C++ has made a point of pride to thoroughly keep backward compatibility so that even the most antique and obsolete C++ source code may compile with just a few warnings (if any). This may be good to someone, but it makes for some funny constructs. E.g.:
[[nodiscard]] bool isEmpty() const override noexcept;Continue reading “What the Hec++k!”
In Italy, June is usually that period of the year when the weather becomes hot and the Italian C++ conference is held in Milan.
That’s why (the conference, not the weather) on Saturday 15th June, my friend and coworker Francesco and I went to the Milan Polytechnic. As for the other conferences I attended, I wrote quite a large volume of notes. Regretfully, as it happened for the other conferences, reorganizing my notes to make some sense of them and present them in an understandable and useful way is out of reach for my homeopathic level of free time. Also, usually you can find online the slides and videos of the talks well before my post. Therefore I decided for the short format I already used a couple of times in the past – a brief overview describing the conference for what you can’t see in the video e a brief summary of the talks I attended to.
Continue reading “it++ 2019”High performance computing made quite some progress lately. Maybe you heard about the reactive manifesto (you should if you are a reader of this blog ;-)), if not, it is an interesting readying … if nothing else as a conversational icebreaker.
Basically the manifesto preaches for systems more responsive and reliable by reacting to requests and dusting it into smaller requests processed by smaller and autonomous systems. This is nothing new, but quite far from traditional processing where single monolitic application took care of requests from reception to response.
There are several programming patterns useful for implementing a reactive system, channels (and streams which are quite the same thing) are one of them. A channel allows you to define a data flow with inputs, computation units, merge, split and collector of information coming in sequence into your system. Once the dataflow is defined, the data flows into it and results come out from the other end.
In this talk Felix Petriconi describes his implementation of channels for C++. Continue reading “++it18 – Channels are Useful, not only for Water”
One of the interesting features of Scala is the trait construct. In Java you have interfaces, a stripped down class that lets you express the requirements for classes that implement it to have a given … interface. Just to keep this from an OO perspective – in a certain system all MovableObjects, must expose an interface to retrieve, say, position, speed and acceleration:
interface MovableObject {
Point getPosition();
Vector getSpeed();
Vector getAcceleration();
}
class Rocket implements MovableObject {
public Point getPosition() {
return m_position;
}
public Vector getSpeed() {
return m_speed;
}
public Vector getAcceleration() {
return m_acceleration;
}
private Point m_position;
private Vector m_speed;
private Vector m_acceleration;
}
In Java the interface has no implementation part, just a list of methods that heir(s) must implement. This is a way to mitigate the lack of multiple inheritance in the Java language – a class can inherit at most from a single base class, but can implement as many interfaces as needed.
Scala traits are different in the sense they can provide an implementation part and fields. Now that I read the wikipedia definition, it looks like that traits are a bit more mathematically sound than interfaces (that’s what you can expect in functional language). Anyway, in plain English, traits are very close to C++ base classes (with some additions that is stuff for another blog post) in the sense that you can both require interface and provide common functionality.
This talk by Vittorio Romeo tackles a subject that could go ignored if your programming style doesn’t feel the need for parallelism (or if it does, pthreads satisfies such a need). Futures have been introduced in C++ with 2011 ISO standard, they hardly can be deemed as novelty.
I’ve been hit by futures some 4 years ago when confronted with a large code base in Scala that exploited futures and actors. So, at least to save my ego, there is nothing wrong if you don’t know what a future is :-).
In brief a future is a way to model a concurrent operation. The future can be either in execution or succeeded with a result value, or failed with an error. As user, you request an async operation and you get a future of the expected result. You can set callbacks on the completion, or combine several futures together either by chaining or by synchronize them. As you may imagine futures vastly simplify the handling of async computation by hiding the synchronization and communication details and by providing a common and clear usage pattern.
So, no excuse, if you need to do async programming, you’d better know this tool.
Continue reading “++it18 – Zero Allocation and No Type Erasure Futures”
This talk presented a very interesting object – the debugger the could go backward. But I don’t want to steal the scene, since Paolo Severini from Microsoft explained everything quite well in his talk. Here you can find the slides. Nedless to say errors are mine.
Everyone knows that debugging is twice as hard as writing the code in the first place. So if you’re as clever as you can be when you write it, how long will it take to debug it?
Debugging could be both hard and time consuming. Bugs can be difficult to replicate, especially those involving memory corruption, race conditions, or those that happen randomly or intermittently.
More often than not, it can be hard to spot the cause from logs or crash dumps. Also step by step debugging may require several restarts.
When doing step-by-step debugging is quite common something like:
step into, step over, step over, step over, step ov.. @!# I should have stepped into!.
The more information we can get from debugging tools the better. A tool that could tell “what just happened” would be very very useful.
Reverse debugging is such a tool. The idea is to record state by state everything happens in the CPU that’s executing the code and then the reverse debugger allows the programmer to play back and forth to spot the problem.
First ideas of such a tool dates back to 70s [NdM. I keep noticing that most of the computer science has been invented back in the 60-70s years. They just lacked the CPU resource for practical applications].
There are several tools for reverse debugging native code:
In this speech I will present TTD.
TTD is the reverse debugger for Windows. It is fully integrated with WinDbg. TTD can be used for all user-mode processes, for multithread and multicore
In order to use TTD, you need to set WinDbg to record data. This option needs to be explicitly enabled since has two main drawbacks – first all the program execution is slowed down and then quite a large amount of data is produced. You can expect anywhere between 5 and 50 Mb/s.
After recording has started, debugging works as expected, plus you can step back and forth and also continue backward or forward.
Take note that TTD works only in the latest versions fo Windows 10.
Interestingly you can perform queries on data collected during execution. Every line in the database matches an assembly instruction executed. So you can query how many times some instruction has been executed, how many exceptions have been thrown and so on.
[NdM: A working demo follows, unfortunately the time is quite constrained. A sample program to recursively compute the byte size of file in a directory tree is presented. The program just crashes when executed. So it is executed in WinDbg with TTD enabled. The execution halts as usual in the middle of exceptionland where nothing can be understood unless you speak advanced assembly. By pressing back a couple of time the debugger, going backwards, reaches the last function executed, the one who crashed. At this point Paolo places a breakpoint at the beginning of the function and – with a good implementation of the WOW effect – runs the program backwards. The execution back-reaches the function beginning and can be step forward to better understand why it crashed]
The presentation has been very interesting to show the tool capabilities. I tried an evaluation of UndoDB back in 2006, but it was as uncomfortable as it could be since it had to be used from command line. WinDbg+TTD is light years ahead in terms of usability.
Also I found very interesting the data query feature. Having the db of the program execution states allows the programmer to get a number of important statistics and events without the need of adding instrumentation or specific breakpoints.
The drawback is the execution time, but I guess that without a specific hardware support this is the price we have to pay still for some time in the future.
I really enjoyed this talk and not only because the speaker is an old friend on mine. I met Alberto back in the Ubisoft days and he was and still is one of the best C++ programmers I know and for sure the only friend who submitted a proposal to the standard committee that got approved. On the other hand I have no friends who submitted a proposal to the standard committee that was refused. So all my friends who submitted… well I’m digressing.
This talk was in Italian, that is welcome, but… weird. Usually all talks are in English even for Italian conference. Anyway here’s my summary, as usual errors and mistakes are mine. You can find Alberto’s slides here.
Alberto is a C++ enthusiast programmer since 1990. In this hour he’s going to talk about coroutines, their history and their relation with C++ and C++ Technical Specifications (TS).
A suspend and resume point is a point in the instruction sequence where the execution can be suspended, other stuff is executed and then the execution is resumed from there.
The most common form of suspend and resume point is a subroutine call. The execution is transferred from the caller to the callee and when the callee is done, the execution is transferred back to the caller.
Coroutines are a generalized form of suspension and resume points, that allow you to manage where you want to suspend, transfer control and resume, without be constrained in the caller/callee format.
When is this pattern useful?
Saturday 23 June 2018, was a sunny day in Milan. The ideal day to close yourself in Milano Bicocca University with many equally nerdy fellow programmers to attend the Italian C++ conference.
I discovered this conference, now in its 3rd edition, by chance, skimming through my neverending queue of unread tweets. A friend of mine and C++ guru twitted about this, so I bookmarked the website and as soon as this year edition was announced I was one of the first to register.
Also it is worth noting that this conference is free (as in beer), lasts a full day and includes the lunch. Talks are interesting and the conference is organized in two tracks with a single starting keynote and a single closing speech.
Although free, there are famous guests – last year Bartosz Milewski – quite a famous functional programmer and category algebra expert – held a speech about C++ and functional programming; this year the keynote was by Peter Sommerlad — C++ expert and member of the C++ ISO committee.
As usual my good intentions are to write and publish about each single talk I attended to.
Comparison with other conferences I attended, such as Scala Days, Scala Italy and Lambda World, is natural, although not always fair – ++it is completely free, while others come to a price (not light in case of Scala Days). I couldn’t fail to notice that age on average was older here than in the functional world. I won’t try an interpretation, but there could be more than one.
But, enough with babbling, let’s start with the keynote (based on my notes taken during the talk, errors or misunderstandings are just mine). You can find Peter Somerlad’s slides here. My comments at the end.
In this talk prof. Somerlad basically tells about his professional career, era by era, providing a short list of lessons learned for each period.
He started programming on a HP pocket calculator. This device had some 50 steps of program instruction and no graphics whatsoever. Also TI calculators were more widespread at time, so he had to translate programs from one programming language to another one. Although very similar (they were all stack based) each calculator had some peculiarities.
One program he developed for fun was a ballistic simulation – define a force and a direction to fire a bullet that has to avoid a wall and hit a target. As said, the pocket gizmo has no graphics, but years later, the very same idea was re-instantiated with some graphics and sounds to become… Angry Birds.
Lessons learned:
The first home computer was a ZX Spectrum – an 8bits computer with rubbery keyboard that connected to a TV set and used a cassette recorder as mass storage. Each key was decorated with several keyword beside the standard letter and symbol you normally find on a key. When pressing a key, depending on the context, one of such a keyword appeared on the screen – automatically typed for you. Context was determined by pressing two shift keys. The correct combo produced the required effect.
At school he worked on a mainframe, that had a curious assembly language. Notable one instruction, by exploiting an indirection bit, could be used to create an infinite loop. He and another one student tried the instruction and the mainframe froze. After few seconds an angry System Administrator entered the room, asking them to quit what they were doing.
Lessons learned:
In 1985 he landed his first job where he had to write a kind of customer management application on an Apple IIe. Now the software had to be professional and one requirement was that a cat could walk on the keyboard without causing a crash of the application. At times no commoditized databases were available and the few db engines available were very expensive. Since the Apple IIe had floppy discs, they may have to merge the two databases on two discs. The solution is to employ a merge sort, so he took the book from Niklaus Wirth “Data+Functions=Algorithms” and copied the merge sort implementation. Unfortunately the merge sort algorithm on the book was not stable, so Peter had to rewrite it from scratch.
Implementing the DB was quite a nightmare for many technical constraints – computer memory, mass storage size and computing power. The PC was an IBM with MS-DOS 1.x. This version had no directories. The Pascal language had no dynamic strings – they had to be statically sized.
As computer technology progressed he switched to a 386 compatible and turned to minix. During this period, Peter used PMate – an editor that sported macros, so that statement frames could be automatically generated.
Lessons learned:
1987 C and Unix courses
During this period Peter started teaching classes about C and Unix. In this classes shell, system programming and C were taught. Also he had to teach how to unlearn things to Cobol Programmers.
In 1988 he wrote some C guidelines – some are still actual today:
Lessons learned:
In 1989 Peter graduated. His master thesis was about a Modula-2 frontend for the Amsterdam Compiler Kit. In this year, also the first experiments with C++ using the Zortech C++ compiler. Writing code with this environment was quite intensive since any error in the code caused the whole machine to die. These were days before Protected Memory.
Lessons Learned:
Conscript at Bundenswehr (the German Army). Peter worked in the Computer Center, on systems for Radar detection and analysis. This was quite a travel back in time. The computer was an 18 bits with core memory, switches to input bits into the main memory and tape for mass storage.
Also Peter found another computer history milestone – the VT100 terminal. This terminal defined what ANSI escape codes are today.
The language used for programming the system was Jovial a mix of pascal and algol that was later replaced by Ada.
Lessons learned
In 1990 he landed a job at Siemens and started working in OO programming environment. Here he worked with Erich Gamma and Andrè Weinard [NdM: I tried to look up this but haven’t found anything relevant, maybe I got the name wrong].
The system Peter was working at was a framework for C++ written in C++, named OOPU, intended to replace TOROS_OBER a C macro based OO system used for car diagnostic computer. Unfortunately OOPU was killed by internal politics. The reason was that a single programming environment had to be used across several business units, and other business units didn’t want to pay for it.
In this context Peter wrote an object inspector, named OBER, that showed the relations among objects (this was before DDD was invented). From architectural point of view this application sported an UI separated from the backend.
In 1991 Peter wrote the first C++ guidelines, although some are still applicable, many are obsolete because the language evolved and things can be done better.
1993 was the year of the PLoP (Pattern Languages of Programming), this was still before the Gang of Four Book. In this conferences many opportunities wer to learn and understand architecture by trying to explain patterns.
1995 is the year of GoF. The most important design pattern are:
Do not use the singleton since C++ has no fixed sequence of initialization. If you need singleton, think about design first. If you need globals, just use globals [NdM, I understand that singleton has some problems, but I don’t see what is the alternatives – globals lacks of encapsulation].
In PLoPs and EuroPLoP, patterns had a community, where knowledge was shared, and it offerd an environment to try out ides. Discussing how stuff has to be written and why.
With other participant of the PLoP/EuroPLoP, Peter wrote the book POSA – Pattern Oriented Software Architecture. This is less famous than Design Pattern by the Gang of Four, but tackled the same matter.
Layers vs. Tiers -Layers are different level of abstraction , you can swap a layer implementation with another; while in tiers, functionality spread on multiple tiers, all the the same level of abstraction employing different technologies.
Using layers the team can specialize, while using tiers the team needs to know everything.
Good abstraction is a skill. Too many layers are a problem if layers do nothing but pass values.
Lessons learned:
In 1997 he moved to Switzerland to take the job of Erich Gamma, more precisely to work on a small server framework www.coast-project.org. While in charge Peter extended and improved the library and developed C++ Unit Test and test automation in doing so. He delivered high-quality solutions using Agile Software development and Wiki wiki web.
In these years Peter came to some interesting conclusions about high quality software. First high quality leads to better developer and better retainment of those developer. Unfortunately high quality leads to canceled maintenance contracts. If nothing goes wrong, the customer doesn’t see the point in paying for something that is never used.
Also, counter intuitively, high quality leads to less business opportunities. Since the contacts between customer and developer are minimal and usually don’t escalate to top level management. So the developer has less chance to meet and know the decision makers of the customer company.
Also CUTE – the test environment – has been created to replace macro-ish test.
Lessons learned
1998 and Agile Development. It became evident that following rituals is not adhering to values. You can do all the scrum rituals, but without committing to its values. When doing agile the most important things are:
Lessons Learned
Since 2004, Peter started teaching agile and continued teaching C++, concurrent and network programming. Also started working on IDE and tools, with the goal to eliminate bad software.
In 2005 many people thought that refactoring tool for C++ was impossible to write. Peter worked to prove them wrong, and Cevelop is the result.
[NdM: The talk ran out of time about now.]
Despite being sympathetic with Peter, since we shared so much in terms of computer history, programming and software engineering, I have to admit that his talk didn’t leave me with much. Also I fail to see the lack of some self-celebration. Maybe the most interesting thought was that quality software leads to less business. The rest of the talk was enjoyable, but I couldn’t see the connection with C++ or the future (or the history) of this language.
It is quite a feeling when you learn that your commit, a couple of months ago, broke the build is such a subtle way that it took so long to be detected. Possibly a more thorough testing and validation of the software would have caught it earlier, nonetheless it’s there and you and your coworkers are working hard to delimit the offending code and better understanding what caused the mess.
It turned out that some perfectly working code was taking too much time in one of the hot spot of our codebase. More precisely that code operated a conversion from an incoming data packet into a format suitable for data processing… about 2500 times a second on a modest hardware.
The code resulting in such poor performances that disrupted the device functionality seemed like a good idea in a very functional idiom –
@inline final val SAMPLES_PER_BLOCK = 6
@inline final val BLOCK_LENGTH = 8
@inline final val BLOCK_HEADER = 2
private def decodeFunctional(packet: Array[Byte]): Array[Int] = {
val lowSamples: Array[Int] = packet.drop(BLOCK_HEADER).map(x => x.toUint)
val msbs: Int = ((packet(0).toUint << 8) | packet(1).toUint) & 0xFFF
Array.tabulate(SAMPLES_PER_BLOCK)(
x => lowSamples(x) | (((msbs << 2 * x) & 0xC00) >> 2)
)
}
Basically the incoming packet holds six 10-bits samples in 8 bytes and some bits-shifting’n’pasting is needed to rebuild the six sample array for later processing.
Once found that the problem was here I decided rewrite the code to get rid of the functional idiom and take an imperative/iterative approach:
private def decodeImperative(packet: Array[Byte]): Array[Int] = {
// the code below is ugly, but it needs to be fast, don't touch,
// unless you know what you are doing
val samples: Array[Int] = new Array[Int](SAMPLES_PER_BLOCK)
val msbs: Int = ((packet(0).toUint << 8) | packet(1).toUint) & 0xFFF
var index: Int = 0
while (index < SAMPLES_PER_BLOCK) {
samples(index) = packet(BLOCK_HEADER + index).toUint |
(((msbs << 2 * index) & 0xC00) >> 2)
index += 1
}
samples
}
(the comment is a colorful note I decide to leave for the posterity :-)).
The idea is to get rid of the read-only constraint imposed by functional programming and save all the temporaries.
Optimizing is a subtle art, and even if intuition may guide you, only measurement can tell whether the optimization succeeded. So I set up some speed test (see “methodology” below) to objectively assess my results.
The speed improvement was astounding: more than 7 times faster (7.6 to be precise).
Energized by this victory I decided to switch to heavy optimization weapon. Not being sure if scala/java compiler optimization really does its optimization homework, I opted for manually unrolling the loop in the code:
private def decodeUnrolled0( packet: Array[Byte] ) : Array[Int] = {
val PACKET0 : Int = packet(0).toUint
val PACKET1 : Int = packet(1).toUint
Array[Int](
packet(BLOCK_HEADER+0).toUint | ((PACKET0 << 6) & 0x300),
packet(BLOCK_HEADER+1).toUint | ((PACKET0 << 8) & 0x300),
packet(BLOCK_HEADER+2).toUint | ((PACKET1 << 2) & 0x300),
packet(BLOCK_HEADER+3).toUint | ((PACKET1 << 4) & 0x300),
packet(BLOCK_HEADER+4).toUint | ((PACKET1 << 6) & 0x300),
packet(BLOCK_HEADER+5).toUint | ((PACKET1 << 8) & 0x300)
)
}
I was somewhat disappointed to learn that this version is only marginally faster (x1.6) than the original functional version. This didn’t really make sense to me, since the loop is completely unrolled and just trivial computation needed to be performed. So I decompiled the .class file:
public int[] CheckPerf$$decodeUnrolled0(byte[] packet)
{
int PACKET0 = UnsignedByteOps(packet[0]).toUint();
int PACKET1 = UnsignedByteOps(packet[1]).toUint();
return (int[])Array..MODULE$.apply(Predef..MODULE$.wrapIntArray(new int[] {
UnsignedByteOps(packet[2]).toUint() | PACKET0 << 6 & 0x300,
UnsignedByteOps(packet[3]).toUint() | PACKET0 << 8 & 0x300,
UnsignedByteOps(packet[4]).toUint() | PACKET1 << 2 & 0x300,
UnsignedByteOps(packet[5]).toUint() | PACKET1 << 4 & 0x300,
UnsignedByteOps(packet[6]).toUint() | PACKET1 << 6 & 0x300,
UnsignedByteOps(packet[7]).toUint() | PACKET1 << 8 & 0x300 }), ClassTag..MODULE$.Int());
}
Instead of the Java array I expected, some Scala internals are used to create the array and then converting it into the Java thing. Possibly this was the cause of the slowness. By looking around at the decompiled code I found another function that just used a Java int array. So I rewrote the unrolled version accordingly –
private def decodeUnrolled1( packet: Array[Byte] ) : Array[Int] = {
val samples = new Array[Int](SAMPLES_PER_BLOCK)
val PACKET0 = packet(0).toUint
val PACKET1 = packet(1).toUint
samples(0) = packet(BLOCK_HEADER + 0).toUint | ((PACKET0 << 6) & 0x300)
samples(1) = packet(BLOCK_HEADER + 1).toUint | ((PACKET0 << 8) & 0x300)
samples(2) = packet(BLOCK_HEADER + 2).toUint | ((PACKET1 << 2) & 0x300)
samples(3) = packet(BLOCK_HEADER + 3).toUint | ((PACKET1 << 4) & 0x300)
samples(4) = packet(BLOCK_HEADER + 4).toUint | ((PACKET1 << 6) & 0x300)
samples(5) = packet(BLOCK_HEADER + 5).toUint | ((PACKET1 << 8) & 0x300)
samples
}
The main difference is that here the array is first declared and allocated, then filled. In the above code the array was created, initialized and returned all in the same statement.
The speed improvement was good, but not much better than my imperative version: x7.77 times faster.
Then a colleague pointed out that I was using the “old” Scala 2.10 compiler and that I should try the latest 2.12 that benefits from a better interoperability with the underlying JVM 1.8.
In the following table you get the performance comparison:
Functional and unrolled0 are quite unsurprising, just what you expect from the next version of the compiler. Imperative approach yields quite a boost – 22 times faster than the functional version on the previous compiler. The shocking surprise is in the last column – the unrolled1 version of the code runs in excess of 1200 times faster than the original code!
Before jumping to conclusions, I performed the same tests on the code translated into C++. Here is the equivalent code:
typedef std::array<int,SAMPLES_PER_BLOCK> OutputArray;
typedef std::array<uint8_t,BLOCK_LENGTH> InputArray;
OutputArray decodeImperative( InputArray const& packet )
{
OutputArray samples;
uint32_t msbs = ((packet[0] << 8) | packet[1]) & 0xFFF;
for( int index=0; index< SAMPLES_PER_BLOCK; ++index )
{
samples[index] = packet[BLOCK_HEADER + index] |
(((msbs << 2 * index) & 0xC00) >> 2);
}
return samples;
}
OutputArray decodeFunctional( InputArray const& packet )
{
OutputArray lowSamples;
std::copy( packet.begin()+2, packet.end(), lowSamples.begin() );
uint32_t msbs = ((packet[0] << 8) | packet[1] ) & 0xFFF;
OutputArray samples;
int index=0;
OutputArray::const_iterator scan = lowSamples.begin();
std::generate(
samples.begin(),
samples.end(),
[&index,&scan,msbs]()
{
return *scan++ | (((msbs << 2*index++) & 0xC00) >> 2);
}
);
return samples;
}
OutputArray decodeUnrolled0( InputArray const& packet )
{
uint8_t PACKET0 = packet[0];
uint8_t PACKET1 = packet[1];
return OutputArray{
packet[BLOCK_HEADER+0] | ((PACKET0 << 6) & 0x300),
packet[BLOCK_HEADER+1] | ((PACKET0 << 8) & 0x300),
packet[BLOCK_HEADER+2] | ((PACKET1 << 2) & 0x300),
packet[BLOCK_HEADER+3] | ((PACKET1 << 4) & 0x300),
packet[BLOCK_HEADER+4] | ((PACKET1 << 6) & 0x300),
packet[BLOCK_HEADER+5] | ((PACKET1 << 8) & 0x300)
};
}
OutputArray decodeUnrolled1( InputArray const& packet )
{
OutputArray samples;
uint8_t PACKET0 = packet[0];
uint8_t PACKET1 = packet[1];
samples[0] = packet[BLOCK_HEADER+0] | ((PACKET0 << 6) & 0x300);
samples[1] = packet[BLOCK_HEADER+1] | ((PACKET0 << 8) & 0x300);
samples[2] = packet[BLOCK_HEADER+2] | ((PACKET1 << 2) & 0x300);
samples[3] = packet[BLOCK_HEADER+3] | ((PACKET1 << 4) & 0x300);
samples[4] = packet[BLOCK_HEADER+4] | ((PACKET1 << 6) & 0x300);
samples[5] = packet[BLOCK_HEADER+5] | ((PACKET1 << 8) & 0x300);
return samples;
}
I’m positive that functional purist will scream in horror at the sight of C++ functional version of the code, but it is the closest thing I could do using the standard library. Should you have any idea on how to improve the functional idiom using the standard library, let me know and I’ll update my test set.
Here are the C++ results –
This is mostly unsurprising but for the the lower right column which seems to indicate that Scala version is faster than C++. I guess that the reason is because I am using timing values that are rebased considering a standard price you pay for executing the code under test and that I don’t want to consider in my measurements (see below about the methodology I used). My interpretation is that the overhead in Scala hides the cost for a trivial expression so that the base time is much more close to the execution time than it is in C++.
Drawing general conclusions from a specific case is all but easy and trivial. Anyway according to data and to intuition, if you want to code using the functional idiom you pay a price that sometimes can be hefty. Paying an execution price, be it memory or cpu cycles, to raise the abstraction level is something that has been with computer programming from the dawn of this craft. I remember echoes from the ranting against high level languages in favor of assembly to avoid paying a 30% penalty. Now those times are gone, but facing a x4-x7 overhead can be something not always you can afford.
(Raising the level of abstraction by employing the functional idiom is something that is not undisputed, but I will spare this for another post).
Writing efficient code is not always intuitive. Constructs that seem innocent enough turns out to be compiled into time consuming lower level code. Peeking at compiled code is always useful. While languages as C++ are accompanied by detailed information about performances of various constructs and library components, Scala (as Java) lacks of these information. Also the huge amount of syntactic sugar employed by Scala does not simplify the life of the programmer that has to optimize code.
Eventually I cannot save to remarks the performances of interpreted languages (and even more so functional interpreted languages), despite of the progress in the JVM technologies, are still way behind native language ones. Even employing a functional approach C++ is some 80 times faster than Scala. You can argue that the C++ code above is not really functional, but there is nothing preventing me to use a mostly functional library for C++, with roughly the same performances of the functional version of the C++ code. It could be more tedious to write because of the lesser grade of syntactic glucose in C++, yet it is still functional.
I’m curious to see how scala-native compares, but this again is a topic for another post.
The last thought I want to report here is that you can make Scala code much faster, by avoiding the functional paradigm and peeking at compiled code. It is something you are not usually expected to do, but and improvement of 1200 times is a worth gain for the most executed code.
A very last thought for those who think that after all nowadays we have such powerful computers that analyzing performances and optimizing is a waste of time. I would like to remember that data processing requires energy and today energy production mostly relies on processes that produce CO2, something that the world doesn’t need. A data processing that runs just twice the speed of another will produce half of the carbon dioxide emissions. The environment benefits for switching to C and C++ should be clear to everyone.
A couple of words on methodology. The compilers were Scala 2.10.6, Scala 2.12.2 and GNU c++ 7.3.1. All the test were performed on Linux (Fedora Core 27) using the command /bin/time –format=”%U %S” command to print the User and Kernel times of the command and summing the two values together. These are times that CPU spent executing the process, not wall clock. The decoding function is executed 0x7FFFFFFF (some 2 billions) times per run and each program is tested for 12 runs. The fastest and the slowest run are discarded and the other are used to compute an average that is considered the measure of the test.
The base time is computed by employing a dummy decoder that does nothing but filling the result array with constants.
This setup should take in account the JIT effect and measure just the code I listed above.
Recently I had to spend some time trying to adapt my imperative/OO background to a piece of code I need to write in functional paradigm. The problem is quite simple and can be described briefly. You have a collection of pairs containing an id and a quantity. When a new pair arrives you have to add the quantity to the pair with the same id in the collection, if exists, otherwise you have to add the pair to the collection.
Functional programming is based on three principles (as seen from an OO programmer) – variables are never modified once assigned, no side-effects (at least, avoid them as much as possible), no loops – work on collections with collective functions. Well maybe I missed something like monad composition, but that’s enough for this post.
Thanks to a coworker I wrote my Scala code that follows all the aforementioned principles and is quite elegant as well. It relies on the “partition” function that transforms (in a functional fashion) a collection into two collections containing the elements of the first one partitioned according to a given criteria. The criteria is the equality of the id so that I find the element with the same id if it exists, or just an empty collection if it doesn’t.
Here’s the code:
type Collection = List[Data]
def merge( collection : Collection, data : Data ) : Collection = {
val result = collection.partition( _.id == data.id )
result match {
case (List(),other) =>
data :: other
case (List(old),other) =>
Data( data.id, data.quantity+old.quantity ) :: other
}
}
Yes, I could have written more concisely, but that would have been too much write-only for me to be comfortable with.
Once the pleasant feeling of elegance wore off a bit I wondered what is the cost of this approach. Each time you invoke merge the collection is rebuilt and, unless the compile optimizer be very clever, also each list item is cloned and the old one goes to garbage recycling.
Partitioning scans and rebuild, but since I’m using an immutable collection, also adding an item to an existing list causes a new list to be generated.
Performance matters in some 20% of your code, so it could acceptable to sacrifice performance in order to get a higher abstraction level and thus a higher coding speed. But then I wonder what about premature pessimization? Premature pessimization, at least in context where I read the them, means the widespread adoption of idioms that lead to worse performances (the case was for C++ use of pre or post increment operator). Premature pessimization may cause the application to run generally slower and makes more difficult to spot and optimize the cause.
This triggered the question – how is language idiomatic approach impacts on performances?
To answer the question I started coding the same problem in different languages.
I started from my language of choice – C++. In this language it is likely you approach a similar problem by using std::vector. This is the preferred collection and the recommended one. Source is like this:
typedef std::vector<Data> Collection;
void
merge( Collection& collection, Data data )
{
auto iter = std::find_if( collection.begin(),
collection.end(),
[&]( Data const& probe ){ return probe.id == data.id; } );
if( iter == collection.end() )
{
collection.push_back( data );
}
else
{
iter->quantity += data.quantity;
}
}
Code is slightly longer (consider that in C++ I prefer opening brace on a line alone, while in Scala “they” forced me to have opening braces at the end of the statement line). Having mutable collections doesn’t require to warp your mind around data to find which aggregate function could transform your input data into the desired output – just find what you are looking for and change it. Seems simpler to explain to a child.
Then I turned to Java. I’m not so fond of this language, but it is quite popular and has a comprehensive set of standard libraries that really allow you to tackle every problem confidently. Not sure what a Java programmer would consider idiomatic, so I staid traditional and went for a generic List. The code follows:
static class Data
{
public int id;
public int quantity;
}
private static void merge( List<Data> collection, Data data )
{
ListIterator<Data> iterator = collection.listIterator();
while( iterator.hasNext() )
{
Data probe = iterator.next();
if( probe.id == data.id )
{
probe.quantity += data.quantity;
return;
}
}
collection.add( 0, data );
}
I’m not sure why the inner class Data needs to be declared static, but it seems that otherwise the instance has a reference to the outer class instance. Anyway – code is decidedly more complex. There is no function similar to C++ find_if nor to Scala partition. The loop is simple, but it offers some chances to add bugs to your code. Anyway explaining the code is straightforward once the iterator concept is clear.
Eventually I wrote a version in C. This language is hampered by the lack of basic standard library – beside some functions on strings and files you have nothing. This could have been fine in the 70s, but today is a serious problem. Yes there are non-standard libraries providing all the needed functionalities, you have plenty of them, gazillions of them, all incompatible. Once you chose one you are locked in… Well clearly C shows the signs of age. So I write my own single linked list implementation:
struct Data
{
int id;
int quantity;
};
struct ListNode
{
struct ListNode* next;
struct Data data;
};
typedef struct ListNode* List;
void merge( List* list, struct Data data )
{
for( struct ListNode* scan = *list; scan != NULL; scan = scan->next )
{
if( scan->data.id == data.id )
{
scan->data.quantity += data.quantity;
return;
}
}
pushFront( list, data );
}
Note that once cleaned of braces, merge function is shorter in C than in Java! This is a hint that Java is possibly verbose.
I just wrote here the merge function. The rest of the sources is not relevant for this post, but it basically consists in parsing the command line (string to int conversion), getting some random numbers and getting the time. The simplest frameworks for this operation are those based on the JVM. The most complex is C++ – it allows a wide range of configuration (random and time), but I had to look up on internet how to do it and… I am afraid I wouldn’t know how to exploit the greater range of options. Well, in my career as a programmer (some 30+ years counting since when I started coding) I think I never had the need to use a specific random number generator, or some clock different from a “SystemTimeMillis” or Wall Clock Time. I don’t mean that because of this no one should ask for greater flexibility, but that I find daunting that every programmer should pay this price because there is case for using a non default RNG.
Anyway back to my test. In the following table I reported the results.
| C++ | Scala | Java | C | C++ | |
| vector | list | ||||
| time (ms) | 170,75 | 11562,45 | 2230,05 | 718,75 | 710,9 |
| lines | 81 | 35 | 69 | 112 | 81 |
Times have been taken performing 100000 insertions with max id 10000. The test has been repeated 20 times and the results have been averaged in the table. The difference in timing between C++ and Scala is dramatic – with the first faster about 70 times the latter. Wildly extrapolating you can say that if you code in C++ you need 1/70 of the hardware you need to run Scala… there’s no surprise (still guessing wildly) that IBM backs this one.
Java is about 5 times faster than Scala. I’m told this is more or less expected and possibly it is something you may be willing to pay for higher level.
In the last column I reported the results for a version of the C++ code employing std::list for a more fair comparison (all the other implementations use a list after all). What I didn’t expected was that C++ is faster (even if slightly) than C despite using the same data structure. It is likely because of some template magic.
The other interesting value I wrote in the table is the number of lines (total, not just the merge function) of each implementation. From my studies (that now are quite aged) I remember that some researches reported that the speed of software development (writing, testing and debugging), stated as lines of code per unit of time, is the same regardless of the language. I’m starting having some doubt because my productivity in Scala is quite low if compared with other languages, but … ipse dixit.
Let’s say that you spend 1 for the Scala program, then you would pay 2.31 for C++, 1.97 for Java and 3.20 for C.
Wildly extrapolating again you could draw a formula to decide whether it is better to code in C++ or in Scala. Be H the cost of the CPU and hardware to comfortably run the C++ program. Be C the cost of writing the program in Scala. So the total cost of the project is:
(C++) H+C×2.31
(Scala) 68×H+C
(C++) > (Scala) ⇒ H+C×2.31 > 68×H+C ⇒ C×1.31 >67×H ⇒ C > 51.14×H
That is, you’d better using Scala when the cost of the hardware you want to use will not exceed the cost of Scala development by a factor of 50. If hardware is going to cost more, then you’d better use C++.
Beside of being a wild guess, this also assumes that there is no hardware constraint and that you can easily scale the hardware of the platform.
(Thanks to Andrea for pointing out my mistake in inequality)