-effects on other platforms. Consequently, for example, under Linux most (all? - all!!) locking is still implemented as macros: on uni-processor system with preemptive kernel feature disabled all in-kernel synchronization would miraculously (thanks to preprocessor) disappear from the whole code base. To make sure that on such platform, kernel would run as efficiently as possible - without any locking overhead, because all the locking is not needed anymore.
And that's single example. There are many macros for special CPU features: depending on platform it would be nop or asm statement or function call. No way around using macros.
I think one of the points the author needed to factor in, is portability of OS. Without that, most metrics are skewed too much.
P.S. Actually, Linux affinity to macros is often (at least from words of kernel developers) stems from poor optimization of inlined functions in GCC. Many macros can be converted to functions - but that would damage overall level of performance. In many places significantly.
Re:question (Score:4, Interesting)
Re:"Code quality" is bunk (Score:3, Interesting)
Not to put too fine a point on it, but this is too much concern over stuff that does not always matter.
I agree "functional" and "reliable" are absolutely important.
"efficient?" Only if efficiency is required or of any concern. How efficient is efficient? It is a balance of efficiency against economy.
"Maintainable?" Sure, most of the time, but not always. Sometimes we toss stuff on purpose. Some good code can be written that is near to impossible to maintain. Some code sacrifices maintainability for performance.
"Portable?" There is little sense in making some code portable. PIC software, for instance, can not be good "PIC" code and portable because the PIC is such an insane type of device.
I think your cute little chart is all well and good for a common segment of software development, but it is hardly an absolute and puts too much emphasis on an arbitrary set of criteria that don't reflect the purpose or economy of software code.