▲ 67 points
back
32 comments
I wonder how long it will be before compilers/interpreters of async-aware languages just do this by default. CPUs and low-level language compilers jump through all kinds of hoops of out-of-order execution, branch prediction, caching, parallel execution, etc.
I picture a day maybe 10 years from now where developers in most languages don't even have to think about these things. All the old-timers will still be structuring their code "as though it didn't exist" whereas the new kids will fly along without even thinking about it. Kind of like garbage collection the first few years.
Analysis of real software that has looked for the maximum theoretical implicit concurrency has found that software to have its implicit concurrency level top out very quickly, often not even making it up to 2x. You don't hear about this today not because nobody has tried it, but because it's been tried and it hasn't been found worth the bother. Amdahl's law is a real jerk sometimes.
If you want concurrent code, you're going to have to write it with an awareness of concurrency issues. There are things that languages can do to make it easier, but in terms of doing it automatically even that hasn't had a great track record. Fortress [1] took a good run at it, but ultimately didn't seem to get that far, at least relative to their ambitions.
Personally I suspect that while Haskell may not be that language today, and may not be that language ever necessarily, that the "magically concurrent" language of the future is going to end up looking a lot like Haskell, in particular, with the ability and need to consider the mathematical properties of your code for things like associativity and commutativity. It's going to take a pretty major sea change in a whole lot of attitudes before "the youngsters" are just casually reeling off group theory names while the "old fogies" are stuck going "What's a monoid again?" Not impossible, but hard to imagine how we get to there from here.
[1]: https://en.wikipedia.org/wiki/Fortress_(programming_language...
Sadly it seems like the youngsters are going in the opposite direction these days. It's the old fogies who have to keep saying, I've got a functor, I've got a monoid, I've got a monoid functor = monad.
Aren't Applicative Functors still all the rage? Monads are not all that well suited to automatic parallelization (as Simon Marlow points out, too).
Applicative seems to be the way of the future, yes.
Promise pipelining is an optimization of that sort: https://capnproto.org/rpc.html which was in the E language around 20 years ago: http://erights.org/elib/distrib/pipeline.html
Even in Haskell you still have to write some boilerplate to get this by default. Languages that don't separate pure computation from IO (and other side-effects) make it even harder on the compiler.
So to answer your question: implementations will do this by default only after pure/constant will become the default for all functions/variables, with side-effecting/mutable clearly marked.
If one looks at it from a 'what a today's compiler can do', then sure, one needs to statically declare everything.
If one looks at things from a 'any technical task can be handled by machine learning, sooner or later' point of view, there seems to be no obvious reason why parallelization, which is a purely technical task, not like say, writing music, could not sooner or later be handled by machine learning algorithms?
Writing music is a pretty technical task, actually. Especially if you optimize for something as tangible as: please human ears, or "fit within the actions in the movie / computer game".
The idea here is that we want to find a way to make automatic parallelization of computer programmes happen as much as possible long before we've solved general AI.
So even in a language like Haskell you still have to pay some attention to make your code amenable to parallelism. See Guy Steele's advice (see https://vimeo.com/6624203).
It's unsafe for a compiler to do this in general (i.e. without annotations) because it can't determine dependencies that are external to the program -- e.g. `one=get(); two=get()`.
The dependency between one and two is not obvious to a compiler when IO is involved, so it has to assume the two `gets' has to be executed sequentially.
> It's unsafe for a compiler to do this in general (i.e. without annotations) because it can't determine dependencies that are external to the program
As I understood the talk, Haxl doesn't address this either. It depends on you describing an IO operation as a type with set of functionality that can be used to identify those dependencies.
> e.g. `one=get(); two=get()`. The dependency between one and two is not obvious to a compiler when IO is involved, so it has to assume the two `gets' has to be executed sequentially.
A "pure" language will address a lot of this due to referential transparency. Which is to say that if `get` takes no arguments, its instructions for where and how to perform an IO operation are entirely static. Given your literal example, the only way the result could be different is if there were some side effect or if a source of data required by `get` could change during execution time.
Haxl's takes this a bit further by suggesting that if you are performing the same IO request twice "at the same time", you expect to get the same result both times, so it memoizes the request for the duration of your set of IO operations.
Within a `do`, as far as I could grok from the talk and glancing at documentation; it's worth noting I don't know Haskell and have never heard of Haxl before today.
Doh, my head-slap moment of the day. I totally overlooked that.
CPUs already have multiple concurrent instructions running. When people write software that runs serially on a macro scale, that is exactly what they get.
Great talk.
For more on concurrency and parallelism in Haskell, check out Parallel and Concurrent Programming in Haskell [0], deemed as the best book on the subject, also written by Simon Marlow.
Haxl is a powerful abstraction with IMHO a beatifuly simple implementation.
However for our use case at LumiGuide (reading and writing registers of modbus devices) it wasn't simple enough. We just needed an abstraction for batching and did not need caching and the other features Haxl provides.
So I wrote monad-batcher which as the name implies only provides a batching abstraction (which can also be used to execute commands concurrently). All the other features can be build on top of monad-batcher as separate layers (separation of concerns).
The library is available on Hackage but needs a bit more documentation (a tutorial would be nice):
I looked through the slides but not the video and the slides ignore the hard problem: how do you schedule these requests? How do you know how many parallel requests you can issue without hammering the database or service? How do you batch queries so that you get acceptable latency and a query size that will not choke the database?
The last question is probably easy for most use cases where you have independent requests coming in (typical web application) - in the context of a single request you can usually get away with batching as much as is possible. But the scheduling problem is very similar to the promises of "free parallelism because Church-Rosser" - actually taking advantage is an open problem. Even when you know how much time each job takes in advance, multiprocessor scheduling is NP-hard.
Anyway, if someone watched the video and the question is addressed there, please let me know so I can watch it.
I think the idea is that with the Haxl approach the scheduling can be dealt with independently from your business logic.
By the way, I don't see how Church-Rosser would give you any free parallelism---even in theory. You'd still have to heed Guy Steele's advice (see https://vimeo.com/6624203).
> By the way, I don't see how Church-Rosser would give you any free parallelism---even in theory.
Church-Rosser theorem means that all possible reduction sequences lead to the same normal form, so you can β reduce the subterms in parallel.
If you look at Paul Hudak's and SPJ's publications from the late 1980s and early 1990s many of them are actually about trying to exploit this implicit parallelism:
http://sunsite.informatik.rwth-aachen.de/dblp/db/indices/a-t...
http://dblp.uni-trier.de/pers/hd/h/Hudak:Paul
https://pdfs.semanticscholar.org/8912/5be7f9c222793c18b99d06...
This turns out to be a hard problem and fell out of fashion as a research topic. The big problem is managing overhead. There has been some recent work to try to incorporate automatic profiling feedback to adjust the parallelism granularity:
http://dominic-mulligan.co.uk/wp-content/uploads/2015/05/S-R...
Yes, but you still need to have those parallel subterms in the first place!
If you eg still use a linked list data structure as an input, you are never gonna be faster than O(n) no matter how clever your implicit parallelism is.
> If you eg still use a linked list data structure as an input, you are never gonna be faster than O(n) no matter how clever your implicit parallelism is.
I think you took away completely the wrong point from Steele's talk. The bottleneck is not in the data structures, it is in the control flow, which is just a trivial way of restating Amdahl's law. In fact Steele co-wrote an excellent expository paper with Daniel Hillis, while they were at Thinking Machines, showing how to do parallel processing on linked lists:
http://cva.stanford.edu/classes/cs99s/papers/hillis-steele-d...
I was in the audience and asked questions, and chatted with him afterwards..
Data flow dictates control flow (to a certain extent).
The paper you linked to is very interesting: but even there they have to augment their 'linked list' with more links. And they assume one processor per list element and don't count the set-up time of having each processor find its list element.
(There's some more complication to it. But I am glad between 1986 when the paper was written and 2009 when he gave the talk, he figured out how to explain some of these issues more simply.)
I have not fully digested yet, but seems very similar to Scala Parallel Collections and Java8 Streams. There are databases which implements such interfaces.
It isn't the slightest bit similar to those! Haxl is a high-level library for specifying (and optimising) concurrent data retrieval.
I'm not sure that's an apt description either. Haxl's features include auto batching of IO through parallelism and a temporal cache...for error logging when failures occur. It's spelled out in the first minute he talks about it. In the IO batching code (the boilerplate) you specify how and what to output in your logging. Haxl can be used without performing optimization (write an inefficient batcher) or doing data retrieval (the I/O is fire and forget and never cache anything). Maybe I am misunderstanding the basic usage.
It's not easy summarising it in one sentence, but 'auto batching' and 'caching' is what I meant by optimising the data retrieval.
EDIT: the Haxl docs describe it as "a library and EDSL for efficient scheduling of concurrent data accesses with a concise applicative API"
Which databases are those and what are the relevant features? Are you talking about FILESTREAM? https://technet.microsoft.com/en-us/library/bb933993(v=sql.1...
Hazelcast Jet, Spark, my project... they work similar way. But as I wrote, I need to read the presentation first.
Jet and Spark are not what most people would call databases, since they have such a heavy stream processing focus and are rarely stores of record.
How does this differ from BEAM langs which already make concurrency "unreasonably easy"?
For one, they serve different purposes. Haxl is specifically for concurrent data retrieval while BEAM is a general purpose platform for fault-tolerant computation.
For another, they operate via different interfaces. BEAM languages communicate concurrently only via a message passing interface. Haxl lets the author write declarative code specifying what to retrieve, then the library executes it concurrently and in parallel under-the-hood.
Both are properly described as "unreasonably easy" - just for different sorts of things, on different platforms, with different interfaces.
Thank you for the breakdown!
Cheers!