The firge, the door and the fire alarm

Today when I prepared breakfast I realized that the fridge wasn't properly closed. It had been a small opening the whole night. That's not particularly good.

But what's worse is that the fridge did not make any noise indicating that the door was open. In fact, it did the opposite: it turned off the only indication there was that the door wasn't properly closed -- it turned of the light inside the fridge. If that light wasn't turned off it would be easier to spot that the door was open.

Turning the light off would not be a big problem if there was some way the fridge alarmed when the door was open. Let's say making a noise when it had been open for more than 1 minute.

I'm sure that it's a feature that the fridge turns off the light by it self when it has been on for too long. But I can't the use-case when it would be a useful. There must be a timer somewhere that turns off the light. That timer should instead trigger an alarm.

Actually there is an noise-making-device in the fridge. But that only used when the temperature in the fridge is above a certain temperature. That noise-making-devices should be triggered by that timer. Why didn't it?

I don't know.

Recently I've noticed more and more weird design choices in everyday things. Like having the handle of a door be shaped as a pipe when you should push the door, and having the handle be shaped like a flat surface when you should pull the door (think about it, a surface is easy to push and a pipe is easy to pull).

Even worse, I've experienced that fire alarms sounds very very much like the break-in alarm.

Mim: the build system you always wanted

The word mim means any of the following:

• an incorrect way of writing 1999 in roman numbers (as in "we gonna party like it's 1999"),
• the Madame Mim from the Disney movie Sword in the Stone,
• Swedish for "mime" meaning "imitating"
• a figure in Norse mythology renowned for his wisdom
• an acronym for "Mim Isn't Make"

it's also the name of the build system I've been thinking and working of for a while. This post is about what Mim is now and what it can become. Text in red describes stuff that isn't implemented yet, text in yellow is things that are kind-of implemented, and text in black describes implemented features. First, I'll describe how Mim is different from other build systems and why it's better.

Problem
The problems with traditional build system are:

• make sure the dependencies are correct and built in the right order (think: make depend && make && make install),
• be sure of that every built file you see is up to date,
• hard understand how a software project is built (what are the artifacts? where are they stored? what are the dependencies?),
• hard to understand which variables a build have (e.g., make DEBUG=1)

Solution
The solution to these problem is (of course) Mim. With Mim you clearly see all artifact the build system produces, how they are built, and when they need to be built. In fact, you can see the artifacts, e.g., by doing ls, before they are built. In other words, using your normal command line tools you can browse the project file tree to find the artifact you need (to run, to view, to copy, etc.) without building anything. Sounds like magic? Keep reading...

When when you found the program you need to get built you simple execute it. No need to issue any command yourself to build it. The program will be automatically built for you and then executed. Similarly, if you have a tool that generates a text file as part of the build, you can do emacs file.txt. The file will be automatically generated and then opened in emacs.

If you need to get more information about a file you can do mim ls filename. You'll get something like this written to the console:
-r-xr-xr-x you users filename (g++ filename.cc -o filename)

that is, you can easily discover to the access rights, the owner, the group, and how the file is built, of any artifact simply by locating it in the file system.

This is especially powerful when you work with project of which source code generation is part of the building process. Understanding what files are generated, how they are generated, can be very hard. With Mim, however, this is very easy, because every file (generated or not) looks the same to you when you browse the file tree. Additionally, you can ask Mim for additional information about a certain artifact file, e.g., get the command for building the file by doing mim cmd file.

Furthermore, getting the dependencies right when the build involves several steps (generating .d files, generating source code, generating .d files for the generated code, compiling the source code, linking the source code against libraries (which themselves has to be built, including its generated source code)) is also very very hard. And understanding it a few month later is even harder. Mim solves all these problems.

Let's take an example. Let's assume we have a directory called hello with a file hello.cc contaning an implementation of Hello, World. There is also a file called Mimfile that tells Mim how to build this program. Doing mim ls in hello gives you:

greeting.txt
hello.cc     codegen -lang en greeting.txt -o hello.cc
hello-en     g++ hello.cc -o hello-en
Mimfile

which tells you that there is only one physical file, greetings.txt, in this directory (except the obligatory Mimfile) and two artifact files, hello.cc and hello-en, of which the latter is an executable. You can also see that there is some code generation involved by looking at the command line for hello.cc. You can tell all this by issuing one simple command, mim ls, without building anything at all.

Let's say you wish to take a peek at the hello.cc file. All you need to do then is to do cat hello.cc. No need to worry that the code you see is out of date because Mim makes sure it's up to date before it's opened.

Being lazy
Being lazy is a good attribute; as long as you're not too lazy. Being lazy in the context of a build system means that you only rebuild a file if is inputs have changed. How do Mim achieve this?

Mim woks by intercepting all read and write accesses to the file system. This means that Mim has perfect knowledge of 1) which files are read when a given artifact if built, and 2) which files that have been updated since the artifact was last built. This means that artifacts are only rebuilt when needed.

This information makes it possible for Mim to easily construct a graph dependencies and what needs to be built. This implies that Mim start building the leaves of the graph as early as possible to reduce the build time.

Instant messaging

In large project there is usually variables that controls how the project is built. Mim supports such variables in a nice way. To list the variables simply do mim conf, you'll get something like this printed to the console:

lang      en   en|fr|swe  Controls in which language "Hello, World" is printed.
debug     off  on|off     Compile with or without debug information.
optimize  0    1|2|3      Optimization level as defined by gcc:s -O flag.

The first column is the the name of the variable, second column is its current value, third is its possible value, and last column is a textual description of it.

Changing the value of the variable lang to swe is done by doing mim conf lang=swe. Setting a variable have instant effect. For example, if a variable affects which files are built, you can see the updated list of files immediately by doing ls without building anything. Example:

$ ls hello*
hello.cc  hello-en
$ mim lang=swe
$ ls
hello.cc  hello-swe

That is, by changing the variable lang from en to swe we get an artifact file called hello-swe instead of hello-en like we got before.

To aid discoverability even more, you can do mim ls hello* lang==* to list all files matching hello* for all values of lang. For the hello example this outputs:

hello-en   lang=en   g++ hello.cc -o hello-en
hello-fr   lang=fr   g++ hello.cc -o hello-fr
hello-swe  lang=swe  g++ hello.cc -o hello-swe

All these features helps you in learning how an a project is built, how to use its build system, how to change it, how those changes affect the build system, and more.

The Tracker
Since Mim intercepts all accesses to the file system, it can help you when you do potentially bad things, for example, if you delete a file that is needed to build an artifact. Mim also updates your Mimfiles automatically when you do non-destructive changes to your source tree. For instance, when you rename an input file to gcc as the following example illustrates this:

$ mim ls hello*
hello.cc  codegen -lang en greeting.txt -o hello.cc
hello-en  g++ hello.cc -o hello-en
$ mv hello.cc hi.cc
$ mim ls hello*
hi.cc     codegen -lang en greeting.txt -o hi.cc
hello-en  g++ hi.cc -o hello-en

Mim can do this automatic update of Mimfiles correctly in many cases, but it's impossible to get it right every time since that would require deep knowledge of all possible build tools (e.g., gcc, ld, m4, etc). However, Mim makes certain (reasonable) assumption on how the build tools behave, and as long as the build tools' conform to these assumptions Mim gets it right most of the time. Anyway, to check that Mim got it right, you simple build of application so this isn't a big problem in practice.

Getting general

Mim is built around the concepts of events and actions triggered by event. So far we have implicitly discussed the event open artifact for reading and the action was update artifact. The possible events that can trigger actions are:

• open physical of artifact file for reading (before-read),
  
• closing a physical or artifact file after write (after-write),
• deleting physical or artifact file or directory (delete-file),
• creating a physical file or directory (create-file), and
  
• moving a physical or artifact file (move-file).
  

Since arbitrary commands can be executed as the result of an event, you can do what ever you heart desire when a file is moved, delete, created, etc. However, actions that interact with Mim would be hard to implement ourself and are thus built into Mim:

• update an artifact (without reading the artifact file)
  
• log to the Mim log (e.g., an error message)
  
• reject event (e.g., reject opening a file).
  

The reject event action is useful, e.g., if you need to verify the consistency of a save file: if the file is consistent it cannot be saved.

Playing nice
So far so good. But what about integrating with other non-miming tools like, say, make. Since all Mim does to build an artifact files is to execute a command line in the shell, Mim can easily be used as a front-end to make. Here is an example of doing mim ls in a directory of a project that does just that:

main.cc
hello     make hello
Makefile
Mimfile

Furthermore, any program can use Mim for building files, because all the program need to do to build a file is to (try to) open it for reading. Mim will then kick in and build the file, without the other program ever nothing anything special. This is very powerful, because every program you have on you computer, cat, emacs, Mozilla, etc, can build any file using Mim.

Saving the environment
Do you have ./configure && make && sudo make install aliased to nike on you system? If you do you're not using Mim.

One reason to use autoconf (i.e., the ./configure script above) is to find the paths to all the necessary tools, libraries, headers, etc, needed to build your project. With autoconf this involves generating Makefiles that are specifically targeted at your system. The problem with autoconf and the generated Makefiles is that you have to be a genius to understand what the heck is going on.

So what do Mim do instead? Easy, you have an artifact file for each tool, library, header, etc, that is needed by your build. Those artifacts are configured (e.g., using autoconf) to link to a specific file on your system. These tools (or, ) are normally placed in a directory called ext (for "external"); thus, to list the dependencies to external tools simply do find ext or ls -R ext. Every file listed is an external dependency. Example:

$ cd my-project/ext
$ ls -R *
ext/bin:
g++  ld  python

ext/include:
3rdparty.h

ext/lib:
3rdparty.so

In this example we have configured the ext directory to contain a compiler, linker, and a python interpreter. So, when we need to compile a .cc file we have the following command line in the Mimfile ext/bin/g++ file.cc. The first time this command line is executed, the artifacts (ext/bin/g++, etc) are not yet configured, so Mim launches the configuration tool (autoconf). After that, the artifacts in the ext directory point to the correct files on the local system and can be executed easily.

A way of thinking of this is that Mim creates a virtual environment (directory structure) that is ideal(ized) for you specific project. All the tools and libraries you need are right there in the ext directory.


Deep understanding
Mim really, really, understands mimfiles. If a mimfile is updated, artifacts potentially need to be rebuilt. However, Mim will only rebuild artifact which command lines has changed or if their dependencies' command lines are changed. Have you ever experienced that your entire project needs to be rebuilt just because you added an a comment to a makefile? That will never happen with Mim.

Mim has even support for refactoring mimfiles via your filesystem! In the hidden directory .mimfile the artifact files' mimfile-data is represented as files. So, if you want to change the name of an artifact file simply do mv .mimfile/old-name .mimfile/new-name; to create a new artifact file that is identical to an existing one do cp .mimfile/original .mimfile/copy; to delete an entry from the mimfile do rm .mimfile/remove-me. Finally, to edit the command line for an artifact you simply edit the corresponding file under .mimfile, e.g., echo 'gcc input-file.c -o %output' > .mimfile/file-to-be-edited.

As mentioned earlier, Mim also has variables that can be used to configure how your project is built. These variables are also represented as files under the directory .mimvars. Here you can easily browse the variables using ls, find, grep, etc, or even your graphical file browser. The get the current value of a variable, simply do cat .mimvars/the-variable and to set the value do something like echo 'new value' > .mimvars/the-variable.

When thing go wrong

Ok, so Mim is nice, I think I've communicated that by now. But sometimes the world is very very unnice. Sometimes the world don't want to compile your code. Then you need to either 1) fix the world, or 2) fix your code. To my experience the latter is slightly easer than the former, although your mileage may vary.

To fix your code you need to know what the compiler complains about, but how do you do this when the files is compiled behind the scenes? Simply do mim log errors to print the error messages to the console, fix the error, and run the program again.

Doing mim log error after every build error gets annoying very fast. Thus there is a variant, mim log errors -f, that is useful when you wish to get continuous feedback. Given the -f flag Mim will run in the background and print the error message to the console as soon as they appear. So, assuming hello.cc contains an error, trying to execute hello will print the flowing:

$ mim errors -f
$ ./hello
hello/hello.cc: In function ‘int main()’:
hello/hello.cc:3: error: ‘printf’ was not declared in this scope

which is is just what gcc outputs, thus, any tool that parses the output from gcc will understand the error output from Mim as well. This is of course also applicable for other tools like make.

Summing up
I've described the ideas behind Mim, its current state, and possible future features (that, by that way should be very reasonable to implement :)). You can find Mim here. You'll need Intel Threading Building Blocks and Lua 5.1. I've developed using Ubuntu and also compiled it on Suse 11.2. The instructions for how to build Mim itself is currently horribly missing, so please contact me if you like to try Mim out.

DSL: Domain Specific Logging

A few years ago, I was asked to to design a logging framework for our new Product X 2.0. Yeah, I know, "design a logging framework?". That's what I thought too. So I didn't really do much design; all I did was to say "let's use java.util.logging", and then I was done designing. Well kind of anyway...

I started to think back on my first impression of the logging framework in Product X 1.0. I remembered that I had a hard time understanding how it was (supposed) to be used from a developer's point-of-view. That is, when faced with a situation where I had to log, I could not easily answer:

• to which severity category does the entry belong (FINEST, FINER, FINE, INFO, WARNING, SEVERE, ERROR, etc)
• what data should be provided in the entry and how it should be formatted, and
• how log entries normally is expressed textually (common phrases, and other conventions used).

In other words, it was too general (i.e., complex) with too much freedom to define your own logging levels, priorities, etc.

That's why I proposed to make logging easier, thus, the logging framework in Product X 2.0 has very few severity levels. This is implemented by simply wrapping a java.util.logging.Logger in a new class that only has the most necessary methods.

This simplification reduced the code noise in the logger interface, which made it a lot easier to choose log-level when writing a new log message in the source. This had an important implication (that we originally didn't think of): given that a certain message had to be logged, we (the developers) agreed more on which level that message should be logged. No more, "oh, this is a log message written by Robert -- he always use high log level for unimportant messages".

Of course, you could still here comments like "oh, this log message is written by John -- he always use ':' instead of '=' when printing variables". This isn't much of an issue if an somewhat intelligent being is reading the logs, e.g., a human. However, if the logs are read by a computer program, this can be a big problem -- this was the case for Product X 2.0.

This variable-printing business could be solved quite easily; we simply added a method to the logger class (MyLogger) called MyLogger variable(final String name, final Object value) that logged a variable. Example:

logger.variable("result", compResult).low("Computation completed."); 

which would result in the following log message:

2008-apr-12 11:49:30 LOW: Compuation completed. [result = 37289.32];

When I did this, I began to think differently about log messages. It put me into a new mind-set -- the pattern matching mind-set. Patterns in log messages started to appear almost everywhere. Actually, most of our logs followed one of the following patterns:

• Started X
• Completed X
• Trying to X
• Failed to X
• Succeeded to X

Here is an example:

try {
  myLogger.tryingTo("connect to remote host").variable(
                    "address", remoteAddress).low();
  // Code for connecting.
  myLogger.succeededTo("connect to remote host").variable(
                       "address", remoteAddress).low();
} catch (final IOException e) {
  myLogger.failedTo("connect to remote host").cause(e).medium();
}

To take this even further, it is possible to only allow certain combinations of succeededTo, tryingTo, cause, etc, by declaring them in different interfaces. Let's assume that it should only be possible to use cause if the failedTo method have been called. Here's the code for doing that:

interface MyLogger {
  // ... other logging methods.
  CauseLogger failedTo(String msg);
} 

interface CauseLogger {
  LevelChooser cause(Throwable cause);
  LevelChooser cause(String message);
}

interface LevelChooser {
  void low();
  void medium();
  void high();
} 

This is essentially what's called fluent interfaces. It complicates the design of the logging framework, but also makes it impossible to misuse it. Good or bad? It's a matter of taste, I think, but it's a good design strategy to have up your sleeve.

The only design pattern is small solutions

I just saw this TED talk (which you need to see too!). It's essentially about how we prefer big complex solutions to any problem we face. Why? Because it makes us feel smart and important. As my head is filled software thoughts, I started to think how this it relates to software design. We software developers really!, really!, really!, like big solutions to small problems: "Oh, you got a program that needs to store some data? You better get yourself a dedicated database machine, a persistence layer, and define a XML schema for communication data format."

We don't need big solutions to small problems. Big solutions are easy to find. Big solutions need man-hours but no understanding. We need small solution to big problems. Small solutions are hard to find. Small solutions need insight into the actual problem we're solving. The actual problem is what's left when we remove all accidental complexity, marketing buzz-words, etc, and think clearly about the original problem.

Small solutions are orthogonal to each other; big solutions are not, they interact in non-obvious ways. Thus, big solutions creates more problems, or as the american journalist Eric Sevareid, said:

The chief cause of problems is solutions

which is more true in software development than in most other areas. Implement small solutions to problems and your future self will thank you. Implement big solutions and you fall for the sirens' calls of the marketeers, or your own wishes to do seemingly cool stuff while looking smart doing it. Do you really need a DSL? A database? Web interface? Reflection? Operator overloading? Meta-programming? Code generation? Ruby? SOAP?

Thinking small have big impact.

If it hurts do it more often

As a kid, most of what you did was related to learning. Some call it "playing" or "being curious", but whatever you call it, it's learning in one form or another. And learning hurts. A lot.

So you start going to school to make learning not hurt so much. As a 7-8 year-old you enjoy school, because it makes learning much easier. Also, you can play with your class mates, like say, play soccer. Oh, by the way, playing soccer makes you run faster longer, gives you better balance, gives you better feeling of how flying objects behaves (like a ball), better timing your own movements to others (like your team mates). It also makes you better winning and loosing (you aren't a bad looser, are you? How about a bad winner?), and it makes you better at focusing on a particular task, and making other (your team mates) focusing on the same task as you. It's all learning.

As you grow older you become more comfortable with most things you do because you know them. When you finally leave school to start working as a programmer you lost all will to do things that hurts... because you lost your will to learn (in comparison to how much you wanted to learn stuff as a kid). You lost your will to discover new things.

Stupidity hurts

So, when stated with a problem that forces you to do something you don't know or something you really don't like, you stall. Stall, stall, stall, because you don't want to do it. You want to do something fun. Something you know how to do. Something that makes you feel secure and comfortable. Something that makes you feel less stupid!

Yes, you are stupid. I'm stupid. We are all stupid. No one know everything, so everyone is stupid at something. And we will remain stupid at those things unless we learn how to do them better. But... Ouch! Remember? Learning hurts! So we think to yourselfs "Ooh, I don't want to do that!", or put slightly different "I may get hurt doing that! I'd rather stay stupid!". Imagine if you thought that way as a 8-year-old kid in school... or at the playground, or at the soccer field. We wouldn't get many marathon runners or Nobel prize winners that way, would we?

The cure

It's really easy: dare to be stupid. Dare to ask stupid questions. After a while, you'll notice that you start to ask really hard questions, and before you know it you'll ask questions no one (you know) have answers to. And then, as it happens, people will ask you instead. If they dare to of course (hint: they should).

In fact, I think that people will be more likely to ask you (stupid) questions because they'll seen or heard (of) you ask (stupid) questions. And you know what, people do as other people do. Especially like people they look up to, and since you know so much (from asking stupid questions) they look up to you. Isn't that neat?

The workplace

If you have colleagues who don't mind potentially looking stupid by asking a question, you've work at a good company, I think. But I'm not really someone who read or thinks about this a lot, so I may be wrong. There may be effect here that I'm clueless about. Anyway, my feelings are that I'd be more at home at a company like that than a company that encourages silence and really intelligent questions only. But I'm just me. You are you, and you may feel entirely different about this.

Design for optimizability

Only optimize when you found a bottleneck by measuring. That's what we all have been taught, right? It's a good rule. But is it really right? What if you're designing a performance critical application? Should you really be ignorant about performance until you measured and found a bottleneck?

I say: don't optimize, but plan for optimizability. I think Dwight D. Eisenhower said it best, though

Plans are nothing; planning is everything.

And now a disclaimer. I am not encouraging over-design in this post. Neither am I proposing that you should analyze every silly detail of your application domain before you start coding. I encourage you to start hacking on you application without worrying about performance, because you know (if you follow the tips that follows and use your brain) that you can optimize it later. Knowing that you can rewrite that slow and memory-hungry Ruby script into a snappy C program (that don't use any heap memory at all) makes it so much easier to sleep at night. Believe me.

Ok, now let's get going!

Caring about planning

Planning for optimizability means that you make sure that:

• seams are placed to allow optimizing,
• system-wide concerns are encapsulated,
• it's possible to do design lowering later on,
• algorithmic decisions are delayed,
• intelligence can be moved from data to code to data,
• it's possible to move run-time aspects to compile-time,
• you know your tools.
  

We will now go through the first three of these bullets one by one and discuss them in detail. I will discuss the remaining four in a later post.

Seams and optimization borders

A seams in software design is a some-what vaguely defined concept. If two components of a software design are loosely coupled, we say that there is a seam between them. In other words, seams separate the design into parts that can vary independently of each other. Inheritance and dependency injection is often used to achieve this in object-oriented systems; function pointers play a similar role in procedural programming.

Since planning for optimizability means that we wish to make it easy to replace slow code with faster, we thus should place seams in the design such that they surround potentially slow code. This implies that design seams also are optimization borders; borders over which optimizations cannot easily be done without affecting other parts of the system.

This doesn't mean that it's impossible to optimize over optimization borders, just that other parts of the system will be affected by it. Thus, optimizing over optimization boarders is harder than optimizing within them. Let's take an example.

Assume that we iterate over a a large list in a performance critical part of our application:

void Foo::iterateOverLargeList(std::list<Thing> l) {
 for (int i = 0; i < l.size(); i++)
   l[i].doEasyCalculation();
}

Where should we place the design seams in this code? Or, more concretely, how do we use virtual functions to enable us to optimize this piece of code as much as possible without affecting the surrounding code? As I see it, we have the following alternatives:

1. make Thing::doEasyCalculation virtual,
   
2. make Foo::iterateOverLargeList virtual, or
   
3. encapsulate std::list<Thing> in a new class ThingList, move iterateOverLargeList to ThingList, and make it virtual.

Of these three alternatives 3) gives us the most freedom to increase the performance of the loop, because it allows us to do not only optimize how we iterate, but also change the underlying data structure. For instance, changing from std::list (a linked list) to std::vector (essentially a linear array) will make the access pattern linear, thus the (L1, L2, L3) cache can do a better job.

However, 3) is also the option that is the hardest to refactor to, because we have to change every part of the application that uses the std::list<Thing> list. So, if our application is large, this may involve a lot of refactoring work. Thus, if we choose 3) early in the development we don't have to refactor so much later. In other words, early understanding of the domain, the application's use-cases, and the application design, is important for us when designing for optimizability.

This example touches on another important tool when designing for optimizability: encapsulation. We will now discuss this in greater detail.

Concerning encapsulation

One of the worst kind of optimizations you can do is to optimize something that affect the entire system, e.g., the data flow. A global property like is a system-wide concern and having the system's performance depend on such property is very very bad. We need to encapsulate these global properties to make sure that there are design seams that enable us to optimize them when we need to.

This means that we need to design our code such that if we need to optimize something, that something is encapsulated into a single class (or few classes). This does not mean that we should encapsulate every little petty detail, though. It does mean, however, that we need to put a bit more effort in understanding our application, and then think about its design, data flow, etc. This is actually something we should do anyway even if we're not designing for optimizability, because understand the current design of the application make it easier to improve it.

When we got a firm understanding of the application and how it will be used, then we can identify the system-wide concerns, and only then can we start to encapsulate them to make them possible to optimize in the future.

Avoid memory copying

Reading data from memory is one of the most costly thing a modern CPU can do. If the data is in (L1) cache reading memory is fast (3-5 cycles or so), but if the data needs to be read from main memory it takes 100 times longer. That's right: 500 cycles to read a single bit of data. Writing data to main memory is also very costly operation, thus, copying is extremely costly operation.

With this in mind, it may be tempting to return a pointer to a shared data buffer instead of returning a copy of that data. For example:

class Thingie {
 int* stuff;
 // constructor that initializes 'stuff'.
 int* getStuff() { return stuff; }
};

This may look efficient since there is no memory copying going on. However, this design is horrible with regard to encapsulation. For instance, what if we need to modify the returned data in some other part of the program without affecting the Thing instance? In that case we need to make a copy and modify the copied data. But what if that copied data is returned just like the int* stuff above? And what if it needs to be modified in some other part of the program? Then we need to make another copy!

In situations like this, you need to take a step back and look at the data flow through your application and then try to optimize it to minimize the number of copies. This is not an easy task, so we really wish to avoid doing it, or at least make it easier for us to do. Is there some way to design our application (with little extra effort) from the start to make optimizations like this possible (without huge redesigns)?

In the example above, the system-wide concern is how the int* stuff data is passed around in the application and how to make sure that it is not copied unnecessarily.

Let's rewrite this example a bit into:

class Thingie2 {
 MyBuffer stuff;
 // constructor that initializes 'stuff'.
 MyBuffer getStuff() { return stuff; }
};

Since the data now is encapsulated in MyBuffer (and the entire application uses MyBuffer to refer the the data), we can now implement a lot of optimizations that affect the data-flow of the application. For instance, we can implement copy-on-write semantics and reference counting; this would be extremely hard to do without using MyBuffer to encapsulate the data-flow, which is a system-wide concern.

Design lowering

The term instruction lowering is an optimization technique used in compilers to generate more efficient code. For example, the C code a = 16 * b; can be implemented (in pseudo-assembler) by a multiplying operation A = MULT(16, B) but also by a simpler and faster shift operation A = SHL(4, B).

With the term design lowering I mean a similar concept: reimplementing the design in a simpler and faster language or platform, or using faster language constructs or data structures. Examples of such design lowering include:

• rewrite Python code in C,
• replace reflection-heavy Java code with simpler non-reflection code,
• using a simple array instead of a hash map or a linked list,
• using the memory layout of C structs as data format instead of XML in files, over network, etc
  
• use hardware acceleration like graphics cards.
  

Design lowering can easily increase the performance 10-100 times or even much more. Lets take an example of this.

A story of left and right

Imagine two small program left and right; left generates data and sends it to right, which sums the received data. Here is the first version of this pair of programs implemented in Python:

--- left.py ---
a = { "first":[1, 123], "middle":[1, 2, 456], "last":[1, 2, 3, 789] }
for i in xrange(0, 1000000):
 print a

--- right.py ---
import sys
total = 0
for line in sys.stdin:
 a = eval(line)
 total = total + sum(a["first"]) + sum(a["middle"]) + sum(a["last"])
 print total

Translating this Python code into C++ idioms this becomes:

--- left.cc ---
struct Data {
 Data() {
   first[0] = 1; first[1] = 123;
   middle[0] = 1; middle[1] = 2; middle[2] = 456;
   last[0] = 1; last[1] = 2; last[2] = 3; last[3] = 789;
 }
 int first[2];
 int middle[3];
 int last[4];
};

int main() {
 Data data;
 for (size_t a = 0; a < 1000000; a++) {
   for (size_t b = 0; b < sizeof(Data); b++)
     putchar(reinterpret_cast<char*>(&data)[b]);
 }
}

--- right.cc ---
struct Data {
 int first[2];
 int middle[3];
 int last[4];
};

int main() {
 Data data;
 long total = 0;
 for (size_t a = 0; a < 1000000; a++) {
   for (size_t b = 0; b < sizeof(Data); b++)
     reinterpret_cast<char*>(&data)[b] = getchar();
   total += data.first[0] + data.first[1] + data.middle[0] + data.middle[1] +
   data.middle[2] + data.last[0] + data.last[1] + data.last[2] + data.last[3];
 }
 printf("%ld\n", total);
}

And running them we get:

$ time ./left.py | ./right.py
1378000000

real 1m3.284s
user 1m14.057s
sys 0m0.256s
$ g++ -O2 right.cc -o right && g++ -O2 left.cc -o left && time ./left | ./right
1378000000

real 0m0.763s
user 0m1.292s
sys 0m0.088s
$ echo 74.067/1.292 | bc
57

Thus, the C++ version of these little programs are 57 times faster than the equivalent Python code. But this is not the neat thing... Let's get back to discussing design lowering.

Where to do design lowering

The neat thing is that you can implement left and right in Python without worrying about performance because they are easily design lowered to C++. I can say this because left and right comes in pair; if you redesign one side you also redesign the other. In other words, the details of the implementations are free to change in ways to make them execute more efficient. Thus, it's easy to do design lowering.

However, this would not be the case if left sent data to an unknown part via a predefined protocol. If this was so, then reimplementing left in C++ would probably be considerably harder. In other words, design lowering left would be less feasible. It would still be possible to do, though.

In summary, design lowering can be used as a plan for optimizability when the design can vary in ways that make it feasible to implement faster in another language (or data structure, etc). How the design can vary depends on many things, though, thus design lowering is not always easy to do without first refactoring the design.

As before, design lowering is not something you get for free, though: you need to ha good understanding of how/where the application will be used. When you understand your application, you're more likely to make correct assumptions of what parts of that can be design lowered and which parts that can't.

Conclusions

Just like optimizing, designing for optimizability requires us to understand the problem. We cannot escape it: to do something good, we need to know what "good" is, and to know what "good" is we need to understand the domain. Beware of over-analyzing though: concentrate on system-wide concerns that are likely to affect performance, and know how the design can vary to enable design lowering. Also, make sure that design seams don't hinder optimizability.

There are still much more to discuss on this topic, so stay tuned!

LLVM talks

A few very interesting presentations of LLVM from the LLVM developer's meeting 2009.