Is Parallel Programming Just Too Hard?

pcause writes "There has been a lot of talk recently about the need for programmers to shift paradigms and begin building more parallel applications and systems. The need to do this and the hardware and systems to support it have been around for a while, but we haven't seen a lot of progress. The article says that gaming systems have made progress, but MMOGs are typically years late and I'll bet part of the problem is trying to be more parallel/distributed. Since this discussion has been going on for over three decades with little progress in terms of widespread change, one has to ask: is parallel programming just too difficult for most programmers? Are the tools inadequate or perhaps is it that it is very difficult to think about parallel systems? Maybe it is a fundamental human limit. Will we really see progress in the next 10 years that matches the progress of the silicon?"

Comments

our brains aren't wired to think in parallel

[rritterson]

I can't speak for the rest of the world, or even the programming community. That disclaimer spoken, however, I can say that parallel programming is indeed hard. The trivial examples, like simply running many processes in parallel that are doing the same thing (as in, for example, Monte Carlo sampling) are easy, but the more difficult examples of parallelized mathematical algorithms I've seen, such as those in linear algebra are difficult to conceptualize, let alone program. Trying to manage multiple threads and process communication in an efficient way when actually implementing it adds an additional level of complexity.

I think the biggest reason why it is difficult is that people tend to process information in a linear fashion. I break large projects into a series of chronologically ordered steps and complete one at a time. Sometimes if I am working on multiple projects, I will multitask and do them in parallel, but that is really an example of trivial parallelization.

Ironically, the best parallel programmers may be those good managers, who have to break exceptionally large projects into parallel units for their employees to simultaneously complete. Unfortunately, trying to explain any sort of technical algorithm to my managers usually exacts a look of panic and confusion.

Re:our brains aren't wired to think in parallel

[ubernostrum]

You may want to look into Erlang, which does two things that will interest you:

• Concurrency is handled by lightweight thread-like pseudo-processes passing messages to one another, and supported directly in the language.
• Shared state between these processes is absolutely forbidden.

There are still concurrent problems which are hard, but generally it boils down to the problem being hard instead of the language making the problem harder to express.

Re:our brains aren't wired to think in parallel

[buswolley]

Our brains are massively parallel, but we do not consciously attend to more than a couple of things at a time.

Re:our brains aren't wired to think in parallel

[jd]

Parallel programming is indeed hard. The standard approach these days is to decompose the parallel problem into a definite set of serial problems. The serial problems (applets) can then be coded more-or-less as normal, usually using message-passing rather than direct calls to communicate with other applets. Making sure that everything is scheduled correctly does indeed take managerial skills. The same techniques used to schedule projects (critical path analysis) can be used to minimize time overheads. The same techniques used to minimize the use of physical resources (SIMPLEX aka Operational Research) works just as well on software for physical resources there.

The first problem is that people who make good managers make lousy coders. The second problem is that people who make good coders make lousy managers. The third problem is that plenty of upper management types (unfortunately, I regard the names I could mention as genuinely dangerous and unpredictable) simply have no understanding of programming in general, never mind the intricacies of parallelism.

However, resource management and coding is not enough. These sorts of problems are typically either CPU-bound and not heavy on the networks, or light on the CPU but are network-killers. (Look at any HPC paper on cascading network errors for an example.) Typically, you get hardware which isn't ideally suited to either extreme, so the problem must be transformed into one that is functionally identical but within the limits of what equipment there is. (There is no such thing as a generic parallel app for a generic parallel architecture. There are headaches and there are high-velocity exploding neurons, but that's the range of choices.)

Re:our brains aren't wired to think in parallel

[poopdeville]

That's debatable. You don't look at text one letter at a time to try to decipher the meaning, do you? You probably look at several letters, or even words, at a time. Understanding how the letters or words relate (spatially, syntactically, semantically) is a parallel process. Unless you're a very slow reader, your eyes have probably moved on from the words you're interpreting before you've understood their meaning. This is normal. This is how you establish a context for a particular word or phrase -- by looking at the surrounding words. Another parallel process.

Every process is serial from a broad enough perspective. Eight hypothetical modems can send 8 bits per second. Or are they actually sending a single byte?

Have some friggin' patience

[Corydon76]

Oh noes! Software doesn't get churned out immediately upon the suggestion of parallel programming! Programmers might actually be debugging their own code!

There's nothing new here: just somebody being impatient. Parallel code is getting written. It is not difficult, nor are the tools inadequate. What we have is non-programmers not understanding that it takes a while to write new code.

If anything, that the world hasn't exploded with massive amounts of parallel code is a good thing: it means that proper engineering practice is being used to develop sound programs, and the jonny-come-lately programmers aren't able to fake their way into the marketplace with crappy code, like they did 10 years ago.

Yes, because programmers are too conservative

[Coryoth]

Parallel programming doesn't have to be quite as painful as it currently is. The catch is that you have to face the fact that you can't go on thinking with a sequential paradigm and have some tool, library, or methodology magically make everything work. And now, I'm not talking about functional programming. Functional programming is great, and has a lot going for it, but solving concurrent programming issues is not one of those things. Functional programming deals with concurrency issues by simply avoiding them. For problems that have no state and can be coded purely functionally this is fine, but for a large number of problems you end up either tainting the purity of your functions, or wrapping things up in monads which end up having the same concurrency issues all over again. It does have the benefit that you can isolate the state, and code that doesn't need it is fine, but it doesn't solve the issue of concurrent programming.

No, the different sorts of paradigms I'm talking about no shared state, message passing concurrency models ala CSP and pi Calculus and the Actor Model. That sort of approach in terms of how to think about the problem shows up in languages like Erlang, and Oz which handle concurrency well. The aim here is to make message passing and threads lightweight and integrated right into the language. You think in terms actors passing data, and the language supports you in thinking this way. Personally I'm rather fond of SCOOP for Eiffel which elegantly integrates this idea into OO paradigms (an object making a method call is, ostensibly, passing a message after all). That's still research work though (only available as a preprocessor and library, with promises of eventually integrating it into the compiler). At least it makes thinking about concurrency easier, while still staying somewhat close more traditional paradigms (it's well worth having a look at if you've never heard of it).

The reality, however, is that these new languages which provide the newer and better paradigms for thinking and reasoning about concurrent code, just aren't going to get developer uptake. Programmers are too conservative and too wedded to their C, C++, and Java to step off and think as differently as the solution really requires. No, what I expect we'll get is kluginess retrofitted on to existing languages in a slipshod way that sort of work, in as much as it is an improvement over previous concurrent programming in that language, but doesn't really make the leap required to make the problem truly significantly easier.

Amdahl's law

[apsmith]

I've worked with parallel software for years - there are lots of ways to do it, lots of good programming tools around even a couple of decades back (my stuff ranged from custom message passing in C to using "Connection-Machine Fortran"; now it's java threads) but the fundamental problem was stated long ago by Gene Amdahl - if half the things you need to do are simply not parallelizable, then it doesn't matter how much you parallelize everything else, you'll never go more than twice as fast as using a single thread.

Now there's been lots of work on eliminating those single-threaded bits in our algorithms, but every new software problem needs to be analyzed anew. It's just another example of the no-silver-bullet problem of software engineering...

Re:Not justifyable

[Grave]

I seem to recall comments from Tim Sweeney and John Carmack that parallelism needed to start from the beginning of the code - IE, if you weren't thinking about it and implementing it when you started the engine, it was too late. You can't just tack it on as a feature. Unreal Engine 3 is a prime example of an engine that is properly parallelized. It was designed from the ground up to take full advantage of multiple processing cores.

If your programmers are telling you they need more time to turn a single-threaded game into a multi-threaded one, then the correct solution IS to push the game out the door, because it won't benefit performance to try to do it at the end of a project. It's a fundamental design choice that has to be made early on.

The Bill comes due

[stox]

After years of driving the programming profession to its least common denominator, and eliminating anything that was considered non-essential, somebody is surprised that current professionals are not elastic enough to quickly adapt to a changing environment in hardware. Whoda thunk it? The ones, you may have left, with some skills are nearing retirement.

Re:Nope.

[Gorshkov]

Then why is it that (as of right now) all the up-modded posts are laid out sequentially down the comment tree?

Because one of the things TFM neglects to mention is that parallel programming, like any other programming method, is suitable for some things and not for others .... and the hard reality is, is that most application programmes you see on the desktop are basically serial, simply because of the way PEOPLE process tasks & information

There is a very real limit as to how much you can parallelize standard office tasks.

Re:Nope.

[EvanED]

Yes, but that's an easy sort of parallelism. Heck, I wrote a fractal generator that did the generation in a separate thread in 11th grade after writing my first Win32 program 4 or 5 months previous. It was also my first experience with threads. I'm not even sure I really knew what they were before that. This isn't *really* paralleling the application in the sense TFA means.

Closer is this: After some more work and a rewrite (for other reasons), I had "Fracked" running n threads, each rendering 1/n of the display. Data parallelism == easy parallelism.

But a lot of problems don't fit these models, and need a LOT of thought put into how to parallelize them. It's likely that some problems in P are not efficiently parallelizable.

That's irrelevant.

[Estanislao Mart�nez]

Our cognitive system does many things at the same time, yes. That doesn't answer the question that's being posed here: whether explicit, conscious reasoning about parallel processing is hard for people.

Actually it's just pipelined

[Moraelin]

Actually, it's more like pipelined. The fact that your eyes already moved to the next letter, just says that the old one is still going through the pipeline. Yeah, there'll be some bayesian prediction and pre-fetching involved, but it's nowhere near consciously doing things in parallel.

Try reading two different texts side by side, at the same time, and it won't work that neatly parallel any more.

Heck, there were some recent articles about why most Powerpoint presentations are a disaster: in a nutshell, because your brain isn't that parallel, or doesn't have the bandwidth for it. If you try to read _and_ hear someone saying something (slightly) different at the same time, you just get overloaded and do neither well. The result is those time-wasting meetings where everyone goes fuzzy-brained and forgets everything as soon as the presentation flipped to the next chart.

To get back to the pipeline idea, the brain seems to be quite the pipelined design. Starting from say, the eyes, you just don't have the bandwidth to consciously process the raw stream of pixels. There are several stages of buffering, filtering out the irrelevant bits (e.g., if you focus on the blonde in the car, you won't even notice the pink gorilla jumping up and down in the background), "tokenizing" it, matching and cross-referencing it, etc, and your conscious levels work on the pre-processed executive summary.

We already know, for example, that the shortest term buffer can store about 8 seconds worth of raw data in transit. And that after about 8 seconds it will discard that data, whether it's been used or not. (Try closing your eyes while moving around a room, and for about 8 seconds you're still good. After that, you no longer know where you are and what the room looks like.)

There's a lot of stuff done in parallel at each stage, yes, but the overall process is really just a serial pipeline.

At any rate, yeah, your eyes may already be up to 8 seconds ahead of what your brain currently processes. It doesn't mean you're that much of a lean, mean, parallel-processing machine, it just means that some data is buffered in transit.

Even time-slicing won't really work that well, because of that (potential) latency and the finite buffers. If you want to suddenly focus on another bit of the picture, or switch context to think of something else, you'll basically lose some data in the process. Your pipeline still has the old data in it, and it's going right to the bit bucket. That or both streams get thrashed because there's simply not enough processing power and bandwidth for both to go through the pipeline at the same time.

Again, you only need to look at the fuzzy-brain effect of bad Powerpoint presentations to see just that in practice. Forced to try to process two streams at the same time (speech and text), people just make a hash of both.

Re:Nope.

[smellotron]

But a lot of problems don't fit these models, and need a LOT of thought put into how to parallelize them. It's likely that some problems in P are not efficiently parallelizable.

I would venture a guess that most problems would benefit from parallelizing basic data structure tasks:

• anything (comparable) can be sorted using divide-and-conquer mergesort
• scanning through an array-based collection (*not* a linked list) can be divided among processors—this is frequently done in hardware, e.g. for CPU cache hashtable lookups

Further, there's a few other obvious ways to parallelize:

• Split program into a chain of filters between producers and consumers, and give each filter its own thread/process. For example, create an event receiver thread, a "do-stuff" thread, and a display thread. At the very least, this will reduce UI response latency.
• Split program processing into "1 event dispatcher + N worker threads", like Apache or Squid. This by itself would be a good way to reduce blocking in most applications. Why should the interface be locked up when expensive processing is happening in a program? Maybe while Photoshop/GIMP runs some filter on my image, I'd like to browse the help documents or scroll around the viewport.
• Re-evaluate any processing as a dependency tree, and code it in something like Twisted, where every piece of code executes nonblocking snippets, and a reactor thread dispatches between them (this is basically like light-light-weight threads)

The reason the problems don't fit these models is moreso that we're used to thinking about algorithms as an ordered list of steps, rather than a set of workers on an assembly line (operating as fast as the slowest individual worker).