Thanks for writing your thoughts about threads. I use threads, and I like them a bit more than most of the other ways to introduce parallelism and concurrency. I only have one quibble with your points:

The "heaviness" of P5 threading is a misnomer. The threads aren't heavy; the implementation of shared memory is heavy. And that could easily be fixed.

As the current implementation of threads tries to simulate parallelism within one interpreter, "fixing" that is not easy as long as you want the promise of implicitly shared things (like the namespace and tied variables most importantly) to remain there. If you move to a less implicit model of sharing, like your usual approach using a queue, and also remove the promise of changes being visible to every thread, or convert it to a threat/warning that changes become visible everywhere, without protection, then I can concur that threads could become less heavy. But as they are now, they are heavy and making the Perl environment look to several threads as if they were the only thread currently running makes them remain heavy.

As the current implementation of threads tries to simulate parallelism within one interpreter,

After 24 hrs of thinking about that statement, I cannot make it gel with my knowledge that each thread starts a new interpreter.

Given it is you, I assume I'm just misunderstanding your drift?

---

With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.

No, it was me who is simply wrong. Actually testing reveals that each thread really has its own namespace, like you say. So I'm less sure of where the actual problem or overhead lies with sharing data across threads, at least where the namespace is involved.

Oh wait, everything on Windows OS is a thread, even the forks. If I had to write a serious program, that could involve crashes or whatever, I would use fork. On Linux, SysV Shared Memory along with a fork, is the best way to go. It just seems to have a cumbersome interface now, it would be nice to have a super-easy SysV shared memory module. I know modules exist, but they are not as easy to use as threads::shared.

Use the Fork, Luke, the Fork. :-)

Try using a fork -- the metal utensil -- to slice a tomato and see the sticky mess you end up with.

Like its namesake, fork is useful for somethings and not so for others.

everything on Windows OS is a thread, even the forks.

Everything on modern *nix is also a thread. Processes are simply single threaded processes.

---

With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.

What are you advocating then, in regards to all this? Are you lobbying for forks and threads that don't use clone(), so you get a pure executable space, unencumbered by the parent process's ENV and filehandles? I think that would be the most desirable way to do it, but you would be forced to deal with the filehandles yourself, having to open them all manually. Most people don't expect to have to manually open stdout or stdin.

---

I'm not really a human, but I play one on earth.
Old Perl Programmer Haiku ................... flash japh

forks?

If the problem is IO bound to a single, local, harddisk, and is uncacheable, then concurrency may not help.

Yes. I appreciated your detailed response.

(Fwiw, in my original comment I meant that a problem that is as you describe is really IO bound, and a problem that isn't, isn't.)

Only mathematicians and computer scientists demand total determinacy; and throw their arms up in refusal to work if they don't get it.

I think you've misinterpreted The Problem with Threads. The author was not demanding total determinacy (which I agree would be profoundly bogus). To quote from the abstract: "Nondeterminism should be explicitly and judiciously introduced where needed, rather than removed where not needed."

don't throw the baby out with the bath water. Flames are dangerous; but oh so very useful.

While your particular choice of metaphors has me imagining a complicated maneuver involving strong folk, an old fashioned bath, a gassy baby, a lighter, and some sort of worthwhile singeing operation, I don't think your point is that far adrift from (Perl) "should not hide the existence of OS-level threads, or fail to provide access to lower level concurrency control constructs".

Futures neither remove the complexity nor solve the problems; they just bury them under the covers forcing everyone to rely upon the efficacy of their implementation and the competence of the implementors.

I confess to being amazed by this statement. Are you really saying that, when solving any concurrent problem correctly (data-wise), Futures are never (or seldom) simpler for the programmers writing, reading, and modifying the code than directly using the underlying low level concurrency constructs (which I'll hand wavily define as threads, locks, cas, etc.)?

Do you think there are any very useful high level concurrency constructs?

The "heaviness" of P5 threading is a misnomer. The threads aren't heavy; the implementation of shared memory is heavy. And that could easily be fixed. ... They've basically stagnated for the past 8 or more years because p5p won't allow change.

I'd appreciate a link to a decent recent technical discussion of the problem of the heaviness of shared memory in P5 and reasonable potential fixes. Failing that, a couple paragraphs that summarize the problem and the fix you're suggesting would be great. Or maybe you could write a meditation that invites monks to focus on the technical issues related to the problem and potential fixes?

• (Fwiw, in my original comment I meant that a problem that is as you describe is really IO bound

  (I had some trouble parsing that sentence; I hope i got you right.)

  The problems I describe where concurrency does help IO bound problems, are "really IO bound".

  1. With multiple physical drives:

     A simple filter -- read modify write (eg. grep) -- can be reading the next record from the source whilst the previous write to the destination completes. Runtime can be more than halved.

  2. Remote drives:

     By overlapping the latencies in the reads and writes; throughput can be increased substantially.

  3. SSDs:

     The write time for SSDs is usually substantially greater than the read time. For tasks that write less records than are written -- eg. line filters etc. -- overlapping the two can increase throughput.

  4. Temporal locality:

     Ditto.

  The point is that your statement implies that all IO bound processes cannot benefit from threading; where the truth is that only a few types cannot realise any benefit.

• I think you've misinterpreted The Problem with Threads. "Nondeterminism should be explicitly and judiciously introduced where needed, rather than removed where not needed."'

  Actually, no I haven't. And your quote proves it.

  The single biggest source of both bugs & design fails, and unrealised performance, is synchronisation and locking. Ie. Attempts to impose determinism.

  People insist on trying to enforce the determinism that when this thread has finished with this piece of data; that thread starts processing it. They'll use semaphores, or locks, or signals, or mutexes, and often some combination of them. And with them come all the nasties wrongly attributed to "threading"; dead-locks, live-locks, and priority inversions; often transient, and always a nightmare to diagnose and fix.

  And even when they get it right, all they've done is create a lock-stepped, and thus sequential process. Almost, if not exactly, as if no threading had been done.

  Except of course, they've carefully orchestrated that there be at least one, and often multiple, context swaps occurring between steps. And then they wonder why it's slowed everything down. By attempting to deterministically control the flow of data between threads, you're working against the system scheduler and guaranteeing that no thread ever uses its full timeslice. Indeed, many times you're actually forcing the scheduler to load a thread only to immediately remove it again because the other thread that you're trying to pass control to hasn't yet run.

  The archetypal example of this is the Global Interpreter Lock as seen in Python, Ruby et al.

  Think of this like a hub airport would be if there was no buffer space -- aprons and gates. Planes would have to circle until one of the connecting flights arrived, then both land, exchange passengers; then both take off and circle again until another connecting flight was available. A stupid, but surprisingly accurate analogy.

  And people forget -- and that includes academics who tend to test and benchmark their algorithms in isolation of other workloads -- that the scheduler isn't just scheduling the threads of their process, but also the threads of every other process in the system. If you fail to utilise the full timeslice, the remainder doesn't automatically get transferred to another thread of your process, but most times gets lost entirely because (the thread of) another process, and often many other processes, get scheduled before your process gets another chance. And even when it (your process) does get another slot, it may be the wrong thread, or even the same thread that just abandoned its last one.

  To make best use of threading, you need to embrace non-determinism, by eschewing locking and synchronisation and the conditional checks associated with them -- unless it is absolutely required, which with appropriately consider algorithms, is very rare -- and allow things to free run.

  By providing low-overhead, flexible sized buffers (queues) between sequential sections of processing, you allow the well-tried and proven system scheduler with its multiple classes, dynamic priorities, dynamic timeslices et al. to give timeslices to threads that need them, in a timely manner.

  Imagine the complexity of locking required by a shared hash with multiple reading and writing threads performing concurrent access. And imagine trying to test it; and then debug it.

  Then, if you are really interested, go off and watch Cliff Click describe his lock-free hash table. Watch it all. See how he scales a hash table to be used concurrently from thousands of threads. See also the discussion of the bottlenecks with the standard Java ConcurrentHashMap.

  Bottom line: Everything in that paper (The Problem with Threads) tackles the "problem" from the completely wrong direction. Had it not been dated 2006, I would have assumed that it had been written circa. 1995 or earlier, such is the level of dated thinking it employs.

• I don't think your point is that far adrift from (Perl) "should not hide the existence of OS-level threads, or fail to provide access to lower level concurrency control constructs".

  The sentence you quote -- "But equally, don't throw the baby out with the bath water. Flames are dangerous; but oh so very useful." -- is (in section 2 of my reply) aimed directly at the paper you cite.

  However, your quote from the P6 spec. is also questionable if the resulting low-level constructs end up being anything like Perl5's.

  Simply exposing the low-level system constructs, especially if they are forced into being POSIX compatible emulations over the underlying OS is arguably worse. You'd just be recreating all the bad bits of the Perl5 implementation.

  And the addition of Futures as the only alternative will do little to make things better. See below.

• Are you really saying that, when solving any concurrent problem correctly (data-wise), Futures are never (or seldom) simpler for the programmers writing, reading, and modifying the code than directly using the underlying low level concurrency constructs ?

  No. I like Futures a lot. They are the best (cleanest; easiest to grasp), high level threading construct I've yet seen. For a certain class of problems, a well implemented -- and that means NOT the obvious, naive implementation -- they can/could be a very simple and effective solution.

  But, I'm am saying that they move the problems somewhere else.

  In effect, all (naive) Futures do, is wrap-over (hide) the mechanics of ThreadCreate() and ThreadJoin(), And all their problems still persist.

  When a Future comes to be realised -- the thread comes to be joined -- if that thread is not yet finished, it will block. And as soon as you have blocking, you have the potential for deadlocking, live-locking and priority inversion. Except now the locking is hidden from, and inaccessible to the programmer.

  And what's worse, the very simplicity of Futures is -- as happened with the synchronized keyword in Java -- going to encourage people to throw threads at everything without understanding or thinking through the consequences. The very act of providing a simple, high-level construct that hides much of the mechanics of threading and locking, will encourage them to be used widely and inappropriately.

  It is possible to implement the hidden lock such that it will detect deadlocks and produce diagnostics, but the runtime penalties of every implementation I have seen are such that for any performance critical application -- why else are you using threading? -- the penalties add up to the point where programmers will need to revert to the low-level constructs in order to regain the level of control they need to realise the performance they are seeking.

  And finally, there are a whole raft of classes of algorithm for which (naive) Futures are not just inappropriate, but completely unusable. For some of these classes, a sophisticated implementation can mitigate that and render them usable; but that is still the subject of research.

  The bottom line is that to implement Futures such that they are reliable, easy to use and sophisticated enough to allow their use for a wide range of problems, is extremely hard. And testing them even harder. If P6 shipped -- should that mythical event ever occur -- with a naive implementation, or a poorly tested sophisticated one, the damage to its reputation would be huge.

  Now think how long (and how many re-writes) it has taken for Java threading to approach some level of real usability.

I'd appreciate a link to a decent recent technical discussion of the problem of the heaviness of shared memory in P5 and reasonable potential fixes. Failing that, a couple paragraphs that summarize the problem and the fix you're suggesting would be great. Or maybe you could write a meditation that invites monks to focus on the technical issues related to the problem and potential fixes?

I started to put something together in reply to Corion's post above, but RL took priority. Give me a week or two and I'll post my demonstration of the problems, highlight (some of) the source(s) of those problems; and offer some suggestions as to how they could be solved (for perl5).

---

With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.

And the academic world will rejoice, proclaim her a genius of our time, and no doubt award her a Nobel prize. (That'd be nice!)

Wish it, and it shall be so :)

FWIW, mathy talk here http://www.mathunion.org/fileadmin/IMU/Prizes/2014/news_release_mirzakhani.pdf

In neither link was I able to find that a "mathematician [invented] a symbol for an asynchronous queue". Can anybody help point to that part?

Good thoughts.   Very.

The most common mis-use of threads that I see is the “flaming arrows strategy.”   For each request, light another flaming arrow (a thread), and shoot it into the air to land where and how it may and then die.   This is sure to overwhelm almost any scheduler, which also has no real way to distinguish one scheduler-unit-of-work from another.   It also burdens the system with setup and teardown of processes, which is often expensive.

A far better (and, far more scalable) approach is to do what’s done in, say, any restaurant:   to have a pool of workers who shift their attentions among a greater number of active orders, which are flowing station-by-station through a flexible and well-defined lifespan.   The workers can be generalists, or specialists, or sometimes both, and the allocation can be adjusted at any time.   The system that is built using threads/processes, is very aware of exactly what business-problem it is constructed to solve.   When a unit of work is obliged to wait, a worker is not.   Instead, the order is briefly “parked.”   Workers live to a ripe old age.   If a single work-unit needs to go to several stations at once (burger, fries, milkshake), it might be opportunistically serviced by three workers simultaneously.   Commitments regarding service level are engineered into the system, so that you can say (and measure), that “95% of the time, all orders will be prepared and served to the customer within 3 minutes.”   Concurrency is used to address the problem, but there is not a one-to-one correspondence between workers and work.   There is a dedicated management role, separate from any worker, which is constantly “riding the faders” to keep everything in balance.

There are plenty of good workload-management systems, including some that are designed to share work in a computing cluster, as well as those that are designed to be self-adapting to changing workload mixes and resource-constraint pressures.   The goal is to find the ever-changing “sweet spot” in which a maximum number of units-of-work are being processed, in the least amount of time, without creating traffic-jams at any point (including the OS scheduler itself).   The principles used are simply taken from the real world of human and industrial processes.

Almost every sentence in that diatribe is wrong.

I can't be bothered to explain why any more cos you'll only regurgitate it back to me in the wrong context a week or two from now when you've forgotten where you read it and what it meant.

---

With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.