Moore’s law was a socially constructed project

Moore’s law was a socially constructed project that depended on the coordinated actions of many independent companies and groups of individuals to last for as long it did.

All products evolve, but what was it about Moore’s law that enabled microelectronics to evolve so much faster and for longer than most other products?

Moore’s observation, made in 1965 based on four data points, was that the number of components contained in a fabricated silicon device doubles every year. The paper didn’t make this claim in words, but a line fitted to four yearly data points (starting in 1962) suggested this behavior continuing into the mid-1970s. The introduction of IBM’s Personal Computer, in 1981 containing Intel’s 8088 processor, led to interested parties coming together to create a hugely profitable ecosystem that depended on the continuance of Moore’s law.

The plot below shows Moore’s four points (red) and fitted regression model (green line). In practice, since 1970, fitting a regression model (purple line) to the number of transistors in various microprocessors (blue/green, data from Wikipedia), finds that the number of transistors doubled every two years (code+data):

Transistors contained in a device over time, plus Moore's original four data-points.

In the early days, designing a device was mostly a manual operation; that is, the circuit design and logic design down to the transistor level were hand-drawn. This meant that creating a device containing twice as many transistors required twice as many engineers. At some point the doubling process either becomes uneconomic or it takes forever to get anything done because of the coordination effort.

The problem of needing an exponentially-growing number of engineers was solved by creating electronic design automation tools (EDA), starting in the 1980s, with successive generations of tools handling ever higher levels of abstraction, and human designers focusing on the upper levels.

The use of EDA provides a benefit to manufacturers (who can design differentiated products) and to customers (e.g., products containing more functionality).

If EDA had not solved the problem of exponential growth in engineers, Moore’s law would have maxed-out in the early 1980s, with around 150K transistors per device. However, this would not have stopped the ongoing shrinking of transistors; two economic factors independently incentivize the creation of ever smaller transistors.

When wafer fabrication technology improvements make it possible to double the number of transistors on a silicon wafer, then around twice as many devices can be produced (assuming unchanged number of transistors per device, and other technical details). The wafer fabrication cost is greater (second row in table below), but a lot less than twice as much, so the manufacturing cost per device is much lower (third row in table).

The doubling of transistors primarily provides a manufacturer benefit.

The following table gives estimates for various chip foundry economic factors, in dollars (taken from the report: AI Chips: What They Are and Why They Matter). Node, expressed in nanometers, used to directly correspond to the length of a particular feature created during the fabrication process; these days it does not correspond to the size of any specific feature and is essentially just a name applied to a particular generation of chips.

Node (nm)                       90      65     40     28      20    16/12     10       7       5
Foundry sale price per wafer  1,650   1,937  2,274  2,891   3,677   3,984   5,992   9,346  16,988
Foundry sale price per chip   2,433   1,428    713    453     399     331     274     233     238
Mass production year          2004    2006   2009   2011    2014    2015    2017    2018   2020
Quarter                        Q4      Q4     Q1     Q4      Q3      Q3      Q2      Q3     Q1
Capital investment per wafer  4,649   5,456  6,404  8,144  10,356  11,220  13,169  14,267  16,746
processed per year
Capital consumed per wafer      411     483    567    721     917     993   1,494   2,330   4,235
processed in 2020
Other costs and markup        1,293   1,454  1,707  2,171   2,760   2,990   4,498   7,016  12,753
per wafer

The second economic factor incentivizing the creation of smaller transistors is Dennard scaling, a rarely heard technical term named after the first author of a 1974 paper showing that transistor power consumption scaled with area (for very small transistors). Halving the area occupied by a transistor, halves the power consumed, at the same frequency.

The maximum clock-frequency of a microprocessor is limited by the amount of heat it can dissipate; the heat produced is proportional to the power consumed, which is approximately proportional to the clock-frequency. Instead of a device having smaller transistors consume less power, they could consume the same power at double the frequency.

Dennard scaling primarily provides a customer benefit.

Figuring out how to further shrink the size of transistors requires an investment in research, followed by designing/(building or purchasing) new equipment. Why would a company, who had invested in researching and building their current manufacturing capability, be willing to invest in making it obsolete?

The fear of losing market share is a commercial imperative experienced by all leading companies. In the microprocessor market, the first company to halve the size of a transistor would be able to produce twice as many microprocessors (at a lower cost) running twice as fast as the existing products. They could (and did) charge more for the latest, faster product, even though it cost them less than the previous version to manufacture.

Building cheaper, faster products is a means to an end; that end is receiving a decent return on the investment made. How large is the market for new microprocessors and how large an investment is required to build the next generation of products?

Rock’s law says that the cost of a chip fabrication plant doubles every four years (the per wafer price in the table above is increasing at a slower rate). Gambling hundreds of millions of dollars, later billions of dollars, on a next generation fabrication plant has always been a high risk/high reward investment.

The sales of microprocessors are dependent on the sale of computers that contain them, and people buy computers to enable them to use software. Microprocessor manufacturers thus have to both convince computer manufacturers to use their chip (without breaking antitrust laws) and convince software companies to create products that run on a particular processor.

The introduction of the IBM PC kick-started the personal computer market, with Wintel (the partnership between Microsoft and Intel) dominating software developer and end-user mindshare of the PC compatible market (in no small part due to the billions these two companies spent on advertising).

An effective technique for increasing the volume of microprocessors sold is to shorten the usable lifetime of the computer potential customers currently own. Customers buy computers to run software, and when new versions of software can only effectively be used in a computer containing more memory or on a new microprocessor which supports functionality not supported by earlier processors, then a new computer is needed. By obsoleting older products soon after newer products become available, companies are able to evolve an existing customer base to one where the new product is looked upon as the norm. Customers are force marched into the future.

The plot below shows sales volume, in gigabytes, of various sized DRAM chips over time. The simple story of exponential growth in sales volume (plus signs) hides the more complicated story of the rise and fall of succeeding generations of memory chips (code+data):

Sales volume, in gigabytes, of various sized DRAM chips over time.

The Red Queens had a simple task, keep buying the latest products. The activities of the companies supplying the specialist equipment needed to build a chip fabrication plant has to be coordinated, a role filled by the International Technology Roadmap for Semiconductors (ITRS). The annual ITRS reports contain detailed specifications of the expected performance of the subsystems involved in the fabrication process.

Moore’s law is now dead, in that transistor doubling now takes longer than two years. Would transistor doubling time have taken longer than two years, or slowed down earlier, if:

  • the ecosystem had not been dominated by two symbiotic companies, or did network effects make it inevitable that there would be two symbiotic companies,
  • the Internet had happened at a different time,
  • if software applications had quickly reached a good enough state,
  • if cloud computing had gone mainstream much earlier.

Where are the industrial strength R compilers?

Why don’t compiler projects for the R language make it into production use? The few that have been written have remained individual experimental products, e.g., RLLVMCompile.

Most popular languages attract many compiler implementations. I’m not saying that any of these implementations have more than a handful of users, that they implement the full language (a full implementation is not common), or that they fulfil any need other than their implementers desire to build something.

A commonly heard reason for the lack of production R compilers is that it is not worth the time and effort, because most of an R program’s time is spent in the library code which is written in a compiled language (e.g., C or Fortran). The fact that it is probably not worth the time and effort has not stopped people writing compilers for other languages, but then I think that the kind of people who use R tend not to be the kind of people who want to spend their time writing compilers. On the whole, they are the kind of people who are into statistics and data analysis.

Is it true that that most R programs spend most of their time executing library code? It’s certainly true for me. But I have noticed that a lot of the library functions executed by my code are written in R. Also, if somebody uses R for all their programming needs (it might be the only language they know), then their code might not be heavily library dependent.

I was surprised to read about Tierney’s byte code compiler, because his implementation is how I thought the R-core’s existing implementation worked (it does now). The internals of R is based on 1980s textbook functional techniques, and like many book implementations of the day, performance is dependent on the escape hatch of compiled code. R’s implementers wisely spent their time addressing user concerns, which revolved around statistics and visual presentation, i.e., not internal implementation technicalities.

Building an R compiler is easy, the much harder and time-consuming part is the runtime system.

Threaded code is a quick and simple approach to compiler implementation. R source gets mapped to a sequence of C function calls, with these functions proving a wrapper to library functions implementing the appropriate basic functionality, e.g., add two vectors. This approach has been the subject of at least one Master’s thesis. Thesis implementations rarely reach production use because those involved significantly underestimate the work that remains to be done, which is usually a lot more than the original implementation.

A simple threaded code approach provides a base for subsequent optimization, with the base having a similar performance to an interpreter. Optimizing requires figuring out details of the operations performed and replacing generic function calls with ones designed to be fast for specific cases, or even better replacing calls with inline code, e.g., adding short vectors of integers. There is a lot of existing work for scripting languages and a few PhD thesis researching R (e.g., Wang). The key technique is static analysis of R source.

Jan Vitek is running what appears to be the most active R compiler research group, at the moment e.g., the Ř project. Research can be good for uncovering language usage and trying out different techniques, but it is not intended to produce industry strength code. Lots of the fancy optimizations in early versions of the gcc C compiler started life as a PhD thesis, with the respective individual sometimes going on to spend a few years creating a production quality version for the released compiler.

The essential ingredient for building a production compiler is persistence. There are an awful lot of details that need to be sorted out (this is why research project code does not directly translate to production code, they ignore ‘minor’ details in order to concentrate on the ‘interesting’ research problem). Is there a small group of people currently beavering away on a production quality compiler for R? If there is, I can understand being discrete, on long-term projects it can be very annoying to have people regularly asking when the software is going to be released.

To have a life, once released, a production compiler needs to attract users, who are often loyal to their current compiler (because they know that their code works for this compiler); there needs to be a substantial benefit to entice people to switch. The benefit of compiling R to machine code, rather than interpreting, is performance. What performance improvement is needed to attract a viable community of users (there is always a tiny subset of users who will pay lots for even small performance improvements)?

My R code is rarely cpu bound, so I am not in the target audience, no matter what the speed-up. I don’t have any insight in the performance problems experienced by the R community, and have no idea whether a factor of two, five, ten or more would be enough.

Scientific management of software production

When Frederick Taylor investigated the performance of workers in various industries, at the start of the 1900’s, he found that workers organise their work to suit themselves; workers were capable of producing significantly more than they routinely produced. This was hardly news. What made Taylor’s work different was that having discovered the huge difference between actual worker output and what he calculated could be achieved in practice, he was able to change work practices to achieve close to what he had calculated to be possible. Changing work practices took several years, and the workers did everything they could to resist it (Taylor’s The principles of scientific management is an honest and revealing account of his struggles).

Significantly increasing worker output pushed company profits through the roof, and managers everywhere wanted a piece of the action; scientific management took off. Note: scientific management is not a science of work, it is a science of the management of other people’s work.

The scientific management approach has been successfully applied to production where most of the work can be reduced to purely manual activities (i.e., requiring little thinking by those who performed them). The essence of the approach is to break down tasks into the smallest number of component parts, to simplify these components so they can be performed by less skilled workers, and to rearrange tasks in a way that gives management control over the production process. Deskilling tasks increases the size of the pool of potential workers, decreasing labor costs and increasing the interchangeability of workers.

Given the almost universal use of this management technique, it is to be expected that managers will attempt to apply it to the production of software. The software factory was tried, but did not take-off. The use of chief programmer teams had its origins in the scarcity of skilled staff; the idea is that somebody who knows what they were doing divides up the work into chunks that can be implemented by less skilled staff. This approach is essentially the early stages of scientific management, but it did not gain traction (see “Programmers and Managers: The Routinization of Computer Programming in the United States” by Kraft).

The production of software is different in that once the first copy has been created, the cost of reproduction is virtually zero. The human effort invested in creating software systems is primarily cognitive. The division between management and workers is along the lines of what they think about, not between thinking and physical effort.

Software systems can be broken down into simpler components (assuming all the requirements are known), but can the implementation of these components be simplified such that they can be implemented by less skilled developers? The process of simplification is practical when designing a system for repetitive reproduction (e.g., making the same widget again and again), but the first implementation of anything is unlikely to be simple (and only one implementation is needed for software).

If it is not possible to break down the implementation such that most of the work is easy to do, can we at least hire the most productive developers?

How productive are different developers? Programmer productivity has been a hot topic since people started writing software, but almost no effective research has been done.

I have no idea how to measure programmer productivity, but I do have some ideas about how to measure their performance (a high performance programmer can have zero productivity by writing programs, faster than anybody else, that don’t do anything useful, from the client’s perspective).

When the same task is repeatedly performed by different people it is possible to obtain some measure of average/minimum/maximum individual performance.

Task performance improves with practice, and an individual’s initial task performance will depend on their prior experience. Measuring performance based on a single implementation of a task provides some indication of minimum performance. To obtain information on an individual’s maximum performance they need to be measured over multiple performances of the same task (and of course working in a team affects performance).

Should high performance programmers be paid more than low performance programmers (ignoring the issue of productivity)? I am in favour of doing this.

What about productivity payments, e.g., piece work?

This question is a minefield of issues. Manual workers have been repeatedly found to set informal quotas amongst themselves, i.e., setting a maximum on the amount they will produce during a shift (see “Money and Motivation: An Analysis of Incentives in Industry” by William Whyte). Thankfully, I don’t think I will be in a position to have to address this issue anytime soon (i.e., I don’t see a reliable measure of programmer productivity being discovered in the foreseeable future).

Performance variation in 2,386 ‘identical’ processors

Every microprocessor is different, random variations in the manufacturing process result in transistors, and the connections between them, being fabricated with more/less atoms. An atom here and there makes very little difference when components are built from millions, or even thousands, of atoms. The width of the connections between transistors in modern devices might only be a dozen or so atoms, and an atom here and there can have a noticeable impact.

How does an atom here and there affect performance? Don’t all processors, of the same product, clocked at the same frequency deliver the same performance?

Yes they do, an atom here or there does not cause a processor to execute more/less instructions at a given frequency. But an atom here and there changes the thermal characteristics of processors, i.e., causes them to heat up faster/slower. High performance processors will reduce their operating frequency, or voltage, to prevent self-destruction (by overheating).

Processors operating within the same maximum power budget (say 65 Watts) may execute more/less instructions per second because they have slowed themselves down.

Some years ago I spotted a great example of ‘identical’ processor performance variation, and the author of the example, Barry Rountree, kindly sent me the data. In the weeks before Christmas I finally got around to including the data in my evidence-based software engineering book. Unfortunately I could not figure out what was what in the data (relearning an important lesson: make sure to understand the data as soon as it arrives), thankfully Barry came to the rescue and spent some time doing software archeology to figure out the data.

The original plots showed frequency/time data of 2,386 Intel Sandy Bridge XEON processors (in a high performance computer at the Lawrence Livermore National Laboratory) executing the EP benchmark (the data also includes measurements from the MG benchmark, part of the NAS Parallel benchmark) at various maximum power limits (see plot at end of post, which is normalised based on performance at 115 Watts). The plot below shows frequency/time for a maximum power of 65 Watts, along with violin plots showing the spread of processors running at a given frequency and taking a given number of seconds (my code, code+data on Barry’s github repo):

Frequency vs Time at 65 Watts

The expected frequency/time behavior is for processors to lie along a straight line running from top left to bottom right, which is roughly what happens here. I imagine (waving my software arms about) the variation in behavior comes from interactions with the other hardware devices each processor is connected to (e.g., memory, which presumably have their own temperature characteristics). Memory performance can have a big impact on benchmark performance. Some of the other maximum power limits have very different, and benchmark, measurements have very different characteristics (see below).

More details and analysis in the paper: An empirical survey of performance and energy efficiency variation on Intel processors.

Intel’s Sandy Bridge is now around seven years old, and the number of atoms used to fabricate transistors and their connectors has shrunk and shrunk. An atom here and there is likely to produce even more variation in the performance of today’s processors.

A previous post discussed the impact of a variety of random variations on program performance.

Below is a png version of the original plot I saw:

Frequency vs Time at all power levels

Modular vs. monolithic programs: a big performance difference

For a long time now I have been telling people that no experiment has found a situation where the treatment (e.g., use of a technique or tool) produces a performance difference that is larger than the performance difference between the subjects.

The usual results are that differences between people is the source of the largest performance difference, successive runs are the next largest (i.e., people get better with practice), and the smallest performance difference occurs between using/not using the technique or tool.

This is rather disheartening news.

While rummaging through a pile of books I had not looked at in many years, I (re)discovered the paper “An empirical study of the effects of modularity on program modifiability” by Korson and Vaishnavi, in “Empirical Studies of Programmers” (the first one in the series). It’s based on Korson’s 1988 PhD thesis, with the same title.

There were four experiments, involving seven people from industry and nine students, each involving modifying a 900(ish)-line program in some way. There were two versions of each program, they differed in that one was written in a modular form, while the other was monolithic. Subjects were permuted between various combinations of program version/problem, but all problems were solved in the same order.

The performance data (time to complete the task) was published in the paper, so I fitted various regressions models to it (code+data). There is enough information in the data to separate out the effects of modular/monolithic, kind of problem and subject differences. Because all subjects solved problems in the same order, it is not possible to extract the impact of learning on performance.

The modular/monolithic performance difference was around twice as large as the difference between subjects (removing two very poorly performing subjects reduces the difference to 1.5). I’m going to have to change my slides.

Would the performance difference have been so large if all the subjects had been experienced developers? There is not a lot of well written modular code out there, and so experienced developers get lots of practice with spaghetti code. But, even if the performance difference is of the same order as the difference between developers, that is still a very worthwhile difference.

Now there are lots of ways to write a program in modular form, and we don’t know what kind of job Korson did in creating, or locating, his modular programs.

There are also lots of ways of writing a monolithic program, some of them might be easy to modify, others a tangled mess. Were these programs intentionally written as spaghetti code, or was some effort put into making them easy to modify?

The good news from the Korson study is that there appears to be a technique that delivers larger performance improvements than the difference between people (replication needed). We can quibble over how modular a modular program needs to be, and how spaghetti-like a monolithic program has to be.

Performance of Java 2D drawing operations (part 3: image opacity)

Series: operations, images, opacity

Not because I was enjoying it, I seemed compelled to continue my quest to understand the performance of various Java 2D drawing operations. I’m hoping to make my game Rabbit Escape faster, especially on the Raspberry Pi, so you may see another post sometime actually trying this stuff out on a Pi.

But for now, here are the results of my investigation into how different patterns of opacity in images affects rendering performance.

You can find the code here: gitlab.com/andybalaam/java-2d-performance.

Results

  • Images with partially-opaque pixels are no slower than those with fully-opaque pixels
  • Large transparent areas in images are drawn quite quickly, but transparent pixels mixed with non-transparent are slow

Advice

  • Still avoid any transparency whenever possible
  • It’s relatively OK to use large transparent areas on images (e.g. a fixed-size animation where a character moves through the image)
  • Don’t bother restricting pixels to be either fully transparent or fully opaque – partially-opaque is fine

Opacity patterns in images

Non-transparent images drew at 76 FPS, and transparent ones dropped to 45 FPS.

I went further into investigating transparency by creating images that were:

  • All pixels 50% opacity (34 FPS)
  • Half pixels 0% opacity, half 100%, mixed up (34 FPS)
  • Double the size of the original image, but the extra area is fully transparent, and the original area is non-transparent (41 FPS)

I concluded that partial-opacity is not important to performance compared with full-opacity, but that large areas of transparency are relatively fast compared with images with complex patterns of transparency and opacity.

Numbers

Transparency and opacity

Test FPS
large nothing 90
large images20 largeimages 76
large images20 largeimages transparentimages 45
large images20 largeimages transparent50pcimages 34
large images20 largeimages transparent0pc100pcimages 34
large images20 largeimages transparentareaimages 41

Feedback please

Please do get back to me with tips about how to improve the performance of my experimental code.

Feel free to log issues, make merge requests or add comments to the blog post.

Performance of Java 2D drawing operations (part 2: image issues)

Series: operations, images

In my previous post I examined the performance of various drawing operations in Java 2D rendering. Here I look at some specifics around rendering images, with an eye to finding optimisations I can apply to my game Rabbit Escape.

You can find the code here: gitlab.com/andybalaam/java-2d-performance.

Results

  • Drawing images with transparent sections is very slow
  • Drawing one large image is slower than drawing many small images covering the same area(!)
  • Drawing images outside the screen is slower than not drawing them at all (but faster than drawing them onto a visible area)

Advice

  • Avoid transparent images where possible
  • Don’t bother pre-rendering your background tiles onto a single image
  • Don’t draw images that are off-screen

Images with transparency

All the images I used were PNG files with a transparency layer, but in most of my experiments there were no transparent pixels. When I used images with transparent pixels the frame rate was much slower, dropping from 78 to 46 FPS. So using images with transparent pixels causes a significant performance hit.

I’d be grateful if someone who knows more about it can recommend how to improve my program to reduce this impact – I suspect there may be tricks I can do around setComposite or setRenderingHint or enabling/encouraging hardware acceleration.

Composite images

I assumed that drawing a single image would be much faster than covering the same area of the screen by drawing lots of small images. In fact, the result was the opposite: drawing lots of small images was much faster than drawing a single image covering the same area.

The code for a single image is:

g2d.drawImage(
    singleLargeImage,
    10,
    10,
    null
)

and for the small images it is:

for (y in 0 until 40)
{
    for (x in 0 until 60)
    {
        g2d.drawImage(
            compositeImages[(y*20 + x) % compositeImages.size],
            10 + (20 * x),
            10 + (20 * y),
            null
        )
    }
}

The single large image was rendered at 74 FPS, whereas covering the same area using repeated copies of 100 images was rendered at 80 FPS. I ran this test several times because I found the result surprising, and it was consistent every time.

I have to assume some caching (possibly via accelerated graphics) of the small images is the explanation.

Drawing images off the side of the screen

Drawing images off the side of the screen was faster than drawing them in a visible area, but slower than not drawing them at all. I tested this by adding 10,000 to the x and y positions of the images being drawn (I also tested subtracting 10,000 with similar results). Not drawing any images ran at 93 FPS, drawing images on-screen at 80 FPS, and drawing them off-screen only 83 FPS, meaning drawing images off the side takes significant time.

Advice: check whether images are on-screen, and avoid drawing them if not.

Numbers

Transparency

Test FPS
large nothing 95
large images20 largeimages 78
large images20 largeimages transparentimages 46

Composite images

(Lots of small images covering an area, or a single larger image.)

Test FPS
large nothing 87
large largesingleimage 74
large compositeimage 80

Offscreen images

Test FPS
large nothing 93
large images20 largeimages 80
large images20 largeimages offscreenimages 83

Feedback please

Please do get back to me with tips about how to improve the performance of my experimental code.

Feel free to log issues, make merge requests or add comments to the blog post.

Performance of Java 2D drawing operations

I want to remodel the desktop UI of my game Rabbit Escape to be more
convenient and nicer looking, so I took a new look at game-loop-style graphics rendering onto a canvas in a Java 2D (Swing) UI.

Specifically, how fast can it be, and what pitfalls should I avoid when I’m doing it?

Results

  • Larger windows are (much) slower
  • Resizing images on-the-fly is very slow, even if they are the same size every time
  • Drawing small images is fast, but drawing large images is slow
  • Drawing rectangles is fast
  • Drawing text is fast
  • Drawing Swing widgets in front of a canvas is fast
  • Creating fonts on-the-fly is a tiny bit slow

Code

You can find the full code (written in Kotlin) at gitlab.com/andybalaam/java-2d-performance.

Basically, we make a JFrame and a Canvas and tell them not to listen to repaints (i.e. we control their drawing).

val app = JFrame()
app.ignoreRepaint = true
val canvas = Canvas()
canvas.ignoreRepaint = true

Then we add any buttons to the JFrame, and the canvas last (so it displays behind):

app.add(button)
app.add(canvas)

Now we make the canvas double-buffered and get hold of a buffer image for it:

app.isVisible = true
canvas.createBufferStrategy(2)
val bufferStrategy = canvas.bufferStrategy
val bufferedImage = GraphicsEnvironment
    .getLocalGraphicsEnvironment()
    .defaultScreenDevice
    .defaultConfiguration
    .createCompatibleImage(config.width, config.height)

Then inside a tight loop we draw onto the buffer image:

val g2d = bufferedImage.createGraphics()
try
{
    g2d.color = backgroundColor
    g2d.fillRect(0, 0, config.width, config.height)

    ... the different drawing operations go here ...

and then swap the buffers:

    val graphics = bufferStrategy.drawGraphics
    try {
        graphics.drawImage(bufferedImage, 0, 0, null)
        if (!bufferStrategy.contentsLost()) {
            bufferStrategy.show()
        }
    } finally {
        graphics.dispose()
    }
} finally {
    g2d.dispose()
}

Results

Baseline: some rectangles

I decided to compare everything against drawing 20 rectangles at random points on the screen, since that seems like a minimal requirement for a game.

My test machine is an Intel Core 2 Duo E6550 2.33GHz with 6GB RAM and a GeForce GT 740 graphics card (I have no idea whether it is being used here – I assume not). I am running Ubuntu 18.04.1 Linux, OpenJDK Java 1.8.0_191, and Kotlin 1.3.20-release-116. (I expect the results would be identical if I were using Java rather than Kotlin.)

I ran all the tests in two window sizes: 1600×900 and 640×480. 640.×480 was embarrassingly fast for all tests, but 1600×900 struggled with some of the tasks.

Drawing rectangles looks like this:

g2d.color = Color(
    rand.nextInt(256),
    rand.nextInt(256),
    rand.nextInt(256)
)
g2d.fillRect(
    rand.nextInt(config.width / 2),
    rand.nextInt(config.height / 2),
    rand.nextInt(config.width / 2),
    rand.nextInt(config.height / 2)
)

In the small window, the baseline (20 rectangles) ran at 553 FPS. In the large window it ran at 87 FPS.

I didn’t do any statistics on these numbers because I am too lazy. Feel free to do it properly and let me know the results – I will happily update the article.

Fewer rectangles

When I reduced the number of rectangles to do less drawing work, I saw small improvements in performance. In the small window, drawing 2 rectangles instead of 20 increased the frame rate from 553 to 639, but there is a lot of noise in those results, and other runs were much closer. In the large window, the same reduction improved the frame rate from 87 to 92. This is not a huge speed-up, showing that drawing rectangles is pretty fast.

Adding fixed-size images

Drawing pre-scaled images looks like this:

g2d.drawImage(
    image,
    rand.nextInt(config.width),
    rand.nextInt(config.height),
    null
)

When I added 20 small images (40×40 pixels) to be drawn in each frame, the performance was almost unchanged. In the small window, the run showing 20 images per frame (as well as rectangle) actually ran faster than the one without (561 FPS versus 553), suggesting the difference is negligible and I should do some statistics. In the large window, the 20 images version ran at exactly the same speed (87 FPS).

So, it looks like drawing small images costs almost nothing.

When I moved to large images (400×400 pixels), the small window slowed down from 553 to 446 FPS, and the large window slowed from 87 to 73 FPS, so larger images clearly have an impact, and we will need to limit the number and size of images to keep the frame rate acceptable.

Scaling images on the fly

You can scale an image on the fly as you draw onto a Canvas. (Spoiler: don’t do this!)

My code looks like:

val s = config.imageSize
val x1 = rand.nextInt(config.width)
val y1 = rand.nextInt(config.height)
val x2 = x1 + s
val y2 = y1 + s
g2d.drawImage(
    unscaledImage,
    x1, y1, x2, y2,
    0, 0, unscaledImageWidth, unscaledImageHeight,
    null
)

Note the 10-argument form of drawImage is being used. You can be sure you have avoided this situation if you use the 4-argument form from the previous section.

Note: the resulting image is the same size every time, and the Java documentation implies that scaled images may be cached by the system, but I saw a huge slow-down when using the 10-argument form of drawImage above.

On-the-fly scaled images slowed the small window from 446 to 67 FPS(!), and the large window from 73 to 31 FPS, meaning the exact same rendering took over twice as long.

Advice: check you are not using one of the drawImage overloads that scales images! Pre-scale them yourself (e.g. with getScaledInstance as I did here).

Displaying text

Drawing text on the canvas like this:

g2d.font = Font("Courier New", Font.PLAIN, 12)
g2d.color = Color.GREEN
g2d.drawString("FPS: $fpsLastSecond", 20, 20 + i * 14)

had a similar impact to drawing small images – i.e. it only affected the performance very slightly and is generally quite fast. The small window slowed from 553 to 581 FPS, and the large window from 87 to 88.

Creating the font every time (as shown above) slowed the process a little more, so it is worth moving the font creating out of the game loop and only doing it once. The slowdown just for creating the font was 581 to 572 FPS in the small window, and 88 to 86 FPS in the large.

Swing widgets

By adding Button widgets to the JFrame before the Canvas, I was able to display them in front. Their rendering and focus worked as expected, and they had no impact at all on performance.

The same was true when I tried adding these widgets in front of images rendered on the canvas (instead of rectangles).

Turning everything up to 11

When I added everything I had tested all at the same time: rectangles, text with a new font every time, large unscaled images, and large window, the frame rate reduced to 30 FPS. This is a little slow for a game already, and if we had more images to draw it could get even worse. However, when I pre-scaled the images the frame rate went up to 72 FPS, showing that Java is capable of running a game at an acceptable frame rate on my machine, so long as we are careful how we use it.

Numbers

Small window (640×480)

Test FPS
nothing 661
rectangles2 639
rectangles20 553
rectangles20 images2 538
rectangles20 images20 561
rectangles20 images20 largeimages 446
rectangles20 images20 unscaledimages 343
rectangles20 images20 largeimages unscaledimages 67
rectangles20 text2 582
rectangles20 text20 581
rectangles20 text20 newfont 572
rectangles20 buttons2 598
rectangles20 buttons20 612

Large window (1200×900)

Test FPS
large nothing 93
large rectangles2 92
large rectangles20 87
large rectangles20 images2 87
large rectangles20 images20 87
large rectangles20 images20 largeimages 73
large rectangles20 images20 unscaledimages 82
large rectangles20 images20 largeimages unscaledimages 31
large rectangles20 text2 89
large rectangles20 text20 88
large rectangles20 text20 newfont 86
large rectangles20 buttons2 88
large rectangles20 buttons20 87
large images20 buttons20 largeimages 74
large rectangles20 images20 text20 buttons20 largeimages newfont 72
large rectangles20 images20 text20 buttons20 largeimages unscaledimages newfont 30

Feedback please

Please do get back to me with tips about how to improve the performance of my experimental code.

Feel free to log issues, make merge requests or add comments to the blog post.

Impact of group size and practice on manual performance

How performance varies with group size is an interesting question that is still an unresearched area of software engineering. The impact of learning is also an interesting question and there has been some software engineering research in this area.

I recently read a very interesting study involving both group size and learning, and Jaakko Peltokorpi kindly sent me a copy of the data.

That is the good news; the not so good news is that the experiment was not about software engineering, but the manual assembly of a contraption of the experimenters devising. Still, this experiment is an example of the impact of group size and learning (through repeating the task).

Subjects worked in groups of one to four people and repeated the task four times. Time taken to assemble a bespoke, floor standing rack with some odd-looking connections between components (the image in the paper shows an image of something that might function as a floor standing book-case, if shelves were added, apart from some component connections getting in the way) was measured.

The following equation is a very good fit to the data (code+data). There is theory explaining why log(repetitions) applies, but the division by group-size was found by suck-it-and-see (in another post I found that time spent planning increased with teams size).

There is a strong repetition/group-size interaction. As the group size increases, repetition has less of an impact on improving performance.

time = 0.16+ 0.53/{group size} - log(repetitions)*[0.1 + {0.22}/{group size}]

The following plot shows one way of looking at the data (larger groups take less time, but the difference declines with practice):

Time taken (hours) for various group sizes, by repetition.

and here is another (a group of two is not twice as fast as a group of one; with practice smaller groups are converging on the performance of larger groups):

Time taken (hours) for various repetitions, by group size.

Would the same kind of equation fit the results from solving a software engineering task? Hopefully somebody will run an experiment to find out :-)

Publishing information on project progress: will it impact delivery?

Numbers for delivery date and cost estimates, for a software project, depend on who you ask (the same is probably true for other kinds of projects). The people actually doing the work are likely to have the most accurate information, but their estimates can still be wildly optimistic. The managers of the people doing the work have to plan (i.e., make worst/best case estimates) and deal with people outside the team (i.e., sell the project to those paying for it); planning requires knowledge of where things are and where they need to be, while selling requires being flexible with numbers.

A few weeks ago I was at a hackathon organized by the people behind the Project Data and Analytics meetup. The organizers (Martin Paver & co.) had obtained some very interesting project related data sets. I worked on the Australian ICT dashboard data.

The Australian ICT dashboard data was courtesy of the Queensland state government, which has a publicly available dashboard listing digital project expenditure; the Victorian state government also has a dashboard listing ICT expenditure. James Smith has been collecting this data on a monthly basis.

What information might meaningfully be extracted from monthly estimates of project delivery dates and costs?

If you were running one of these projects, and had to provide monthly figures, what strategy would you use to select the numbers? Obviously keep quiet about internal changes for as long as possible (today’s reduction can be used to offset a later increase, or vice versa). If the client requests changes which impact date/cost, then obviously update the numbers immediately; the answer to the question about why the numbers changed is that, “we are responding to client requests” (i.e., we would otherwise still be on track to meet the original end-points).

What is the intended purpose of publishing this information? Is it simply a case of the public getting fed up with overruns, with publishing monthly numbers is seen as a solution?

What impact could monthly publication have? Will clients think twice before requesting an enhancement, fearing public push back? Will companies doing the work make more reliable estimates, or work harder?

Project delivery dates/costs change because new functionality/work-to-do is discovered, because the appropriate staff could not be hired and other assorted unknown knowns and unknowns.

Who is looking at this data (apart from half a dozen people at a hackathon on the other side of the world)?

Data on specific projects can only be interpreted in the context of that project. There is some interesting research to be done on the impact of public availability on client and vendor reporting behavior.

Will publication have an impact on performance? One way to get some idea is to run an A/B experiment. Some projects have their data made public, others don’t. Wait a few years, and compare project performance for the two publication regimes.

Time taken to compile a source file

How long will it take to compile a source file?

When computers were a lot slower than they are today, this question was of general interest. Job scheduling is more effective when reliable runtime estimates are available, and developers want to know if there is enough time to get a coffee before the compile finishes.

An embarrassing fact about compile time performance, used to be that a large percentage of compile time was spent doing lexical analysis [“The cost of lexical analysis”, I cannot find an online copy]. Why was this embarrassing? Compiler writers like to boast about all the fancy optimizations their compiler does; but doing fancy stuff consumes lots of resources, so why were compilers spending so much of their time doing simple things like lexical analysis? The reality was that fancy compiler optimizations were not commercially viable until developer computers contained tens of megabytes of memory, i.e., very few pre-1990 compilers did any real optimization (people are still fussing over lexer performance).

An analysis of the data in Captain Dennis Miller’s Masters thesis (late Rome period), finds compile time is proportional to the square root of the number of tokens in the source (code+data); more complicated models are a slightly better fit. Where did square root come from? I expected a linear relationship, but would be willing to go with log. The measurements are from Ada compilers in the mid 1980s. I know several people who worked on Ada compilers during that time, and they were implementing the latest fancy optimizations (Ada was going to be the next big thing and the venture capital was flowing; big companies, with big computers were going to be paying lots of money to use Ada, but then microcomputers came along). I think that square root is driven by OS resource limitations, the compilers are using lots of memory and a noticeable amount of time is spent swapping.

So computers got a lot faster and people lost interest in estimates of how long it would take to compile individual files. I have not seen any interest in predicting how long it would take to compile whole projects (just complaints about how long it takes). There has been some work on progress indicators, updated as compilation progresses, which is a step in the right direction. Perhaps somebody has recorded compile time information and thrown machine learning at it; I usually ignore machine learning papers applied to software engineering and perhaps I have missed something. Pointers to project compile time prediction work welcome.

Then along came just-in-time compilation. Now people want to estimate how long it will take to generate machine code from some intermediate form, that is being interpreted.

The plot below (thanks to Rafael Auler for kindly supplying the data from his paper) shows the time taken to generate code from functions containing a given number of LLVM instructions (an intermediate code), at optimization level O3. The red line is a regression fit to one of the ‘arms’ and shows constant time for less than 100’ish instructions and then a linear relationship. I have no idea why the time is roughly constant for a large number of functions.

Time taken to convert functions containing a given number of LLVM instructions to machine code

There is a lot of variation for function containing the same number of instructions. This is to be expected when lots of different optimizations are being tried; sometimes a function will contain lots of the kind of code that a particular optimization spends lot of times process and sometimes the code will not contain anything interesting (i.e., no optimizations are found).

Main memory: the crucial component that vendors don’t mention

CPU performance hogs the limelight when people discuss the year-on-year increases in computing power that used to occur.

This focus on cpu performance was/is driven by marketing, the people with the money either don’t want customers thinking about the performance impact of main memory size or speed, or want them to treat the processor as the most important component of a computer. Vendors want processor performance to drive customer purchase decisions.

Hardware manufacturers used to entice new customers with low cost machines, containing minimal memory. Once a customer started to use their shiny new computer, they found that it did save them lots of time and money, but also they needed more memory (which could only be brought from the manufacturer and was not cheap).

The plot below shows the prices IBM charged for System 360s, in 1966. Anti-trust investigations uncover all kinds of interesting data, like selling low-spec equipment at a loss to entice customers and make life difficult for competitors (code+data for all plots).

Profit margin on IBM 360s sold with various memory sizes

The plot below (data from the 19 Aug 1985 issue of ComputerWorld) shows how the price of computers increased as the minimum about of memory they supported increased.

Yes, in 1985 top end computers came with over 50M of memory; but most customers thought themselves lucky if they had a few megabytes.

If the processor is slow, it just takes longer for programs to run. If the computer does not have enough memory, programs cannot run. For most applications memory requirements are addressed first, followed by processor performance; memory requirements is the number one issue. The optimizations that commercial compilers could perform were limited by the memory capacity of developer machines.

List price of computers, in 1985, supporting the given minimum amount of  memory

Intel’s main line of business used to be selling memory chips, but these chips became commodity items as more companies entered the market; Intel bet the farm on selling processors and the rest is history. As a seller of a unique product it was/is in Intel’s interest to spend lots of money on marketing the benefits of processor performance; sellers of commodity items (such as memory chips) don’t have nearly as much to gain from generic product marketing, because customers may choose to buy from other sellers (in such markets sellers have to concentrate on marketing themselves).

Memory capacity/speed and cpu speed are two aspects of system performance; they need to be balanced to meet customer drive application requirements. The plot below shows the SPEC cpu integer performance of 4,332 systems running at various clock rates; the colors denote the different peak memory transfer rates of the memory chips in these systems (code+data).

SPEC cpu integer performance vs. cpu clock rate

These days (and perhaps in the past, I don’t have any data), memory performance is a much better predictor of system performance, but vendors don’t have an incentive to market this fact.

Adding a concurrency limit to Python’s asyncio.as_completed

Series: asyncio basics, large numbers in parallel, parallel HTTP requests, adding to stdlib

In the previous post I demonstrated how the limited_as_completed method allows us to run a very large number of tasks using concurrency, but limiting the number of concurrent tasks to a sensible limit to ensure we don’t exhaust resources like memory or operating system file handles.

I think this could be a useful addition to the Python standard library, so I have been working on a modification to the current asyncio.as_completed method. My work so far is here: limited-as_completed.

I ran similar tests to the ones I ran for the last blog post with this code to validate that the modified standard library version achieves the same goals as before.

I used an identical copy of timed from the previous post and updated versions of the other files because I was using a much newer version of aiohttp along with the custom-built python I was running.

server looked like:

#!/usr/bin/env python3

from aiohttp import web
import asyncio
import random

async def handle(request):
    await asyncio.sleep(random.randint(0, 3))
    return web.Response(text="Hello, World!")

app = web.Application()
app.router.add_get('/{name}', handle)

web.run_app(app)

client-async-sem needed me to add a custom TCPConnector to avoid a new limit on the number of concurrent connections that was added to aiohttp in version 2.0. I also need to move the ClientSession usage inside a coroutine to avoid a warning:

#!/usr/bin/env python3

from aiohttp import ClientSession, TCPConnector
import asyncio
import sys

limit = 1000

async def fetch(url, session):
    async with session.get(url) as response:
        return await response.read()

async def bound_fetch(sem, url, session):
    # Getter function with semaphore.
    async with sem:
        await fetch(url, session)

async def run(r):
    with ClientSession(connector=TCPConnector(limit=limit)) as session:
        url = "http://localhost:8080/{}"
        tasks = []
        # create instance of Semaphore
        sem = asyncio.Semaphore(limit)
        for i in range(r):
            # pass Semaphore and session to every GET request
            task = asyncio.ensure_future(
                bound_fetch(sem, url.format(i), session))
            tasks.append(task)
        responses = asyncio.gather(*tasks)
        await responses

loop = asyncio.get_event_loop()
loop.run_until_complete(asyncio.ensure_future(run(int(sys.argv[1]))))

My new code that uses my proposed extension to as_completed looked like:

#!/usr/bin/env python3

from aiohttp import ClientSession, TCPConnector
import asyncio
import sys

async def fetch(url, session):
    async with session.get(url) as response:
        return await response.read()

limit = 1000

async def print_when_done():
    with ClientSession(connector=TCPConnector(limit=limit)) as session:
        tasks = (fetch(url.format(i), session) for i in range(r))
        for res in asyncio.as_completed(tasks, limit=limit):
            await res

r = int(sys.argv[1])
url = "http://localhost:8080/{}"
loop = asyncio.get_event_loop()
loop.run_until_complete(print_when_done())
loop.close()

and with these, we get similar behaviour to the previous post:

$ ./timed ./client-async-sem 10000
Memory usage: 73640KB	Time: 19.18 seconds
$ ./timed ./client-async-stdlib 10000
Memory usage: 49332KB	Time: 18.97 seconds

So the implementation I plan to submit to the Python standard library appears to work well. In fact, I think it is better than the one I presented in the previous post, because it uses on_complete callbacks to notice when futures have completed, which reduces the busy-looping we were doing to check for and yield finished tasks.

The Python issue is bpo-30782 and the pull request is #2424.

Note: at first glance, it looks like the aiohttp.ClientSession‘s limit on the number of connections (introduced in version 1.0 and then updated in version 2.0) gives us what we want without any of this extra code, but in fact it only limits the number of connections, not the number of futures we are creating, so it has the same problem of unbounded memory use as the semaphore-based implementation.

Making 100 million requests with Python aiohttp

Series: asyncio basics, large numbers in parallel, parallel HTTP requests, adding to stdlib

I’ve been working on how to make a very large number of HTTP requests using Python’s asyncio and aiohttp.

Paweł Miech’s post Making 1 million requests with python-aiohttp taught me how to think about this, and got us a long way, with 1 million requests running in a reasonable time, but I need to go further.

Paweł’s approach limits the number of requests that are in progress, but it uses an unbounded amount of memory to hold the futures that it wants to execute.

We can avoid using unbounded memory by using the limited_as_completed function I outined in my previous post.

Setup

Server

We have a server program “server”:

(Note it differs from Paweł’s version because I am using an older version of aiohttp which has fewer convenient features.)

#!/usr/bin/env python3.5

from aiohttp import web
import asyncio
import random

async def handle(request):
    await asyncio.sleep(random.randint(0, 3))
    return web.Response(text="Hello, World!")

async def init():
    app = web.Application()
    app.router.add_route('GET', '/{name}', handle)
    return await loop.create_server(
        app.make_handler(), '127.0.0.1', 8080)

loop = asyncio.get_event_loop()
loop.run_until_complete(init())
loop.run_forever()

This just responds “Hello, World!” to every request it receives, but after an artificial delay of 0-3 seconds.

Synchronous client

As a baseline, we have a synchronous client “client-sync”:

#!/usr/bin/env python3.5

import requests
import sys

url = "http://localhost:8080/{}"
for i in range(int(sys.argv[1])):
    requests.get(url.format(i)).text

This waits for each request to complete before making the next one. Like the other clients below, it takes the number of requests to make as a command-line argument.

Async client using semaphores

Copied mostly verbatim from Making 1 million requests with python-aiohttp we have an async client “client-async-sem” that uses a semaphore to restrict the number of requests that are in progress at any time to 1000:

#!/usr/bin/env python3.5

from aiohttp import ClientSession
import asyncio
import sys

limit = 1000

async def fetch(url, session):
    async with session.get(url) as response:
        return await response.read()

async def bound_fetch(sem, url, session):
    # Getter function with semaphore.
    async with sem:
        await fetch(url, session)

async def run(session, r):
    url = "http://localhost:8080/{}"
    tasks = []
    # create instance of Semaphore
    sem = asyncio.Semaphore(limit)
    for i in range(r):
        # pass Semaphore and session to every GET request
        task = asyncio.ensure_future(bound_fetch(sem, url.format(i), session))
        tasks.append(task)
    responses = asyncio.gather(*tasks)
    await responses

loop = asyncio.get_event_loop()
with ClientSession() as session:
    loop.run_until_complete(asyncio.ensure_future(run(session, int(sys.argv[1]))))

Async client using limited_as_completed

The new client I am presenting here uses limited_as_completed from the previous post. This means it can make a generator that provides the futures to wait for as they are needed, instead of making them all at the beginning.

It is called “client-async-as-completed”:

#!/usr/bin/env python3.5

from aiohttp import ClientSession
import asyncio
from itertools import islice
import sys

def limited_as_completed(coros, limit):
    futures = [
        asyncio.ensure_future(c)
        for c in islice(coros, 0, limit)
    ]
    async def first_to_finish():
        while True:
            await asyncio.sleep(0)
            for f in futures:
                if f.done():
                    futures.remove(f)
                    try:
                        newf = next(coros)
                        futures.append(
                            asyncio.ensure_future(newf))
                    except StopIteration as e:
                        pass
                    return f.result()
    while len(futures) > 0:
        yield first_to_finish()

async def fetch(url, session):
    async with session.get(url) as response:
        return await response.read()

limit = 1000

async def print_when_done(tasks):
    for res in limited_as_completed(tasks, limit):
        await res

r = int(sys.argv[1])
url = "http://localhost:8080/{}"
loop = asyncio.get_event_loop()
with ClientSession() as session:
    coros = (fetch(url.format(i), session) for i in range(r))
    loop.run_until_complete(print_when_done(coros))
loop.close()

Again, this limits the number of requests to 1000.

Test setup

Finally, we have a test runner script called “timed”:

#!/usr/bin/env bash

./server &
sleep 1 # Wait for server to start

/usr/bin/time --format "Memory usage: %MKB\tTime: %e seconds" "$@"

# %e Elapsed real (wall clock) time used by the process, in seconds.
# %M Maximum resident set size of the process in Kilobytes.

kill %1

This runs each process, ensuring the server is restarted each time it runs, and prints out how long it took to run, and how much memory it used.

Results

When making only 10 requests, the async clients worked faster because they launched all the requests simultaneously and only had to wait for the longest one (3 seconds). The memory usage of all three clients was fine:

$ ./timed ./client-sync 10
Memory usage: 20548KB	Time: 15.16 seconds
$ ./timed ./client-async-sem 10
Memory usage: 24996KB	Time: 3.13 seconds
$ ./timed ./client-async-as-completed 10
Memory usage: 23176KB	Time: 3.13 seconds

When making 100 requests, the synchronous client was very slow, but all three clients worked eventually:

$ ./timed ./client-sync 100
Memory usage: 20528KB	Time: 156.63 seconds
$ ./timed ./client-async-sem 100
Memory usage: 24980KB	Time: 3.21 seconds
$ ./timed ./client-async-as-completed 100
Memory usage: 24904KB	Time: 3.21 seconds

At this point let’s agree that life is too short to wait for the synchronous client.

When making 10000 requests, both async clients worked quite quickly, and both had increased memory usage, but the semaphore-based one used almost twice as much memory as the limited_as_completed version:

$ ./timed ./client-async-sem 10000
Memory usage: 77912KB	Time: 18.10 seconds
$ ./timed ./client-async-as-completed 10000
Memory usage: 46780KB	Time: 17.86 seconds

For 1 million requests, the semaphore-based client took 25 minutes on my (32GB RAM) machine. It only used about 10% of my CPU, and it used a lot of memory (over 3GB):

$ ./timed ./client-async-sem 1000000
Memory usage: 3815076KB	Time: 1544.04 seconds

Note: Paweł’s version only took 9 minutes on his laptop and used all his CPU, so I wonder whether I have made a mistake somewhere, or whether my version of Python (3.5.2) is not as good as a later one.

The limited_as_completed version ran in a similar amount of time but used 100% of my CPU, and used a much smaller amount of memory (162MB):

$ ./timed ./client-async-as-completed 1000000
Memory usage: 162168KB	Time: 1505.75 seconds

Now let’s try 100 million requests. The semaphore-based version lasted 10 hours before it was killed by Linux’s OOM Killer, but it didn’t manage to make any requests in this time, because it creates all its futures before it starts making requests:

$ ./timed ./client-async-sem 100000000
Command terminated by signal 9

I left the limited_as_completed version over the weekend and it managed to succeed eventually:

$ ./timed ./client-async-as-completed 100000000
Memory usage: 294304KB	Time: 150213.15 seconds

So its memory usage was still very bounded, and it managed to do about 665 requests/second over an extended period, which is almost identical to the throughput of the previous cases.

Conclusion

Making a million requests is usually enough, but when we really need to do a lot of work while keeping our memory usage bounded, it looks like an approach like limited_as_completed is a good way to go. I also think it’s slightly easier to understand.

In the next post I describe my attempt to get something like this added to the Python standard library.

Python 3 – large numbers of tasks with limited concurrency

Series: asyncio basics, large numbers in parallel, parallel HTTP requests, adding to stdlib

I am interested in running large numbers of tasks in parallel, so I need something like asyncio.as_completed, but taking an iterable instead of a list, and with a limited number of tasks running concurrently. First, let’s try to build something pretty much equivalent to asyncio.as_completed. Here is my attempt, but I’d welcome feedback from readers who know better:

# Note this is not a coroutine - it returns
# an iterator - but it crucially depends on
# work being done inside the coroutines it
# yields - those coroutines empty out the
# list of futures it holds, and it will not
# end until that list is empty.
def my_as_completed(coros):

    # Start all the tasks
    futures = [asyncio.ensure_future(c) for c in coros]

    # A coroutine that waits for one of the
    # futures to finish and then returns
    # its result.
    async def first_to_finish():

        # Wait forever - we could add a
        # timeout here instead.
        while True:

            # Give up control to the scheduler
            # - otherwise we will spin here
            # forever!
            await asyncio.sleep(0)

            # Return anything that has finished
            for f in futures:
                if f.done():
                    futures.remove(f)
                    return f.result()

    # Keep yielding a waiting coroutine
    # until all the futures have finished.
    while len(futures) > 0:
        yield first_to_finish()

The above can be substituted for asyncio.as_completed in the code that uses it in the first article, and it seems to work. It also makes a reasonable amount of sense to me, so it may be correct, but I’d welcome comments and corrections.

my_as_completed above accepts an iterable and returns a generator producing results, but inside it starts all tasks concurrently, and stores all the futures in a list. To handle bigger lists we will need to do better, by limiting the number of running tasks to a sensible number.

Let’s start with a test program:

import asyncio
async def mycoro(number):
    print("Starting %d" % number)
    await asyncio.sleep(1.0 / number)
    print("Finishing %d" % number)
    return str(number)

async def print_when_done(tasks):
    for res in asyncio.as_completed(tasks):
        print("Result %s" % await res)

coros = [mycoro(i) for i in range(1, 101)]

loop = asyncio.get_event_loop()
loop.run_until_complete(print_when_done(coros))
loop.close()

This uses asyncio.as_completed to run 100 tasks and, because I adjusted the asyncio.sleep command to wait longer for earlier tasks, it prints something like this:

$ time python3 python-async.py
Starting 47
Starting 93
Starting 48
...
Finishing 93
Finishing 94
Finishing 95
...
Result 93
Result 94
Result 95
...
Finishing 46
Finishing 45
Finishing 42
...
Finishing 2
Result 2
Finishing 1
Result 1

real    0m1.590s
user    0m0.600s
sys 0m0.072s

So all 100 tasks were completed in 1.5 seconds, indicating that they really were run in parallel, but all 100 were allowed to run at the same time, with no limit.

We can adjust the test program to run using our customised my_as_completed function, and pass in an iterable of coroutines instead of a list by changing the last part of the program to look like this:

async def print_when_done(tasks):
    for res in my_as_completed(tasks):
        print("Result %s" % await res)
coros = (mycoro(i) for i in range(1, 101))
loop = asyncio.get_event_loop()
loop.run_until_complete(print_when_done(coros))
loop.close()

But we get similar output to last time, with all tasks running concurrently.

To limit the number of concurrent tasks, we limit the size of the futures list, and add more as needed:

from itertools import islice
def limited_as_completed(coros, limit):
    futures = [
        asyncio.ensure_future(c)
        for c in islice(coros, 0, limit)
    ]
    async def first_to_finish():
        while True:
            await asyncio.sleep(0)
            for f in futures:
                if f.done():
                    futures.remove(f)
                    try:
                        newf = next(coros)
                        futures.append(
                            asyncio.ensure_future(newf))
                    except StopIteration as e:
                        pass
                    return f.result()
    while len(futures) > 0:
        yield first_to_finish()

We start limit tasks at first, and whenever one ends, we ask for the next coroutine in coros and set it running. This keeps the number of running tasks at or below limit until we start running out of input coroutines (when next throws and we don’t add anything to futures), then futures starts emptying until we eventually stop yielding coroutine objects.

I thought this function might be useful to others, so I started a little repo over here and added it: asyncioplus/limited_as_completed.py. Please provide merge requests and log issues to improve it – maybe it should be part of standard Python?

When we run the same example program, but call limited_as_completed instead of the other versions:

async def print_when_done(tasks):
    for res in limited_as_completed(tasks, 10):
        print("Result %s" % await res)
coros = (mycoro(i) for i in range(1, 101))
loop = asyncio.get_event_loop()
loop.run_until_complete(print_when_done(coros))
loop.close()

We see output like this:

$ time python3 python-async.py
Starting 1
Starting 2
...
Starting 9
Starting 10
Finishing 10
Result 10
Starting 11
...
Finishing 100
Result 100
Finishing 1
Result 1

real	0m1.535s
user	0m1.436s
sys	0m0.084s

So we can see that the tasks are still running concurrently, but this time the number of concurrent tasks is limited to 10.

See also

To achieve a similar result using semaphores, see Python asyncio.semaphore in async-await function and Making 1 million requests with python-aiohttp.

It feels like limited_as_completed is more re-usable as an approach but I’d love to hear others’ thoughts on this. E.g. could/should I use a semaphore to implement limited_as_completed instead of manually holding a queue?

Basic ideas of Python 3 asyncio concurrency

Series: asyncio basics, large numbers in parallel, parallel HTTP requests, adding to stdlib

Python 3’s asyncio module and the async and await keywords combine to allow us to do cooperative concurrent programming, where a code path voluntarily yields control to a scheduler, trusting that it will get control back when some resource has become available (or just when the scheduler feels like it). This way of programming can be very confusing, and has been popularised by Twisted in the Python world, and nodejs (among others) in other worlds.

I have been trying to get my head around the basic ideas as they surface in Python 3’s model. Below are some definitions and explanations that have been useful to me as I tried to grasp how it all works.

Futures and coroutines are both things that you can wait for.

You can make a coroutine by declaring it with async def:

import asyncio
async def mycoro(number):
    print("Starting %d" % number)
    await asyncio.sleep(1)
    print("Finishing %d" % number)
    return str(number)

Almost always, a coroutine will await something such as some blocking IO. (Above we just sleep for a second.) When we await, we actually yield control to the scheduler so it can do other work and wake us up later, when something interesting has happened.

You can make a future out of a coroutine, but often you don’t need to. Bear in mind that if you do want to make a future, you should use ensure_future, but this actually runs what you pass to it – it doesn’t just create a future:

myfuture1 = asyncio.ensure_future(mycoro(1))
# Runs mycoro!

But, to get its result, you must wait for it – it is only scheduled in the background:

# Assume mycoro is defined as above
myfuture1 = asyncio.ensure_future(mycoro(1))
# We end the program without waiting for the future to finish

So the above fails like this:

$ python3 ./python-async.py
Task was destroyed but it is pending!
task: <Task pending coro=<mycoro() running at ./python-async:10>>
sys:1: RuntimeWarning: coroutine 'mycoro' was never awaited

The right way to block waiting for a future outside of a coroutine is to ask the event loop to do it:

# Keep on assuming mycoro is defined as above for all the examples
myfuture1 = asyncio.ensure_future(mycoro(1))
loop = asyncio.get_event_loop()
loop.run_until_complete(myfuture1)
loop.close()

Now this works properly (although we’re not yet getting any benefit from being asynchronous):

$ python3 python-async.py
Starting 1
Finishing 1

To run several things concurrently, we make a future that is the combination of several other futures. asyncio can make a future like that out of coroutines using asyncio.gather:

several_futures = asyncio.gather(
    mycoro(1), mycoro(2), mycoro(3))
loop = asyncio.get_event_loop()
print(loop.run_until_complete(several_futures))
loop.close()

The three coroutines all run at the same time, so this only takes about 1 second to run, even though we are running 3 tasks, each of which takes 1 second:

$ python3 python-async.py
Starting 3
Starting 1
Starting 2
Finishing 3
Finishing 1
Finishing 2
['1', '2', '3']

asyncio.gather won’t necessarily run your coroutines in order, but it will return a list of results in the same order as its input.

Notice also that run_until_complete returns the result of the future created by gather – a list of all the results from the individual coroutines.

To do the next bit we need to know how to call a coroutine from a coroutine. As we’ve already seen, just calling a coroutine in the normal Python way doesn’t run it, but gives you back a “coroutine object”. To actually run the code, we need to wait for it. When we want to block everything until we have a result, we can use something like run_until_complete but in an async context we want to yield control to the scheduler and let it give us back control when the coroutine has finished. We do that by using await:

import asyncio
async def f2():
    print("start f2")
    await asyncio.sleep(1)
    print("stop f2")
async def f1():
    print("start f1")
    await f2()
    print("stop f1")
loop = asyncio.get_event_loop()
loop.run_until_complete(f1())
loop.close()

This prints:

$ python3 python-async.py
start f1
start f2
stop f2
stop f1

Now we know how to call a coroutine from inside a coroutine, we can continue.

We have seen that asyncio.gather takes in some futures/coroutines and returns a future that collects their results (in order).

If, instead, you want to get results as soon as they are available, you need to write a second coroutine that deals with each result by looping through the results of asyncio.as_completed and awaiting each one.

# Keep on assuming mycoro is defined as at the top
async def print_when_done(tasks):
    for res in asyncio.as_completed(tasks):
        print("Result %s" % await res)
coros = [mycoro(1), mycoro(2), mycoro(3)]
loop = asyncio.get_event_loop()
loop.run_until_complete(print_when_done(coros))
loop.close()

This prints:

$ python3 python-async.py
Starting 1
Starting 3
Starting 2
Finishing 3
Result 3
Finishing 2
Result 2
Finishing 1
Result 1

Notice that task 3 finishes first and its result is printed, even though tasks 1 and 2 are still running.

asyncio.as_completed returns an iterable sequence of futures, each of which must be awaited, so it must run inside a coroutine, which must be waited for too.

The argument to asyncio.as_completed has to be a list of coroutines or futures, not an iterable, so you can’t use it with a very large list of items that won’t fit in memory.

Side note: if we want to work with very large lists, asyncio.wait won’t help us here – it also takes a list of futures and waits for all of them to complete (like gather), or, with other arguments, for one of them to complete or one of them to fail. It then returns two sets of futures: done and not-done. Each of these must be awaited to get their results, so:

asyncio.gather

# is roughly equivalent to:

async def mygather(*args):
    ret = []
    for r in (await asyncio.wait(args))[0]:
        ret.append(await r)
    return ret

I am interested in running very large numbers of tasks with limited concurrency – see the next article for how I managed it.