Personal tools


From HaskellWiki

Revision as of 09:41, 11 January 2007 by Simonpj (Talk | contribs)

Jump to: navigation, search

Beautiful concurrency

I am writing a chapter for a book called "Beautiful code", edited by Greg Wilson. My draft chapter is about Software Transactional Memory in Haskell.

I would welcome any comments or questions you have on the paper, or constructive suggestions for improving it; the more concrete the better.

The book is aimed at a general audience of programmers, not Haskell geeks, so I have tried to explain everything necessary as I go along. So if you are not a Haskell expert, your input would be particularly valuable to me.

You can email me directly (, or add Wiki talk notes below.

If you give your real name somewhere in your text (or email it to me), I'll add you to the acknowledgements at the end of the chapter.

Simonpj 14:26, 22 December 2006 (UTC) To add a note, begin with four tilde signs ~~~~; the Wiki will fill in your user name and date.

ArthurVanLeeuwen 16:25, 22 December 2006 (UTC) (SLPJ: done) There's a couple of typos in the paper as it is now. More importantly, the footnote on page 2 has (hGetLine h "hello") where it should state (hPutStr h "hello").

Neil Mitchell 16:28, 22 December 2006 (UTC) (SLPJ: done) Sect 1, Para 2. If we want to write parallel program[s] - missing s.

Steven807 17:14, 22 December 2006 (UTC) (SLPJ: done) There is no definition or description of check

Tibbe 18:33, 22 December 2006 (UTC); (SLPJ: done)The footnote on page 2 now has a incorrectly capitalized T in hPutSTr.

Fanf 18:51, 22 December 2006 (UTC) (SLPJ: done)page 1 para 2 "as well shall see" should be "as we shall see"

Fanf 18:51, 22 December 2006 (UTC) (SLPJ: done)page 3 "a bit like *t in C" should be "a bit like t* in C" since t is a type

Garious 18:56, 22 December 2006 (UTC) (SLPJ: done) page 10 "at which point An elf" should be "at which point an elf"

Garious 18:58, 22 December 2006 (UTC) (SLPJ: done) page 10 "Here, then is a possible" should be "Here then, is a possible"

Garious 19:16, 22 December 2006 (UTC) (SLPJ: done) page 11 "Again, we define Group is declared" whaaa? maybe: "Group is declared as a data type with constructor MkGroup. MkGroup is passed the Group's capacity, and a TVar containing the number of current members and the Group's two Gates."

Garious 19:16, 22 December 2006 (UTC) (SLPJ: done) page 11 "Creating a new Group is simply..." Is a process of three somewhat-abstract steps simple enough to say 'simply'? Instead, maybe show the code and let the reader think, "Hey, that's simple!"

Rycee 20:46, 22 December 2006 (UTC) (SLPJ: done) Page 4. You probably want to change "... the action a does ..." to "... the action act does ..."

Rycee 20:46, 22 December 2006 (UTC) (SLPJ: done) Page 8. Typographic nitpick: The space after "i.e." looks wide, perhaps you forgot to write "i.e.\ " to force a regular (non sentence ending) space in LaTeX?

MichalPalka 22:32, 22 December 2006 (UTC) You could add a reference to SQL and triggers. They are similar in that it is programming with transations and seeing familiar names will make applied programmers feel more comfortable. (SLPJ: hard to know what is "familiar" to a random reader. For a reader unfamiliar with SQL and triggers, referring to that might just make it more confusing. What do others think?)

DavidHouse 22:55, 22 December 2006 (UTC) (SLPJ: done) Page 3, "That is, return v is a function that, when performed, does no side effects and returns v." Your use of 'that is' implies that you could deduce the fact that it does no side effects from the type signature just given, which isn't true. It's an auxiliary property of 'return'. Maybe just miss out the 'that is'.

DavidHouse 23:04, 22 December 2006 (UTC)(SLPJ: done) Bottom of Page 4, "Then 'atomically act' grabs the lock, performs the action 'at',". Missing 'c' out of 'at'.

DavidHouse 23:23, 22 December 2006 (UTC) (SLPJ: done) Page 5, "Then 'withdraw' is an STM action that adds amount to the balance in the account." 1) For consistency with the rest of the paper, it should be "Then 'withdraw account ammount'..." 2) withdraw subtracts, not adds, to the amount in the account.

DavidHouse 23:40, 22 December 2006 (UTC) (SLPJ: done at defn of hEchoLine) Page 9, you probably want to mention that ++ is string concatenation.

DavidHouse 23:45, 22 December 2006 (UTC) (SLPJ: done)Page 10, I would expect to see the type signatures for the Gate interface rewritten alongside their definitions.

DavidHouse 23:47, 22 December 2006 (UTC) (SLPJ: difficut to do this crisply, but I take the point) Page 9/10, algebraic datatypes are a little weird for the non-initiated, especially with constructors that don't appear to do anything ("Where do you define MkGate?" etc.). You might want to liken them to C's structs, and explain the constructors as tags?

DavidHouse 23:57, 22 December 2006 (UTC) (SLPJ: done) Page 13, I don't think you sufficiently explain 'sequence'. You may wish to add a sentence along the lines of "'sequence' is a function that takes a list of IO actions and returns the action that, when executed, runs each of the actions you passed it in turn."

Dalejordan 02:15, 23 December 2006 (UTC) (SLPJ: done) For the sake of us newbs you might mention in Section 2.1 how all these actions ever get performed. Also, in your description of nTimes I think it would be clearer to say it creates a composite action that performs its argument action n times, rather than say it performs it (directly) n times, even though the value of n is not yet known. Another example of the "beauty" of first class actions (and recursion).

Brecknell 03:29, 23 December 2006 (UTC) (SLPJ: done) Page 7: withdraw should subtract, not add amount to the balance (sadly). Also, since this withdraw has different semantics to the version on p5, you may want to consider giving it a different name.

Brecknell 03:29, 23 December 2006 (UTC) (SLPJ: done) Page 8, towards the end of section 2.4: Two instances of "withdraw" should be "withdraw2", one in the function definition and one in the comment.

Brecknell 05:04, 23 December 2006 (UTC) (SLPJ: done) In the problem definition, "Santa repeatedly sleeps until wakened...", but with the definition of "forever" given in the text, and without some sort of asynchronous signalling, it looks to me like Santa will sleep for at least as long as his random number generator tells him to. Perhaps move the sleep out of forever and into elf and reindeer, then Santa will just sleep in awaitGroup.

Mvanier 19:21, 23 December 2006 (UTC) (SLPJ: done) In section 3.3, "sequence" might as well just be "sequence_". It's confusing to return a list of unit values and not do anything with them. In section 5, I think that the problem with composing lock-based concurrent functions could be explained better; the statement "instead we must expose the locking protocol" could be expanded to show why this is the case.

ChrisKuklewicz 22:17, 24 December 2006 (UTC) In learning Haskell STM back in October of 1995 I had written a solution to the same Santa problem. The main suggestion I have is that the elf and reindeer are given two actions (use Gate inGate) and (useGate outGate) and are required to properly bracket their actual action between these two. If the (joinGroup group) call returned a single scoping combinator (bracket_ (useGate inGate) (useGate outGate)) then the concurrency model for the elf and reindeer would be even to get correct (and beautiful?). Furthermore they could be passed (joinGroup group) instead of (group). This prevents the elf and reindeer from calling any other group functions.

Conal 16:39, 25 December 2006 (UTC) (SLPJ: done) Page 11, first full paragraph, "then waits to the TVar to be decremented to zero." Replace "waits to" with "waits for".

Brecknell 03:41, 27 December 2006 (UTC) This was my first exposure to STM, and I found it clear and easy enough to follow. However, I think it is mainly the discussion of bank accounts that conveys the "beauty" of STM, since that's where the important concepts are demonstrated: composability of transactions, implicit locking and deadlock avoidance, etc. In the Santa Claus problem, those concepts are somewhat obscured. I think the Santa Claus problem is still useful for demonstrating how STM can handle a tricky concurrency scenario with relative ease, but it didn't give me the sense of clarity that the discussion of bank accounts did. Maybe the Santa Claus solution needs more focus on the abstractions possible in STM, how they can help modularise concurrent programs, and why STM is better for solving concurrency problems than the concurrency primitives available elsewhere.

Gaal 08:20, 27 December 2006 (UTC) (SLPJ: done) The binding and usage of the reindeer group on page 12 are inconsistent:

 ; rein_gp <- newGroup 9
 ; sequence [ reindeer gp n | n <- [1 .. 9]]
                    -- ^^
 ; forever (santa elf_group rein_group)

Fernando 14:11, 29 December 2006 (UTC) SLPJ: done In page 3, in the definition of incRef, the last line should probably read, writeIORef var (val+1), instead of writeIORef (val+1). The latter version won't compile.

PaulRBrown 07:11, 2 January 2007 (UTC) It would be convenient if the download link for the source code were active. (Otherwise, reviewers have to cut and paste the fragments in by hand.) SLPJ: try [1]

PaulRBrown 07:11, 2 January 2007 (UTC) I had trouble with the motivation for the Santa Claus example, and I ended up resorting to scratch paper to sketch out the problem and the model. That section would benefit from some additional introduction, perhaps a diagram. (Something schematic with "(n of m) || (p of q)" would be nice.) You could also explain it in terms of concurrency APIs that people are familiar with, e.g., count-down semaphores in Java's concurrency package. In fact, the additional development of the counting barrier prior to the Santa Claus problem would help overall clarity. SLPJ: Hmm

DavidHouse 10:30, 2 January 2007 (UTC) I actually had trouble with this one too. I think even the addition of a sentence like the following would help a great deal: "We want to model this setup in Haskell, where 'sleeping' is blocking and 'waking up' is doing something." SLPJ: where exactly? I'm not convinced this is going to help, but concrete suggestions are always useful.

EricWilligers 12:43, 4 January 2007 (UTC) Page 4: As page 14 currently contains a leap in difficulty with the definition of choose, perhaps foldr1 could be introduced before the last paragraph of 2.1, with an example call. This would also be an opportunity to introduce [ ] in a type signature - it does appear with regards to "sequence" but is easy to miss as the return value isn't discussed. A simple example would avoid (op) notation and would make no sense for empty lists. e.g. max_n values = foldr1 max_2 values

EricWilligers 12:43, 4 January 2007 (UTC) Page 10: Missing closing ) at the end of useGate

EricWilligers 12:43, 4 January 2007 (UTC) Page 12: elf_group => elf_gp, rein_group => rein_gp

EricWilligers 12:43, 4 January 2007 (UTC) Page 13: elf-gp 1 => elf_gp 1

EricWilligers 12:43, 4 January 2007 (UTC) Page 14 and page 8 claim all the code is presented. To meet this claim:- Perhaps the following can appear on p9, just before "Since IO actions are first-class, ..."

deliverToys :: String -> IO ()
deliverToys s = putStr (s ++ " delivering toys\n")

Perhaps the following can appear on p13, just before "Working inside-out, ..."

reindeer :: Group -> Int -> IO ()
reindeer gp id = forkIO (forever (reindeer1 gp id))

Alternately, perhaps deliverToys and reindeer could appear in section 3.5 so they don't affect the flow in 3.1 and 3.3.

Brian Hulley 22:06, 4 January 2007 (UTC) Page 5: Trivial typo "be execute simultaneously" should be "execute simultaneously" or "be executed simultaneously". I spent ages trying to understand why the 'brutal lock' method wouldn't ensure isolation then I realised it was that I'd just assumed all other threads would be stopped dead in their tracks while the atomic block was running whereas if I'd read the text more carefully I'd have seen that the lock indeed only applies to actions wrapped in
. I think if you said something like "if a thread is accessing an IORef while holding the
lock there is nothing to stop a *different* thread from accessing the same IORef directly at the same time" it would help emphasise the fact that other threads can still be running even though only one thread can be in an atomic block at any given time.

Brian Hulley 09:49, 5 January 2007 (UTC)

  • Page 6: re-execution: this looks like you could end up in an infinite loop since if thread A needs to do something complicated in an atomic block while thread B just sits in a tight loop continuously incrementing the same TVar (used by A) then A would probably always have an invalid view of memory by the time it was ready to validate. Does the implementation address this problem?
  • Page 10: No definition for
    . Section 2.4 only defines
    so perhaps section 3.2 should include:
    check c = if not c then retry() else return ()
(to avoid introducing yet another Haskell function
  • Page 12: Use of
    and list comprehensions seems more complicated than something like:
   mapM_ (elf elf_gp) [1..10]
  • Page 16: Section 5 on locks. It seems to me that STM solves these problems by just conceptually using a single lock (the 'brute force' lock on
    ) but a clever thing about STM is that it gets the simplicity of just having a single global lock without the performance penalty due to the use of logging (ie the global lock is only held for the duration of the validate and commit instead of for the whole duration of the atomic block).

These are just minor points though - I thought it was a really good intro to STM - I've never used/looked at STM before now and yet after reading just this one chapter I feel I've got enough understanding to start using it - Thanks.

EricWilligers 01:02, 7 January 2007 (UTC) I needed to change elf and reindeer due to a compilation error - Expected type: IO (), Inferred type: IO ThreadId. Each of the following worked:-

elf :: Group -> Int -> IO ()
elf gp id = do { _ <- forkIO (forever (elf1 gp id)) ; return () }
elf :: Group -> Int -> IO ThreadId
elf gp id = forkIO (forever (elf1 gp id))
elf gp id = forkIO (forever (elf1 gp id))

The last option avoids the need to mention ThreadId.

Danwang74 04:06, 7 January 2007 (UTC) I think the paper conflates the use of the STM programming model (atomics an various combinators) with the implementation details by focusing only on optimistic methods of implementing the STM programming model. There are fundemental problems with optimistic methods of STM such as situations which require IO as well as implementation challenges which have not been solved.

Pessimistic methods (i.e. lock based methods) can implement the STM programming model efficently. There have been several papers that make progress in this area that I'm sure you are aware of yet they get no mention in the discussion.

Also the comparison of the programming model using bare locks and the STM programming model implemented with optimistics methods is a completely unfair comparison. It is like comparing raw x86 assembly to Haskell compiled to a RISC processor. Programming in raw unstructured x86 assembly is surely a bad idea because the programming model is too low-level. But there is nothing "broken" with x86 assembly when compared to a RISC processor, ignoring aesthetics. In fact one could compile Haskell to x86 assembly with the programmer unaware. The modularity claims are also unfair. Is Haskell compiled to RISC assembly more modular than Haskell compiled to x86?

Locks can be used modularly, just as you can programing in raw x86 assembly modularly with some effort. Unfortunately use of locks in a modular way is not well supported. One way to do this is to use the strawman acquire one global lock implementation of atomics. The whole discussion confuses the programming model and the runtime implementation of it.

In fact the STM name is confusing itself. There's an "atomic" programming model possibly implemented via STM. It would be more insightful to talk more about alternative methods of implementing the atomic programming model, independent of STM. It is ironic and confusing to use STM to refer to the programming model which he authors emphasize could be implemented with locks. There needs to be a clearer seperation between the programming model the implementation details.

Gknauth 19:32, 7 January 2007 (UTC) (I'm definitely a Haskell newb.) page 2, Sec. 2.1, 1st paragraph

 and returns a value of type IO ().
 returns a value of type ().

I read this, and wondered, "Is it type IO (), or type () ?"

Gknauth 19:32, 7 January 2007 (UTC) page 3, line 4, "Then it will performs" s.b. "Then it will perform"

Gknauth 19:32, 7 January 2007 (UTC) page 3, 2nd paragraph. I'm not sure what "IO a" means. From the previous page, I thought that IO implied there would be a side effect. Now you're saying the "return v" action "does no side effects."

Gknauth 19:32, 7 January 2007 (UTC) page 3, paragraph 6, "Gentle..." After 5 dense paragraphs, suddenly this one seems clear. Maybe you intended this. I'm still scratching my head, wondering if IO and IORef are reserved words, or whether they're things you've defined somewhere but not yet revealed.

Gknauth 19:32, 7 January 2007 (UTC) page 4, end of Section 2.1

I'm still left wondering if IO is a marker that signals "side effect(s) can happen," or if it's a type. Put another way, I don't know if IO is an adjective, noun or verb, and I don't know if it's a part of the language or something Joe Haskell wrote for demonstration.

Gknauth 19:32, 7 January 2007 (UTC) page 5, footnote 2. Why do you say "But it's too late now!" Are so many people in the world using Haskell that it's impossible to tweak nomenclature? I thought Haskell was a young up and coming language, not an old language set in its ways. Also, programmers are used to being able to set things right themselves if the language somehow has gotten something wrong. The exclamation implies that once a design mistake is made, however small, programmers just have to lump it.

Gknauth 19:32, 7 January 2007 (UTC) page 5, code example 3. cts -- shorthand for cents? Why not cents?

Gknauth 19:32, 7 January 2007 (UTC) page 6, paragraph 3, "When..." This is very clear. I could substitute that paragraph for ten times as many words I've seen in database texts.

Gknauth 19:32, 7 January 2007 (UTC) page 6, "This process is called re-execution." I read that and thought, "What is to keep a failed transaction from re-executing forever and never being satisfied?" But it looks like you get to that in Section 2.4.

Gknauth 19:32, 7 January 2007 (UTC) page 10, code examples at bottom. Where did check(...) come from? How does it work? Does it block? Does it abort? Is it part of the language? Is it defined later in the chapter? Based on the last sentence of the first full paragraph on page 11, I'm guessing it blocks.

Gknauth 19:32, 7 January 2007 (UTC) page 11, paragraph 3, "Again...", rewrite: "we define Group is declared as a fresh data type..."

Gknauth 19:32, 7 January 2007 (UTC) page 14, last paragraph in section 3.4. That is very cool. The only extra thing I wanted was a picture of how all the pieces you built combined. I was curious which ones you thought of first. I wondered about the order in which you conceived of pieces, as it compares with the order in which you described them.

Gknauth 19:32, 7 January 2007 (UTC) page 17, conclusion, 1st paragraph, last sentence. I can definitely see the need for transactional memory techniques in the years ahead, but I don't know much about how other languages are addressing the issue. Maybe you could provide some pointers. If Haskell is ahead of the others, then so much the better for Haskell, but at the moment I have nothing to compare.

EngineerScotty 17:23, 8 January 2007 (UTC) Page 2. The problem is stated such that no thread may "observe a state in which money has left one account but not arrived in the other". This may be stating the obvious; but should the make-before-break condition also be proscribed--no thread should be able to observe a state in which money has arrived in one account without having left the other, as well? A minor quibble, but one if which not addressed will cause me to an open an account with your bank.  :)

BoAdler 20:10, 10 January 2007 (UTC) page 3. You describe "IORef t" as being "a bit like *t in C", but that notation in C is used to dereference pointers. I know this is pedantic, but it should be "(t *)". I got thrown for a second, and had to remind myself that 't' was a type, not a variable.