Ask HN: On Rob Pike's Concurrency is not Parallelism?
But there isn't much detail in slides on how to get parallelism once you've achieved concurrency. In the Lesson slide (num 52), he says Concurrency - "Maybe Even Parallel". But the question is - When and How can concurrency correctly and efficiently lead to Parallelism?
My guess is that, under the hood Rob's pointing out that developers should work at the level of concurrency - and parallelism should be language's/vm's concern (gomaxprocs?). Just care about intelligent decomposition into smaller units, concerned only about correct concurrency - parallelism will be take care by the "system".
Please shed some light.
Slides: http://concur.rspace.googlecode.com/hg/talk/concur.html#title-slide HN Discussion: http://news.ycombinator.com/item?id=3837147
70 comments
[ 2.9 ms ] story [ 145 ms ] threadIn Go, it is easy to create multiple tasks/workers each with a different job. This is implicitly parallelizable - each task can (but doesn't have to) work within their own thread. The only time when the workers can't run in parallel is when they are waiting on communication from another worker or outside process.
This is opposed to data level parallelism where each thread is doing the nearly exactly same instructions on different input, with little to no communication between the threads. An example would be to increase the blue level on each pixel in an image. Each pixel can be operated on individually and be performed in parallel.
So - the push is for more task-based parallelism in programs. It is very flexible in that it can run actually in parallel or sequentially and it won't matter on the outcome of the program.
Still, I think it's clear that an x264 with N threads per frame with no per-pixel SIMD would lose to the current x264 with N threads per frame. The key is that x264 is making good use of task parallelism and data parallelism.
Our current hardware tends to offer great data parallelism for homogenous task queues and task parallelism for heterogenous task queues. Given that, we task parallelism needs a lot more consideration from a human. It's also the case that our current implementations of data parallelism tend to focus on shared memory computations, so their scope is a lot more limited than the distributed-system-conflated discipline of task-oriented concurrency, where we're currently having an explosion of engineering.
I've mentioned before that I don't think SIMD is going away anytime soon. It has so many upsides (cache locality, simple implementation in hardware due to the single-instruction nature of it) that I think designs that don't take advantage of it will always be at a disadvantage for the foreseeable future.
I think that is about what you are saying.
Concurrency is when you write your program using multiple threads. Your program looks and means vastly different if you use threads.
You use concurrency not for performance gain, but for clarity of your program. You use parallelism for performance gain, to utilize all your processors.
You can expand the term "thread" to mean "concurrency in general" but even then it isn't true that the main purpose of writing concurrent code is clarity. When people say "look at this concurrent program I wrote" they rarely [1] say "look at how well the code expresses the problem". What they overwhelmingly say is "look at this benchmark".
[1] Joe Armstrong talks about how Erlang lets you represent processes more like they happen in the real world. But that's a niche view. Most people think about concurrency as a platform issue, where the platform is multicore hardware or distributed systems, and otherwise wouldn't bother with it.
(future (do (foo 1) (bar 2)) Runs the expression on another thread. Futures intentionally compose well with the built-in clojure concurrency tools.
Clojure has no dislike of threads. It scorns locks, and code that is safe in one thread, but unsafe in multi-threaded situations.
The goal of concurrency is to model a problem that is easier or better or more natural to model concurrently (that is, different parts are running simultaneously). For example, if you are simulating agents in some virtual world (eg a game), then it may make sense that these agents are being modeled in a way that they are all running simultaneously and the processing of one does not block the processing another. This could be done by timeslicing available processing between each agent (either by using the processor/OS pre-emptive multitasking, if available, or through cooperative multitasking[1]), or is could be done by physically running multiple agents at the same time on different processors or cores or hardware threads (parallelism). The main point is that concurrency may be parallel, but does not have to be and the reason you want concurrency is because it is a good way to model the problem.
The goal of parallelism is to increase performance by running multiple bits of code in parallel, at the same time. Concurrency only calls for the illusion of parallelism, but parallelism calls for real actual multiple things running at the exact same time and so you must have multiple processors or cores or computers or whatever hardware resources for parallelism, while concurrency can be simulated on single core systems. Parallel code is concurrent code, but concurrent code is not necessarily parallel code.
Distributed programming is parallel programming where the code is running in parallel, but distributed over multiple computers (possibly over the internet at distant locations) instead of running locally on one multi-core machine or a HPC cluster.
From stackoverflow[2]:
As for when can concurrency be turned into parallelism, that depends. Assuming that the hardware resources exist (or that it simply falls back to time-sliced concurrency if they do not), parallelism can be achieved if there are multiple things that can execute independently. There are at least three types of parallelism and if your problem, code and/or data fit one of these, then your concurrent code may be executed in parallel.1. You have a number of items of data and each one can be processed independently and at the same time. This is the classic embarrassingly parallel data parallelism. For example, you have a large array of input data and an algorithm that needs to run on each one, but they do not interact. Calculating pixel colours on your screen, or handling HTTP requests, for example.
2. You have two or more independent tasks doing different things that are run in parallel. For example, you have one thread handling the GUI and another thread handling audio. Both need to run at the same time, but both run independent of each other with minimal communication (which can happen over a queue, perhaps).
3. Sometimes you have a stream of data which must be processed by a number of tasks one after the other. Each task can be run in parallel so that if you have a stream of data, item[0], item[1], item[2], etc (where 0 is first in the stream, 1 is next and so on) and a number of tasks that need to run in order: A, B, C - then you can run A, B and C in parallel such that A processes item[0] while B and C are idle, then B processes item[0] an A processes item[1] and C is idle. Then C processes item[0] while B processes item[1] and A processes item[2] and so on. This is called pipelining and as you probably know is a very common technique inside processors.
Of course, all three can be combined.
[1] Could be...
Two operations are concurrent if they have no causal dependency between them.
That's it, really. f(a) and g(b) are concurrent so long as a does not depend on g and b does not depend on f. If you've seen special relativity before, think of "concurrency" as meaning "spacelike"--events which can share no information with each other save a common past.
The concurrency invariant allows a compiler/interpreter/cpu/etc to make certain transformations of a program. For instance, it can take code like
and generate ... perhaps because b becomes available before a does. Both programs will produce identical functional results. Side effects like IO and queue operations could strictly speaking be said to violate concurrency, but in practice these kinds of reorderings are considered to be acceptable. Some compilers can use concurrency invariants to parallelize operations on a single chip by taking advantage of, say, SIMD instructions or vector operations: Or more often: where f could be something like "multiply by 2".Concurrency allows for cooperative-multitasking optimizations. Unix processes are typically concurrent with each other, allowing the kernel to schedule them freely on the CPU. It also allows thread, CPU, and machine-level parallelism: executing non-dependent instructions in multiple places at the same wall-clock time.
In practice, languages provide a range of constructs for implicit and explicit concurrency (with the aim of parallelism), ranging from compiler optimizations that turn for loops into vector instructions, push matrix operations onto the GPU and so on; to things like Thread.new, Erlang processes, coroutines, futures, agents, actors, distributed mapreduce, etc. Many times the language and kernel cooperate to give you different kinds of parallelism for the same logical concurrency: say, executing four threads out of 16 simultaneously because that's how many CPUs you have.What does this mean in practice? It means that the fewer causal dependencies between parts of your program, the more freely you, the library, the language, and the CPU can rearrange instructions to improve throughput, latency, etc. If you build your program out of small components that have well-described inputs and outputs, control the use of mutable shared variables, and use the right synchronization primitives for the job (shared memory, compare-and-set, concurrent collections, message queues, STM, etc.), your code can go faster.
Hope this helps. :)
No one is making "excuses". It's important to understand these problems. Not understanding concurrency, parallelism, their relationship, and Amdahl's Law is what has Node.js in such trouble right now.
this is the sanest and most pragmatic way server a web server from multiple threads
it's all serialization - but that's not a bottleneck for most web servers.
I disagree, especially for a format like JSON. In fact, every web app server I've dug into spends a significant amount of time on parsing and unparsing responses. You certainly aren't going to be doing computationally expensive tasks in Node, so messaging performance is paramount.
i'd love to hear your context-switching free multicore solution.
I claimed no such thing: only that multiprocess IPC is more expensive. Modulo syscalls, I think your best bet is gonna be n-1 threads with processor affinities taking advantage of cas/memory fence capabilities on modern hardware.
this is the sanest and most pragmatic way server a web server from multiple threads
What is this I can't even.
Node.js: https://gist.github.com/3200829
Clojure: https://gist.github.com/3200862
Note that I picked the really small messages here--integers, to give node the best possible serialization advantage.
Note the high sys time: that's IPC. Node also uses only 75% of each core. Why? 96 context switches per second.Compare that to a multithreaded Clojure program which uses a LinkedTransferQueue--which eats 97% of each core easily. Note that the times here include ~3 seconds of compilation and jvm startup.
Why is this version over 3 times faster? Partly because it requires only 4 context switches per second. And queues are WAY slower than compare-and-set, which involves basically no context switching: TL;DR: node.js IPC is not a replacement for a real parallel VM. It allows you to solve a particular class of parallel problems (namely, those which require relatively infrequent communication) on multiple cores, but shared state is basically impossible and message passing is slow. It's a suitable tool for problems which are largely independent and where you can defer the problem of shared state to some other component, e.g. a database. Node is great for stateless web heads, but is in no way a high-performance parallel environment.Also Node starts up in 35ms and doesn't require all those parentheses - both of which are waaaay more important.
A great example of this is resque. It'd be great if we could have multiple resque workers per process per job type. This would save a ton of resources and greatly improve processing for very expensive classes of jobs. It's a very real-world consideration. But instead, because our architecture follows this "share nothing in my code, pass that buck to someone else" model like a religion, we waste a lot of computing resources and lose opportunities for better reliability.
What I find most confusing about this argument is that I challenge you to find me a website written in your share nothing architecture that, at the end of the day, isn't basically a bunch of CRUD and UX chrome around a big system that does in-process parallelism for performance and consistency considerations. Postgresql, MySQL, Zookeeper, Redis, Riak, DynamoDB ... all these things are where the actual heavy lifting gets done.
Given how pivotal these things are to modern websites, it's bizarre for you to suggest it is not something to consider.
Web serving is OK & all, but I'd love if node could be an ideal runtime for petri-nets and webworker meshes too.
In addition to that, processes severely inhibit the usefulness of in-process caches. Where threads would allow a single VM to have a large in-process cache, processes generally prevent such collaboration and mean you can only have multiple, duplicated, smaller in-process caches. (Yes, you could use SysV shared memory, but that's also fraught with issues)
The same goes for any type of service you would like to run inside a particular web server that could otherwise be shared among multiple threads.
But you say things like this and it worries me. Because a lot of people look up to you and either you said this because you feel defensive about your project or you said it because you genuinely don't understand the cases we're talking about here. And this is a problem because a lot of people look up to you and what you say, so when you say something as baffling as this response, you run the risk of leading a lot of people astray.
I was sort of at a loss for how to reply in the time I have to spare for Hacker News, but thankfully Aphyr did for me. But let me clarify what I said a bit, since I was a bit terse: The problem Node.js has is a social one. A lot of node hackers take the stance, "I thought Node.js solved the problems threads presented," (https://groups.google.com/d/msg/nodejs/eVBOYiI_O_A/kv6iiDyy9...) like there is a single axis of superiority and Node.js sits above the methods that came before. But the reality is that Node.js is really just another possible implementation in the Ruby/Python/Pike/-Perl-¹ space, and shares most of the same characteristics as those languages.
So you have a lot of people who are aces at front-end programming in the browser thinking they have a uniformly superior tool for tackling high-performance server problems, but really they don't have that; they just have a tool with familiar syntax. And so they fearlessly (and perhaps admirably) charge into the breech of platform programming without realizing that the way people scale big projects involves a lot of tools, a lot of thought about failure modes, and a lot of well-established algorithms with very specific tradeoffs.
And so this is Node.js's problem. It's just another gun in the gunfight, but its community thinks its a cannon. In a world where high-performance parallel VMs like Java or Erlang have very powerful and helpful languages like Clojure or Scala on top, we're in a funny situation. It becomes increasingly difficult to justify all these GIL-ridden implementations of languages.
Which is not to say these implementations don't have their place (and Node.js is hardly the first javascript implementation outside of a browser), but increasingly they are losing their place in the pieces of your code expected to shuffle data around efficiently in the backend of modern distributed applications.
¹ Correction, perl 6 doesn't plan share this behavior. What I read suggests it's not done yet.
Yet Node itself can not and should not solve the deployment problem: node is a javascript runtime, and contrary to earlier claims I'd declare not opinionated, not one to make this decision for us. The scaling out story is indeed not easy: even tasks like validating session credentials need to be mastered by the team (persist creds into a data-store, or use group communication/pubsub: building for Node is a lot like building for multi-machine Java). The level of DIY-it-ness here are indeed colossal.
What I'd contrast against your view- and I agree with most of your premise, that node is extremely seductive and dangerous and many are apt to get in way way way over their head- is that the comforts you describe are what kill these other languages, what strange and prevent us from becoming better more understanding programmers. Ruby, python, php, less so perl, the webdev that goes on there happens by and large at extreme levels of abstraction: developers flock to the known explored center, the tools with the most, the places that seem safest.
The dangerous dangerous scenario presented by most good web development tools is that it is the tools that know how to run things. Contrary to the charge into the breech throw up ad-hoc platforms in production every day mentality (of node), these (ruby, php, python) platforms stagnate, they fall to the ruin as their tooling strives towards ever reaching greater heights: the tools accrue more and more responsibility, there are better carved out & expected ways to do things, and incidental complexity, the scope of what must be known, how far one has to travel, to get from writing a page to it getting shipped over the wire or executing, balloons.
If anything, Node's core lesson to the world has been about how much is not required. Connect, the only & extremely extremely low-lifed common denominator of Node web world, is the meager-est, tiniest smallest iota of a pluggable middleware system (if only Senchalabs had been courteous enough to be more up front about it being a complete and total rip off Commons-Chain & to not add a thing, I would not bloody loath it). That pattern? bool execute(Context context). Did you handle this request? No? Ok, next. You need to deploy a bunch of processes on a bunch of boxes? You an probably write up something perfectly adequate in a week. Don't have a week? Go find a module: certainly Substack has at least one for whatever your cause (here it's Fleet, https://github.com/substack/fleet).
Node modules are wonderful. They all have some varyingly long list of dependencies, usually the tree is 4-8 different things, but the total amount of code being executed from any given module is almost always short of a couple dozen KB: your engineering team can come in and understand anything in a day or three, and gut it and rebuild it in another day or two. Modules, unlike how development processes have shaped up in hte past decade, are wonderfully delightfully stand-alone: there are no frameworks, no crazy deployment systems, no bloody tooling one is writing to: it's just a couple of functions one can use. The surface area, what is shown, is tiny, is isolated, is understandable, there's no great deep mesh. This runs so contrary to the Drupal, to the Rails, to the Cakes or Faces of the world where one is not writing a language, they're at the eight degree of abstraction writing tools for a library that implements enhancements for a framework that is a piece of an ioc container that runs on a application server that...
I have a lot of problems with Node myself - but the single event loop per process is not one of them. I think that is a good programming model for app developers. I love Go so much (SO MUCH), but I cannot get past the fact that goroutines share memory or that it's statically typed. I love Erlang but I cannot get the past the syntax. I do not like the JVM because it takes too long to startup and has a bad history of XML files and IDE integration - which give me a bad vibe. Maybe you don't care about Erlang's syntax or static typing but this is probably because you're looking at it from the perspective of an engineer trying to find a good way to implement your website today. This is the source of our misunderstanding - I am not an app programmer arguing what the best platform to use for my website--I'm a systems person attempting to make programming better. Syntax and overall vibe are important to me. I want programming computers to be like coloring with crayons and playing with duplo blocks. If my job was keeping Twitter up, of course I'd using a robust technology like the JVM.
Node's problem is that some of its users want to use it for everything? So what? I have no interest in educating people to be well-rounded pragmatic server engineers, that's Tim O'Reilly's job (or maybe it's your job?). I just want to make computers suck less. Node has a large number of newbie programmers. I'm proud of that; I want to make things that lots of people use.
The future of server programming does not have parallel access to shared memory. I am not concerned about serialization overhead for message passing between threads because I do not think it's the bottleneck for real programs.
I also thought this. For some reason though the more I use it the more I am coming around to its syntax. There are lots of gotchas when compared with other popular languages, but still its syntax has grown on me recently.
At the opposite end of the syntactic spectrum is Lisp. The more I use Clojure the more I am loving it as well.
"The future of server programming does not have parallel access to shared memory."
I agree. The future of server programming is also not spawning multiple child processes. The question I have is how far off is that future? I know that most of my web applications today simply run in multiple spawned OS processes.
What about STM in Clojure? It's technically parallel access to shared memory, but the transactional nature obviates the need for mutexes and all the crap that makes shared memory a pain in the ass.
Barring a massive shift in hardware architectures, shared access by cooperating threads is your only option for high-performance shared state. Look at core counts vs core flops for the last five years. Look at Intel's push for NUMA architectures. This is a long-scale trend forced by fundamental physical constraints with present architectures, and I don't see it changing any time soon.
Anyone telling you shared state is irrelevant is just pushing the problem onto someone else: e.g., a database.
I cannot understand this hangup about syntax. Syntax is the easiest thing to learn with a new language, you just look it up. It's the semantics and usage where the real problems are.
Erlang doesn't have static typing, it is dynamically typed. It has always been dynamically typed.
And however you look at it doing highly concurrent systems with processes is much easier. And who says that processes imply "parallel access to shared memory"? Quite the opposite actually.
Really the only really tricky thing for a newbie is strings.
[1] as larry wall famously said, "lisp has all the visual appeal of oatmeal with fingernail clippings mixed in."
[2] http://blog.opalang.org/2012/02/opa-090-new-syntax.html
I'm not the author of the comment above, but I think Erlang's syntax is effective in that it strongly emphasizes computation by pattern matching. If you write very imperative code in it (as people tend to, coming from Ruby or what have you), yes, it will look gnarly. Good Erlang code looks qualitatively different. There are pretty good examples of hairy, imperative Erlang code being untangled in this blog post: http://gar1t.com/blog/2012/06/10/solving-embarrassingly-obvi...
The . , ; thing is a bit of a hack, admittedly -- I suspect that comes from using Prolog's read function to do parsing for the original versions of Erlang (which was a Prolog DSL), and reading every clause of a function definition at the same time. Prolog ends every top-level clause with a period, not "; ; ; ; .". (Not sure, but a strong hunch, supported by Erlang's history.) I got used to it pretty quickly, though.
It is very simple really, think of sentences in English and it all becomes trivially simple.How many sentences end in a ';'?. And you almost never need to explicitly specify blocks.
Here http://ferd.ca/on-erlang-s-syntax.html are some alternate ways of looking at it.
(Hello, Robert! :) )
Anyways, my experience is that all languages will get people criticizing them. And, in my experience, those kinds of criticisms should almost always be categorized as "does not want to talk about language FOO" and a proper response is probably something like "if you don't want to give that subject the respect it deserves, let's change the subject to something you find interesting".
Syntax benefits are not all about subjectivity. Anyone claiming this has effectively lost the plot and decided to go turtle in the discussion.
There is no getting around the difference in semantics. For this I think that having a different syntax is actually better. Also there are things in Erlang which are hard to fit syntactically into the syntax of OO languages, for example pattern-matching.
http://www.youtube.com/watch?v=zY-IueSMAPc
http://elixir-lang.org/
My favorite: Joxa a Clojure inspired (really just inspired ;)) lisp http://joxa.org/
Then there are Elixir (mentioned already) and Reia, Both seem to be inspired by Ruby. http://reia-lang.org/ http://elixir-lang.org/
Erlang has a far superior computational model an implementation than Node, it can handle way more requests faster and is newcomer friendlier as any web request is simply a message received.
Programming is a verb describing what people do with programming systems, and we're nowhere near the point where it can be done automatically yet. You cannot remove people from the equation and claim to be interacting with it.
It really kinda sucks for them right now.
Serious hard-core engineers that need serious tools are actually pretty well served by current tools. No, no, they're not perfect. But we're way better off than somebody who's just beginning in terms of what tools are aimed at us.
Personally I like static typing, especially if somewhat optional, it's something we effectively do through documentation anyway (via JSdoc or similar), but makes it concrete.
I dont think light-weight threads sharing memory is so bad, symmetric coroutines are more or less the same as an event loop IMO, the thought put into working with them is more or less identical, just without callback hell and odd error-handling, but I suppose going all-out with message passing could be fine. I think that's still a bit of an implementation detail unless you get rid of the concept of a process all together and start just having a sea of routines that talk to each other.
Thank you for the nice write up.
https://github.com/halayli/lthread
lthread supports concurrency and parallelism using pthreads. Each lthread scheduler runs its own lthreads concurrently, or better said, one at a time. But from an observer's perspective they look like they are running in parallel.
Now if you create 2 pthreads on a 2 core machine and each runs an lthread scheduler then you have true parallelism because you can have 2 lthreads running in parallel at the same time. One by each scheduler.
I feel this is a closer context to what Rob is discussing than what I found in the comments here.
Concurrency is a property of a program's semantics, usually seen in a 'thread' abstraction. The most important part of concurrency is nondeterminism. Concurrency might permit parallelism depending on hardware, language runtime, OS, etc.
Parallelism is a property of program execution and means multiple operations happening at once, in order to speed up execution. A program written to take advantage of parallelism can be deterministic, but often is accomplished by way of concurrency in OS threads. Because most languages still suck.
Examples of Concurrency:
1. I surf the web And I run an installer for another program.
2. One gopher brings empty carts back, while another brings full carts to the incinerator.
The idea of concurrency is that two completely separate tasks are being done at the same time. There may be synchronization points between the two tasks, but the tasks themselves are dissimilar.
Viewed in one way moving empty wheelbarrows may be completely different from moving filled ones.
Viewed in another way, they might seem very similar.
Concurrency has to do with task parallelism.
Parallel has to do with data parallelism.
There's a gray line between the two where you can't clearly differentiate between them.