I am processing a stream of data pulled, usually, from a RxNode.fromReadableStream() where there is an intermediate step that is asynchronous and may only run with a certain concurrency, and at the same time the output step can only run in serial, in discrete batches.
Basically a classic transformation pipeline:
cat file | parallel -P5 transform | xargs -n1000 process
With promises it would be something like (extremely simplified):
lines.on('readable', function read() {
let line;
while (line = lines.read()) {
queue.push(line);
if (queue.length >= 10) {
return Promise.map(queue, item => someAsyncTransformations(item), {concurrency: 5})
.then(items => someFinalAsyncProcessing(items))
.then(read);
}
}
});
By inserting the controller after the transforms, I can ensure that the transforms execute independently, giving me pipelining; and the combination of defers and merging seems to let me control asynchronicity and concurrency.
However, I'm using controlled() to make sure the map step effectively pauses processing until done. It looks really ugly — isn't there a better, non-manual way to implement lossless backpressure?
Also, I believe I have to use the same control trick with the transformers, otherwise the entire stream will be transformed in one go and blow memory usage. Is there a more idiomatic way?
Finally: The above actually doesn't work if the size of the input isn't divisible by the batch size. For some reason if I do request() with more than the number of the remaining items, the entire pipeline just quits; onComplete isn't even called. Actaully, if I initally call request() with a large number, the entire pipeline repeats seemingly indefinitely with the same contents. Bug or intended behaviour?
Edit: I found I could use concatMap. Nevermind.
Related
I have a complex data processing pipeline using streams, in which I have a readable stream input, a writable stream output, and a series of transform streams (let's call them step1, step2, step3, and step4). While step1, step3, and output are stateless, relying only on the data chunks coming in to produce their output, chunk for chunk, step2 and step4 are aggregation steps, collecting data from multiple chunks to produce their output, and often having outputs that overlap time-wise (e.g. chunk1, chunk3 and chunk5 might produce output1, chunk2 and chunk4 might produce output2, and so on).
Currently, the pipeline is structured as follows:
input.pipe(step1).pipe(step2).pipe(step3).pipe(step4).pipe(output);
This pipeline is very computationally expensive, and as such I'd like to split it across multiple instances, preferably running on multiple cores. Node.js streams guarantee order preservation, so Node.js seems to balance the message passing such that data chunks coming out of one step first get passed into the next step first, and this is a property I'd need to have on whatever method I come up with of making this computation concurrent.
I'm definitely not asking for hand-holding, more if anyone has solved this problem before, and the general approach used for this kind of thing. I'm not really sure where to start.
Although I haven't been able to complete the order preservation, the streaming framework I'm supporting, scramjet, will get you really close to achieve your goal.
I'll nudge you here towards the best solution:
let seq = 0;
source.pipe(new DataStream())
.map(data => {data, itr: seq++}) // mark your order
.separate(x => x % 8) // separate into 8 streams
.cluster((stream) => { // spawn subprocesses
// do your multi threaded transforms here
}, {threads: 8})
.mux((a, b) => a.itr - b.itr) // merge in the order above
At some point I will introduce reordering, but in order to keep abstraction I can't take too many shortcuts, but you can take yours like the 2^52 limitation of count in the above example (seq will run out of bit space to increment then).
This should lead you towards some solution.
I'm using d3.js 3.5.6. How do we tick the force layout in our own render loop?
It seems that when I call force.start(), that automatically starts the force layout's own internal render loop (using requestAnimationFrame).
How do I prevent d3 from making a render loop, so that I can make my own render and call force.tick() myself?
This answer is plain wrong. Don't refer to it, don't use it.
I wrote a new one explaining how to do this correctly. I remember spending days digging into this as I though I had discovered an error. And, judging by the comments and the upvotes, I have managed to trick others—even including legends like Lars Kotthoff—to follow me down this wrong road. Anyways, I have learned a lot from my mistake. You only have to be ashamed of your errors if you do not take the chance to learn from them.
As soon as this answer is unaccepted I am going to delete it.
At first I was annoyed by the lack of code in the question and considered the answer to be rather easy and obvious. But, as it turned out, the problem has some unexpected implications and yields some interesting insights. If you are not interested in the details, you might want to have a look at my Final thoughts at the bottom for an executable solution.
I had seen code and documentation for doing the calculations of the force layout by explicitly calling force.tick.
# force.tick()
Runs the force layout simulation one step. This method can be used in conjunction with start and stop to compute a static layout. For example:
force.start();
for (var i = 0; i < n; ++i) force.tick();
force.stop();
This code always seemed dubious to me, but I took it for granted because the documentation had it and Mike Bostock himself made a "Static Force Layout" Block using the code from the docs. As it turns out, my intuition was right and both the Block as well as the documentation are wrong or at least widely off the track:
Calling start will do a lot of initialization of your nodes and links data (see documentation of nodes() and links(). You cannot just dismiss the call as you have experienced yourself. The force layout won't run without it.
Another thing start will eventually do is to fire up the processing loop by calling requestAnimationFrame or setTimeout, whatever is available, and provide force.tick as the callback. This results in an asynchronous processing which will repeatedly call force.tick, whereby doing the calculations and calling your tick handler if provided. The only non-hacky way to break this loop is to set alpha to below the hard-coded freezing point of 0.005 by calling force.alpha(0.005) or force.stop(). This will stop the loop on the next call to tick. Unless the timer is stopped this way, it will continue looping log0.99 (0.005 / 0.1) ≈ 298 times until alpha has dropped below the freezing point.
One should note, that this is not the case for the documentation or the Block. Hence, the tick-loop started by force.start() will continue running asynchronously and do its calculations.
The subsequent for-loop might or might not have any effect on the result of the force layout. If the timer happens to be still running in the background, this means concurrent calls to force.tick from the timer as well as from the for-loop. In any case will the calculations be stopped once alpha has dropped low enough when reaching a total of 298 calls to tick. This can be seen from the following lines:
force.tick = function() {
// simulated annealing, basically
if ((alpha *= 0.99) < 0.005) {
timer = null;
event.end({type: "end", alpha: alpha = 0});
return true;
}
// ...
}
From that point on you can call tick as often as you like without any change to the layout's outcome. The method is entered, but, because of the low value of alpha, will return immediately. All you will see is a repeated firing of end events.
To affect the number of iterations you have to control alpha.
The fact that the layout in the Block seems static is due to the fact that no callback for the "tick" event is registered which could update the SVG on every tick. The final result is only drawn once. And this result is ready after just 298 iterations, it won't be changed by subsequent, explicit calls to tick. The final call to force.stop() won't change anything either, it just sets alpha to 0. This does not have any effect on the result because the force layout has long come to an implicit halt.
Conclusion
Item 1. could be circumvented by a clever combination of starting and stopping the layout as in Stephen A. Thomas's great series "Understanding D3.js Force Layout" where from example 3 on he uses button controls to step through the calculations. This, however, will also come to a halt after 298 steps. To take full control of the iterations you need to
Provide a tick handler and immediately stop the timer by calling force.stop() therein. All calculations of this step will have been completed by then.
In your own loop calculate the new value for alpha. Setting this value by force.alpha() will restart the layout. Once the calculations of this next step are done, the tick handler will be executed resulting in an immediate stop as seen above. For this to work you will have to keep track of your alpha within your loop.
Final thoughts
The least invasive solution might be to call force.start() as normal and instead alter the force.tick function to immediately halt the timer. Since the timer in use is a normal d3.timer it may be interrupted by returning true from the callback, i.e. from the tick method. This could be achieved by putting a lightweight wrapper around the method. The wrapper will delegate to the original tick method, which is closed over, and will return true immediately afterwards, whereby stopping the timer.
force.tick = (function(forceTick) {
return function() { // This will be the wrapper around tick which returns true.
forceTick(); // Delegate to the original tick method.
return true; // Truth hurts. This will end the timer.
}
}(force.tick)); // Pass in the original method to be closed over.
As mentioned above you are now on your own managing the decreasing value of alpha to control the slowing of your layout's movements. This, however, will only require simple calculus and a loop to set alpha and call force.tick as you like. There are many ways this could be done; for a simple showcase I chose a rather verbose approach:
// To run the computing steps in our own loop we need
// to manage the cooling by ourselves.
var alphaStart = 0.1;
var alphaEnd = 0.005;
var alpha = alphaStart;
var steps = n * n;
var cooling = Math.pow(alphaEnd / alphaStart, 1 / steps);
// Calling start will initialize our layout and start the timer
// doing the actual calculations. This timer will halt, however,
// on the first call to .tick.
force.start();
// The loop will execute tick() a fixed number of times.
// Throughout the loop the cooling of the system is controlled
// by decreasing alpha to reach the freezing point once
// the desired number of steps is performed.
for (var i = 0; i < steps; i++) {
force.alpha(alpha*=cooling).tick();
}
force.stop();
To wrap this up, I forked Mike Bostock's Block to build an executable example myself.
You want a Static Force Layout as demonstrated by Mike Bostock in his Block. The documentation on force.tick() has the details:
# force.tick()
Runs the force layout simulation one step. This method can be used in conjunction with start and stop to compute a static layout. For example:
force.start();
for (var i = 0; i < n; ++i) force.tick();
force.stop();
As you have experienced yourself you cannot just dismiss the call to force.start() . Calling .start() will do a lot of initialization of your nodes and links data (see documentation of nodes() and links()). The force layout won't run without it. However, this will not start the force right away. Instead, it will schedule the timer to repeatedly call the .tick() method for asynchronous execution. It is important to notice that the first execution of the tick handler will not take place before all your current code has finished. For that reason, you can safely create your own render loop by calling force.tick().
For anyone interested in the gory details of why the scheduled timer won't run before the current code has finished I suggest thoroughly reading through:
DVK's answer (not the accepted one) to "Why is setTimeout(fn, 0) sometimes useful?".
John Reisig's excellent article on How JavaScript Timers Work.
(I need a process.nextTick equivalent on browser.)
I'm trying to get the most out of javascript performance so I made a simple counter ...
In a second I make continuous calls to a function that just adds one to a variable.
The code: codepen.io/rafaelcastrocouto/pen/gDFxt
I got about 250 with setTimeout and 70 with requestAnimationFrame in google chrome / win7.
I know requestAnimationFrame goes with screen refresh rate so, how can we make this faster?
PS: I'm aware of asm.js
Well, there's setImmediate() which runs the code immediately, ie as you'd expect to get with setTimeout(0).
The difference is that setTimeout(0) doesn't actually run immediately; setTimeout is "clamped" to a minimum wait time (4ms), which is why you're only getting a count of 250 in your test program. setImmediate() really does run immediately, so your counter test will be orders of magnitude higher using it.
However you may want to check browser support for setImmediate -- it's not available yet in all browsers. (you can use setTimeout(0) as a fallback of course though, but then you're back to the minimum wait time it imposes).
postMessage() is also an option, and can achieve much the same results, although it's a more complex API as it's intended for more doing a lot more than just a simple call loop. Plus there are other considerations to think of when using it (see the linked MDN article for more).
The MDN site also mentions a polyfill library for setImmediate which uses postMessage and other techniques to add setImmediate into browsers that don't support it yet.
With requestAnimationFrame(), you ought to get 60 for your test program, since that's the standard number of frames per second. If you're getting more than that, then your program is probably running for more than an exact second.
You'll never get a high figure in your count test using it, because it only fires 60 times a second (or fewer if the hardware refresh frame-rate is lower for some reason), but if your task involves an update to the display then that's all you need, so you can use requestAnimationFrame() to limit the number of times it's called, and thus free up resources for other tasks in your program.
This is why requestAnimationFrame() exists. If all you care about is getting your code to run as often as possible then don't use requestAnimationFrame(); use setTimeout or setImmediate instead. But that's not necessarily the best thing for performance, because it will eat up the processor power that the browser needs for other tasks.
Ultimately, performance isn't just about getting something to run the maximum number of times; it's about making the user experience as smooth as possible. And that often means imposing limits on your call loops.
Shortest possible delay while still being asynchronous is from MutationObserver but it is so short that if you just keep calling it, the UI will never have chance to update.
So trick would be to use MutationObserver to increment value while using requestAnimationFrame once in a while to update UI but that is not allowed.
See http://jsfiddle.net/6TZ9J/1/
var div = document.createElement("div");
var count = 0;
var cur = true;
var now = Date.now();
var observer = new MutationObserver(function () {
count++;
if (Date.now() - now > 1000) {
document.getElementById("count").textContent = count;
} else {
change();
}
});
observer.observe(div, {
attributes: true,
childList: true,
characterData: true
});
function change() {
cur = !cur;
div.setAttribute("class", cur);
}
change();
Use postMessage() as described in this blog.
The problem is as such:
In a js and asm.js based multiplayer game I've got two loops.
One handles the actual game ticks, like unit position, velocity and combat.
The other handles rendering of this world onto the canvas for the user to see.
What I'd like to happen is when the processor/GPU(they made those the same thing on some machines now, can't say I'm happy about that) gets encumbered too much the rendering loop should skip and thus stop changing the canvas. I.e. freezing the game screen in a lag pike.
Meanwhile the little processing power left is used to successfully complete the actual game tick preventing de-synchronisation with other game clients.
(It's an RTS-like game when it comes to load so the user input instead of positions of all objects are sent over the net).
Failing this the client would have to be kicked by the other clients or all clients would have to pause for him to reconnect and resync. i.e. bad bad bad!
A sloppy makeshift way to do this would probably be by using timestamps and terminate the graphic loop if it won't be complete by a certain time. One would presumably do this by determining max execution time for the packet types on the stack of the loop and immediately terminate the loop if the "time to execute value" of all packets together is too great to be dealt with within the resource capacity the timestamps are indicating by slowdown measurement. Hell, maybe that's radical but perhaps even skip-terminating the graphic loop when any slowdown is detected just to be sure to avoid desync.
So priorotizing one loop over another(both handling ticks) and making the second one skip if a shortage in resource is detected to ensure the first one always completes it's tick within each timeframe(10 ticks per second here).
Any possibilities or best practice methods you guys can inform me on?
EDIT: Please focus on the ability to measure availability of cpu resources and the skipping/termination for one tick of the graphic loop if these resources would not available enough to finish both loops (i.e. if the loops won't finish in the 100ms timeframe after which the next loop tick should already be firing, don't start/terminate the graphics loop).
One solution would be to use a web worker to do your world update loop and then the normal javascript loop to do the render. You would need to hand the state back and forthright o /from the web worker but the render loop would only draw on the updated data.
An advantage is that you could still have reactionary display code on the main ui loop.
This also have the advantage of the fact that the web worker could be using a different core and with multiple web workers you could use multiple extra cores
Fo the logical loop, i would take setInterval, and for the paint - requestAnimationFrame. And even more - the callback at requestAnimationFrame also receives a timestamp, so you can track timestamps and skip single frame if some lack appear.
the processor is able to handle other tasks while also rendering the animation
This statement is wrong - processor can handle only one task, and requestAnimationFrame is not actually the Rendering, it is your callback - generic javascript. You can think about it like a setTimeout. The only difference is, that it tries to run the callback on next free framerate's frame. That's why it is much better than setTimeout. So for the animations you must use the requestAnimationFrame. Other good part about it is, when the webpage is in background(other tab opened). Then the callback wont be called until it comes to the foreground. This saves processor time, as nothing is calculated in that callback.
Going back to your question: You now but, that only one callback can be processed in a time, so if the processor is in a particular time busy with the logical function, then the callback of the animation loop won't be fired. In that case it calls 'lag'. But as I understood, it is actually the desired behavior - to give the logical callback function more time. But there is other side. What if your animation function is busy, when the time for logical function came to be fired? In this case it will be fired only when animation function ends. There is nothing to do about it. If your animation function is 'heavy', you could only try to split it for 2 frames. One frame - prepare everything for render, the second one - render.
But anyway, you never become millisecond-perfect interval or timeout in javascript. As it want be called until event-loop is not free. To get the idea:
var seconds = 0;
setInterval(function(){ seconds++; var x = 10e8; while(--x); }, 1000);
Depends on you CPU, but after 10 seconds time, variable 'seconds' will be much less then 10.
And one more thing, if you really rely on time, then it is safer to use Date.now() to synchronize next logical tick:
var setLogicalLoop = (function(){
var _startedAt,
_stop,
_ms;
function frame(){
if (_stop === true)
return;
// calculations
var dt = Date.now() - _startedAt,
diff = dt % _ms;
setTimeout(frame, ms - diff);
};
return function (callback, ms){
_startedAt = Date.now();
_stop = false;
setTimeout(frame, ms);
return function(){
_stop = true;
};
};
});
// -> start
var stopLoop = setLogicalLoop(myFunction, ms);
// -> stop
stopLoop();
Sometimes i amaze myself by writting a piece of code which can be worst performant like the one below i guess. Code simply reads all files (files are text only) in directory and prints each line of each file on console after interval of one 1/10 thsecond. Now with below approach , if there are 1 million lines in the files , there will be 1 million setTimeout functions defined. and every 1/10th second one setTimeout will call its respective function. I was just curious how below code will impact the performance ? Is it ok to have millions of callbacks defined in nodejs env ? what are your thoughts about below piece of code.
function scanDir(dir){
fs.readdir(dir , function(err , list){
var interval = 0;
list.forEach(function(file, index){
lineReader.eachLine(dir + "/" + file, function(line, last) {
interval += 100;
(function(line){
setTimeout(function(){
console.log(line+"\n\r");
},interval)
})(line);
if (last) {
return false; // stop reading
}
});
})
})
}
You can't have millions of callbacks defined, you'll run out of space on the call stack after about 12,000 or so. This is the error that you'll get:
RangeError: Maximum call stack size exceeded
Now, you can increase the call stack size if you really need to, by using the following command line argument, but I don't think it would ever support millions of callbacks.
node --max-stack-size=val
Without really understanding what you are trying to accomplish, it's hard to suggest additional improvements in your code. If you really have a million lines, printing out the lines is going to take more than a day at 1 every 1/10 sec, why would you want to do this?