Rants from the Ballmer Peak rss feed

40 years later we still can't be friends


The computer (or electronic device) you are likely using to read this article is also very likely to be doing some form of multitasking, where the machine gives the illusion of performing different tasks simultaneously. At the low level a single processor is only able to perform a single task, operation, action, whatever you want to call it. However, processors are so fast that switching from one task to another looks to us, humans, like they are doing multiple things.

For instance, even if you are waiting for a web page or program to load, and you see some sort of visual indicator, like a spinning wheel, as a user you may think the computer is busy doing one task, the loading of the web page or program. But at the low level the computer is doing much more, in fact, by simple virtue of drawing and updating an animated icon, updating the position of the mouse on the screen, and why not even playing some kpop music in the background while you patiently wait. In contrast, if the computer could not multitask, while the web page or program loaded, you could not do anything, not even move the mouse.

According to the erudites from wikipedia (citation needed), time sharing was a computing model which between 1960 and 1970 established itself as the way to share resources on big mainframes. Unix and many of its descendants, like the popular Linux inherited this computing model, since it was accepted as valid. Most of operating systems of today follow this computing model, since one of the troubles with the computing industry is its really impressive momentum and resistance to change.

There's a war going on inside your machine

There are many ways to explain how programs running on a computer share resources, but actually most ignore the fact that the programs are not sharing, nor willing to share those resources. Operating systems are really like guardians who provide access to a single resource (the CPU) in turns. If you have a single process running, it will get all the time slices of the CPU. But if you have two processes, the operating system will try to distribute equally the CPU among them. Depending on the type of the programs, they might not even use the CPU at all because they may be waiting for user input. The process gets the time slice of the CPU but yields it back to the OS. Hence modern machines have many processes, but they are actually sleeping, waiting for some event which will trigger a reaction.

The problem is of course the active CPU hogs. These could be playing music in the background, video playback, compressing images or rendering frames for a video game (games are constantly redrawing the game world on the screen for the user). But even if you are actively running a single program, it might run many different sub threads to perform its tasks. For this reason most task managers can display the list of programs running, and how many threads or children have they spawned. So you don't actually need multiple processes to trigger time sharing behaviour, it's enough for a single process with multiple threads.

The problem with this inherited approach is the way multitasking is expressed and handled in software. The most popular ways to split a task are to either fork a process, or spawn a thread. In both cases the source program decides how many processes or threads to create, and then coordinates the communication to control and complete the task. And here lies the problem, a process can query the number of available cores on the system and decide to spawn an equal number of threads to perform a task which can be subdivided (this is called parallelization). Why is this a problem at all? There are two:

Both of these can be handled perfectly by the operating system, since it knows all the necessary data to decide best how to divide tasks. Yet our current software threading model forces the programmer to decide without this information.

Task switching killing your scalability

Let's write a photo processing software! Or maybe video. Anyway, this kind of software operates on bidimensional images which can usually be split into smaller chunks and dealt with mostly individually without dependencies. Tasks inside the computer don't magically migrate to other cores. If we write this software in single threaded mode, the four core machine will have one core working at the maximum, and three idling. What a waste. No problem, we subdivide the image and feed the chunks to the four cores. Now the performance is nearly four times that of the original single threaded code (we have a small overhead for splitting/controlling tasks).

Cool, now we can batch process porn pictures at the speed of light. But it takes time to go through our folder of midget porn, and we want to do other things in the meantime. Let's compress some video! Video edition can also benefit from parallelization, since at the basic level the individual images can also be split into chunks to feed different cores. Again, our video program detects four cores, splits the images in a queue and starts processing them four at a time. See the problem?

Now there are two processes on the quad core machine, each of them requesting to have the four cores for itself, but in total that means running eight threads at the same time. Unless we are running JesusOS which can multiply cores out of nowhere, the OS is just going to switch tasks between each core. Big deal, right? Yes, it's a big deal. When you start to measure performance of such programs in combination you realize that task switching is not free: the CPU has to change a lot of internal state and then the next task has to recover it. It takes time. And the more processes you run the worse it gets. So we end up with a machine which for each process overspawns many threads instead of getting one thread per core. Where doing tasks serially would take A + B + C seconds, now we have A + B + C + task switching overhead seconds, and the task switching overhead part can grow quite a lot, especially the more processes there are.

This considers a situation where the number of processing units is static all the time, but things can be harder especially on mobile devices where the hardware may decide to disable one or more processing units to save battery. Plugin in the laptop might give it a performance boost, and viceversa. For these situations the programming model we have dragged for over forty years is completely useless, there is no provision for changing the number of threads on the fly, you need clever programmers to implement such behaviour themselves, but there are clearly none since we haven't solved this yet, have we?

Take this ticket and wait for your turn

While most of the world was indulging in criticizing Apple for having economic success, they silently released Grand Central Dispatch (GCD) which is yet another task parallelism tool based on kqueue. GCD changes the way the programmer thinks about multithreading. Instead of saying "hey, I want 4 threads doing this much stuff", the programmer says "hey, I have these many tasks which can run parallel to each other without dependencies, run them please". This is a big change. While it can be argued that queues are easier to handle than threads, what this change means to the user is that the OS can now decide how many threads to allocate for a process. The OS doesn't face the simultaneous attack of dozens of processes, instead it sees dozens of processes waiting for their queues to finish. The OS can decide then to pick as many tasks from their queues and not worry (mostly) about switching threads.

Of course the devil is in the details. What if you are not subdividing your tasks well enough that they block the queues for other processes? What if the chunk of code in the queue blocks for disk I/O? What if... queues are not for solving the inherent threading problems OSes will keep having for the foreseeable future. But they help a lot in allowing them to decide what to run and when. In the example give above, the OS could decide to take only two tasks at the same time from the image process queue and two tasks from the video queue, and if any process finishes, the new slots can be given to the reminder tasks in other processes' queues. Similar scenario happens if the platform you are running enables/disables more processing cores. Have you imagined a hardware where you can plug in a card and double the processing speed of the running processes without them having to restart to take advantage of the change? Now you could.

In fact, all of this is in the past. Note that GCD was introduced in the year 2009. Since then, Apple has been pushing API changes all over their iOS and OSX frameworks to include blocks and queues where they make sense. Even if programmers of these platforms don't explicitly use queues for their programs, most of the libraries they will surely use are going to take advantage of these task parallelization techniques, thus gaining the advantages mentioned here. And of course, whenever they need to run something in the background, the Objective-C language and APIs will prod them towards queues rather than threads or processes.

The benefits from using queues are not invisible or theoretical. Already in November of 2010, Robbie Hanson (aka Deusty) wrote a blog post explaining the benefits of migrating its HTTP server (CocoaHTTPServer) to queues. Claimed performance improvements range from doubling to quadrupling, but the most impressive is the nearly linear scalability when the number of concurrent connections was increased. This is the golden dream: increase number of tasks with nearly zero overhead. And Robbie is collaborating to other pieces of software you might not expect could benefit from queues, like YapDatabase, built on top of SQLite and providing smooth database operations not blocking the user interface to preserve the fluidity of the user's interaction.

A bleak future

Yet here we are, nearly four years later still waiting for the revolution to happen. You could only hope the competition would clone this approach to threaded code as the phone industry copied the iPhone, but I haven't seen yet any other mainstream programming language embedding such functionality in its core language and standard library. And if you think that's bad, we still haven't talked about another pressing issue related to inter process hostility. If only I had the memory to remember what it was all about…

$ nim c work_faster.nim
work_faster.nim(1, 7) Error: cannot open 'threads'

See the article index or browse articles by tags: multitasking , user-experience , objc , apple .
Published on: 07/10/2013 23:29. Last update: 19/10/2013 23:39. rss feed
Copyright 2021 by Grzegorz Adam Hankiewicz.
Generated with ipsum genera. Look at the source code.