Related
This is the first time I am trying to profile a multi-threaded program.
I suspect the problem is it waiting for something, but I have no clue what, the program never reaches 100% of CPU, GPU, RAM or I/O use.
Until recently, I've only worked on projects with single-threading, or where the threads were very simple (example: usually an extra thread just to ensure the UI is not locked while the program works, or once I made a game engine with a separate thread to handle .XM and .IT files music, so that the main thread could do everything, while the other thread in another core could take care of decoding those files).
This program has several threads, and they don't do parallel work on the same tasks, each thread has its own completely separate purpose (for example one thread is dedicated to handling all sound-related API calls to the OS).
I downloaded Microsoft performance tools, there is a blog by an ex-Valve employee that explains that they work to do this, but although I even managed to make some profiles and whatnot, I don't really understood what I am seeing, it is only a bunch of pretty graphs to me (except the CPU use graph, that I already knew from doing sample-based profiling on single-threaded apps), so, how I find why the program is waiting on something? Or how I find what is it waiting for? How I find what thread is blocking the others?
I look at is as an alternation between two things:
a) measuring overall time, for which all you need is some kind of timer, and
b) finding speedups, which does not mean measuring, in spite of what a lot of people have been told.
Each time you find a speedup, you time the results and do it again.
That's the alternation.
To find speedups, the method I and many people use is random pausing.
The idea is, you get the program running under a debugger and manually interrupt it, several times.
Each time, you examine the state of every thread, including the call stack.
It is very crude, and it is very effective.
The reason this works is that the only way the program can go faster is if it is doing an activity that you can remove, and if that saves a certain fraction of time, you are at least that likely to see it on every pause.
This works whether it is doing I/O, waiting for something, or computing.
It sees things that profilers do not expose, because they make summaries from which speedups can easily hide.
Performance Wizard in Visual Studio Performance and Diagnostics Hub has "Resource contention data" profiling regime which allows to analyze concurrency contention among threads, i.e. how the overall performance of a program is impacted by threads waiting on other threads. Please refer to this blog post for more details.
PerfView is an extremely powerful profiling tool which allows one to analyze the impact of service threads and tasks to the overall performance of the program. Here is the PerfView Tutorial available.
Normally it is said that multi threaded programs are non-deterministic, meaning that if it crashes it will be next to impossible to recreate the error that caused the condition. One doesn't ever really know what thread is going to run next, and when it will be preempted again.
Of course this has to do with the OS thread scheduling algorithm and the fact that one doesn't know what thread is going to be run next, and how long it will effectively run.
Program execution order also plays a role as well, etc...
But what if you had the algorithm used for thread scheduling and what if you could know when what thread is running, could a multi threaded program then become "deterministic", as in, you'll be able to reproduce a crash?
Knowing the algorithm will not actually allow you to predict what will happen when. All kinds of delays that happen in the execution of a program or thread are dependent on environmental conditions such as: available memory, swapping, incoming interrupts, other busy tasks, etc.
If you were to map your multi-threaded program to a sequential execution, and your threads in themselves behave deterministically, then your whole program could be deterministic and 'concurrency' issues could be made reproducible. Of course, at that point they would not be concurrency issues any more.
If you would like to learn more, http://en.wikipedia.org/wiki/Process_calculus is very interesting reading.
My opinion is: technically no (but mathematically yes). You can write deterministic threading algorithm, but it will be extremely hard to predict state of the application after some sensible amount of time that you can treat it is non-deterministic.
There are some tools (in development) that will try to create race-conditions in a somewhat predictable manner but this is about forward-looking testing, not about reconstructing a 'bug in the wild'.
CHESS is an example.
It would be possible to run a program on a virtual multi-threaded machine where the allocation of virtual cycles to each thread was done via some entirely deterministic process, possibly using a pseudo-random generator (which could be seeded with a constant before each program run). Another, possibly more interesting, possibility would be to have a virtual machine which would alternate between running threads in 'splatter' mode (where almost any variable they touch would have its value become 'unknown' to other threads) and 'cleanup' mode (where results of operations with known operands would be visible and known to other threads). I would expect the situation would probably be somewhat analogous to hardware simulation: if the output of every gate is regarded as "unknown" between its minimum and maximum propagation times, but the simulation works anyway, that's a good indication the design is robust, but there are many useful designs which could not be constructed to work in such simulations (the states would be essentially guaranteed to evolve into a valid combination, though one could not guarantee which one). Still, it might be an interesting avenue of exploration, since large parts of many programs could be written to work correctly even in a 'splatter mode' VM.
I don't think it is practicable. To enforce a specific thread interleaving we require to place locks on shared variables, forcing the threads to access them in a specific order. This would cause severe performance degradation.
Replaying concurrency bugs is usually handled by record&replay systems. Since the recording of such large amounts of information also degrades performance, the most recent systems do partial logging and later complete the thread interleavings using SMT solving. I believe that the most recent advance in this type of systems is Symbiosis (published in this year's PLDI conference). Tou can find open source implementations in this URL:
http://www.gsd.inesc-id.pt/~nmachado/software/Symbiosis_Tutorial.html
This is actually a valid requirement in many systems today which want to execute tasks parallelly but also want some determinism from time to time.
For example, a mobile company would want to process subscription events of multiple users parallelly but would want to execute events of a single user one at a time.
One solution is to of course write everything to get executed on a single thread. Another solution is deterministic threading. I have written a simple library in Java that can be used to achieve the behavior I have described in the above example. Take a look at this- https://github.com/mukulbansal93/deterministic-threading.
Now, having said that, the actual allocation of CPU to a thread or process is in the hands of the OS. So, it is possible that the threads get the CPU cycles in a different order every time you run the same program. So, you cannot achieve the determinism in the order the threads are allocated CPU cycles. However, by delegating tasks effectively amongst threads such that sequential tasks are assigned to a single thread, you can achieve determinism in overall task execution.
Also, to answer your question about the simulation of a crash. All modern CPU scheduling algorithms are free from starvation. So, each and every thread is bound to get guaranteed CPU cycles. Now, it is possible that your crash was a result of the execution of a certain sequence of threads on a single CPU. There is no way to rerun that same execution order or rather the same CPU cycle allocation order. However, the combination of modern CPU scheduling algorithms being starvation-free and Murphy's law will help you simulate the error if you run your code enough times.
PS, the definition of enough times is quite vague and depends on a lot of factors like execution cycles need by the entire program, number of threads, etc. Mathematically speaking, a crude way to calculate the probability of simulating the same error caused by the same execution sequence is on a single processor is-
1/Number of ways to execute all atomic operations of all defined threads
For instance, a program with 2 threads with 2 atomic instructions each can be allocated CPU cycles in 4 different ways on a single processor. So probability would be 1/4.
Lots of crashes in multithreaded programs have nothing to do with the multithreading itself (or the associated resource contention).
Normally it is said that multi threaded programs are non-deterministic, meaning that if it crashes it will be next to impossible to recreate the error that caused the condition.
I disagree with this entirely, sure multi-threaded programs are non-deterministic, but then so are single-threaded ones, considering user input, message pumps, mouse/keyboard handling, and many other factors. A multi-threaded program usually makes it more difficult to reproduce the error, but definitely not impossible. For whatever reasons, program execution is not completely random, there is some sort of repeatability (but not predictability), I can usually reproduce multi-threaded bugs rather quickly in my apps, but then I have lots of verbose logging in my apps, for the end users' actions.
As an aside, if you are getting crashes, can't you also get crash logs, with call stack info? That will greatly aid in the debugging process.
I was reading the SQLite FAQ, and came upon this passage:
Threads are evil. Avoid them.
I don't quite understand the statement "Thread are evil". If that is true, then what is the alternative?
My superficial understanding of threads is:
Threads make concurrence happen. Otherwise, the CPU horsepower will be wasted, waiting for (e.g.) slow I/O.
But the bad thing is that you must synchronize your logic to avoid contention and you have to protect shared resources.
Note: As I am not familiar with threads on Windows, I hope the discussion will be limited to Linux/Unix threads.
When people say that "threads are evil", the usually do so in the context of saying "processes are good". Threads implicitly share all application state and handles (and thread locals are opt-in). This means that there are plenty of opportunities to forget to synchronize (or not even understand that you need to synchronize!) while accessing that shared data.
Processes have separate memory space, and any communication between them is explicit. Furthermore, primitives used for interprocess communication are often such that you don't need to synchronize at all (e.g. pipes). And you can still share state directly if you need to, using shared memory, but that is also explicit in every given instance. So there are fewer opportunities to make mistakes, and the intent of the code is more explicit.
Simple answer the way I understand it...
Most threading models use "shared state concurrency," which means that two execution processes can share the same memory at the same time. If one thread doesn't know what the other is doing, it can modify the data in a way that the other thread doesn't expect. This causes bugs.
Threads are "evil" because you need to wrap your mind around n threads all working on the same memory at the same time, and all of the fun things that go with it (deadlocks, racing conditions, etc).
You might read up about the Clojure (immutable data structures) and Erlang (message passsing) concurrency models for alternative ideas on how to achieve similar ends.
What makes threads "evil" is that once you introduce more than one stream of execution into your program, you can no longer count on your program to behave in a deterministic manner.
That is to say: Given the same set of inputs, a single-threaded program will (in most cases) always do the same thing.
A multi-threaded program, given the same set of inputs, may well do something different every time it is run, unless it is very carefully controlled. That is because the order in which the different threads run different bits of code is determined by the OS's thread scheduler combined with a system timer, and this introduces a good deal of "randomness" into what the program does when it runs.
The upshot is: debugging a multi-threaded program can be much harder than debugging a single-threaded program, because if you don't know what you are doing it can be very easy to end up with a race condition or deadlock bug that only appears (seemingly) at random once or twice a month. The program will look fine to your QA department (since they don't have a month to run it) but once it's out in the field, you'll be hearing from customers that the program crashed, and nobody can reproduce the crash.... bleah.
To sum up, threads aren't really "evil", but they are strong juju and should not be used unless (a) you really need them and (b) you know what you are getting yourself into. If you do use them, use them as sparingly as possible, and try to make their behavior as stupid-simple as you possibly can. Especially with multithreading, if anything can go wrong, it (sooner or later) will.
I would interpret it another way. It's not that threads are evil, it's that side-effects are evil in a multithreaded context (which is a lot less catchy to say).
A side effect in this context is something that affects state shared by more than one thread, be it global or just shared. I recently wrote a review of Spring Batch and one of the code snippets used is:
private static Map<Long, JobExecution> executionsById = TransactionAwareProxyFactory.createTransactionalMap();
private static long currentId = 0;
public void saveJobExecution(JobExecution jobExecution) {
Assert.isTrue(jobExecution.getId() == null);
Long newId = currentId++;
jobExecution.setId(newId);
jobExecution.incrementVersion();
executionsById.put(newId, copy(jobExecution));
}
Now there are at least three serious threading issues in less than 10 lines of code here. An example of a side effect in this context would be updating the currentId static variable.
Functional programming (Haskell, Scheme, Ocaml, Lisp, others) tend to espouse "pure" functions. A pure function is one with no side effects. Many imperative languages (eg Java, C#) also encourage the use of immutable objects (an immutable object is one whose state cannot change once created).
The reason for (or at least the effect of) both of these things is largely the same: they make multithreaded code much easier. A pure function by definition is threadsafe. An immutable object by definition is threadsafe.
The advantage processes have is that there is less shared state (generally). In traditional UNIX C programming, doing a fork() to create a new process would result in shared process state and this was used as a means of IPC (inter-process communication) but generally that state is replaced (with exec()) with something else.
But threads are much cheaper to create and destroy and they take less system resources (in fact, the operating itself may have no concept of threads yet you can still create multithreaded programs). These are called green threads.
The paper you linked to seems to explain itself very well. Did you read it?
Keep in mind that a thread can refer to the programming-language construct (as in most procedural or OOP languages, you create a thread manually, and tell it to executed a function), or they can refer to the hardware construct (Each CPU core executes one thread at a time).
The hardware-level thread is obviously unavoidable, it's just how the CPU works. But the CPU doesn't care how the concurrency is expressed in your source code. It doesn't have to be by a "beginthread" function call, for example. The OS and the CPU just have to be told which instruction threads should be executed.
His point is that if we used better languages than C or Java with a programming model designed for concurrency, we could get concurrency basically for free. If we'd used a message-passing language, or a functional one with no side-effects, the compiler would be able to parallelize our code for us. And it would work.
Threads aren't any more "evil" than hammers or screwdrivers or any other tools; they just require skill to utilize. The solution isn't to avoid them; it's to educate yourself and up your skill set.
Creating a lot of threads without constraint is indeed evil.. using a pooling mechanisme (threadpool) will mitigate this problem.
Another way threads are 'evil' is that most framework code is not designed to deal with multiple threads, so you have to manage your own locking mechanisme for those datastructures.
Threads are good, but you have to think about how and when you use them and remember to measure if there really is a performance benefit.
A thread is a bit like a light weight process. Think of it as an independent path of execution within an application. The thread runs in the same memory space as the application and therefore has access to all the same resources, global objects and global variables.
The good thing about them: you can parallelise a program to improve performance. Some examples, 1) In an image editing program a thread may run the filter processing independently of the GUI. 2) Some algorithms lend themselves to multiple threads.
Whats bad about them? if a program is poorly designed they can lead to deadlock issues where both threads are waiting on each other to access the same resource. And secondly, program design can me more complex because of this. Also, some class libraries don't support threading. e.g. the c library function "strtok" is not "thread safe". In other words, if two threads were to use it at the same time they would clobber each others results. Fortunately, there are often thread safe alternatives... e.g. boost library.
Threads are not evil, they can be very useful indeed.
Under Linux/Unix, threading hasn't been well supported in the past although I believe Linux now has Posix thread support and other unices support threading now via libraries or natively. i.e. pthreads.
The most common alternative to threading under Linux/Unix platforms is fork. Fork is simply a copy of a program including it's open file handles and global variables. fork() returns 0 to the child process and the process id to the parent. It's an older way of doing things under Linux/Unix but still well used. Threads use less memory than fork and are quicker to start up. Also, inter process communications is more work than simple threads.
In a simple sense you can think of a thread as another instruction pointer in the current process. In other words it points the IP of another processor to some code in the same executable. So instead of having one instruction pointer moving through the code there are two or more IP's executing instructions from the same executable and address space simultaneously.
Remember the executable has it's own address space with data / stack etc... So now that two or more instructions are being executed simultaneously you can imagine what happens when more than one of the instructions wants to read/write to the same memory address at the same time.
The catch is that threads are operating within the process address space and are not afforded protection mechanisms from the processor that full blown processes are. (Forking a process on UNIX is standard practice and simply creates another process.)
Out of control threads can consume CPU cycles, chew up RAM, cause execeptions etc.. etc.. and the only way to stop them is to tell the OS process scheduler to forcibly terminate the thread by nullifying it's instruction pointer (i.e. stop executing). If you forcibly tell a CPU to stop executing a sequence of instructions what happens to the resources that have been allocated or are being operated on by those instructions? Are they left in a stable state? Are they properly freed? etc...
So, yes, threads require more thought and responsibility than executing a process because of the shared resources.
For any application that requires stable and secure execution for long periods of time without failure or maintenance, threads are always a tempting mistake. They invariably turn out to be more trouble than they are worth. They produce rapid results and prototypes that seem to be performing correctly but after a couple weeks or months running you discover that they have critical flaws.
As mentioned by another poster, once you use even a single thread in your program you have now opened a non-deterministic path of code execution that can produce an almost infinite number of conflicts in timing, memory sharing and race conditions. Most expressions of confidence in solving these problems are expressed by people who have learned the principles of multithreaded programming but have yet to experience the difficulties in solving them.
Threads are evil. Good programmers avoid them wherever humanly possible. The alternative of forking was offered here and it is often a good strategy for many applications. The notion of breaking your code down into separate execution processes which run with some form of loose coupling often turns out to be an excellent strategy on platforms that support it. Threads running together in a single program is not a solution. It is usually the creation of a fatal architectural flaw in your design that can only be truly remedied by rewriting the entire program.
The recent drift towards event oriented concurrency is an excellent development innovation. These kinds of programs usually prove to have great endurance after they are deployed.
I've never met a young engineer who didn't think threads were great. I've never met an older engineer who didn't shun them like the plague.
Being an older engineer, I heartily agree with the answer by Texas Arcane.
Threads are very evil because they cause bugs that are extremely difficult to solve. I have literally spent months solving sporadic race-conditions. One example caused trams to suddenly stop about once a month in the middle of the road and block traffic until towed away. Luckily I didn't create the bug, but I did get to spend 4 months full-time to solve it...
It's a tad late to add to this thread, but I would like to mention a very interesting alternative to threads: asynchronous programming with co-routines and event loops. This is being supported by more and more languages, and does not have the problem of race conditions like multi-threading has.
It can replace multi-threading in cases where it is used to wait on events from multiple sources, but not where calculations need to be performed in parallel on multiple CPU cores.
I have a multithreaded server C++ program that uses MSXML6 and continuously parses XML messages, then applies a prepared XSLT transform to produce text. I am running this on a server with 4 CPUs. Each thread is completely independent and uses its own transform object. There is no sharing of any COM objects among the threads.
This works well, but the problem is scalability. When running:
with one thread, I get about 26 parse+transformations per second per thread.
with 2 threads, I get about 20/s/thread,
with 3 threads, 18/s/thread.
with 4 threads, 15/s/thread.
With nothing shared between threads I expected near-linear scalability so it should be 4 times faster with 4 threads than with 1. Instead, it is only 2.3 times faster.
It looks like a classic contention problem. I've written test programs to eliminate the possibility of the contention being in my code. I am using the DOMDocument60 class instead of the FreeThreadedDOMDocument one in order to avoid unnecessary locking since the documents are never shared between threads. I looked hard for any evidence of cache-line false sharing and there isn't any, at least in my code.
Another clue, the context switch rate is > 15k/s for each thread.
I am guessing the culprit is the COM memory manager or the memory manager within MSXML. Maybe it has a global lock that has to be acquired and released for every memory alloc/deallocation. I just can't believe that in this day and age, the memory manager is not written in a way that scales nicely in multithreaded multi-cpu scenarios.
Does anyone have any idea what is causing this contention or how to eliminate it?
It is fairly common for heap-based memory managers (your basic malloc/free) to use a single mutex, there are fairly good reasons for it: a heap memory area is a single coherent data structure.
There are alternate memory management strategies (e.g. hierachical allocators) that do not have this limitation. You should investigate customizing the allocator used by MSXML.
Alternatively, you should investigate moving away from a multi-threaded architecture to a multi-process architecture, with separate processes for each MSXML worker. Since your MSXML worker take string data as input and output, you do not have a serialization problem.
In summary: use a multiprocess architecture, it's a better fit to your problem, and it will scale better.
MSXML uses BSTRs, which use a global lock in its heap management. It caused us a ton of trouble for a massively multiuser app a few years ago.
We removed our use of XML in our app, you may not be able to do this, so you might be better off using an alternative XML parser.
Thanks for the answers. I ended up implementing a mix of the two suggestions.
I made a COM+ ServicedComponent in C#, hosted it as a separate server process under COM+, and used the XSLCompiledTransform to run the transformation. The C++ server connects to this external process using COM and sends it the XML and gets back the transformed string. This doubled the performance.
How do you make your application multithreaded ?
Do you use asynch functions ?
or do you spawn a new thread ?
I think that asynch functions are already spawning a thread so if your job is doing just some file reading, being lazy and just spawning your job on a thread would just "waste" ressources...
So is there some kind of design when using thread or asynch functions ?
If you are talking about .Net, then don't forget the ThreadPool. The thread pool is also what asynch functions often use. Spawning to much threads can actually hurt your performance. A thread pool is designed to spawn just enough threads to do the work the fastest. So do use a thread pool instead of spwaning your own threads, unless the thread pool doesn't meet your needs.
PS: And keep an eye out on the Parallel Extensions from Microsoft
Spawning threads is only going to waste resources if you start spawning tons of them, one or two extra threads isn't going to effect the platforms proformance, infact System currently has over 70 threads for me, and msn is using 32 (I really have no idea how a messenger can use that many threads, exspecialy when its minimised and not really doing anything...)
Useualy a good time to spawn a thread is when something will take a long time, but you need to keep doing something else.
eg say a calculation will take 30 seconds. The best thing to do is spawn a new thread for the calculation, so that you can continue to update the screen, and handle any user input because users will hate it if your app freezes untill its finished doing the calculation.
On the other hand, creating threads to do something that can be done almost instantly is nearly pointless, since the overhead of creating (or even just passing work to an existing thread using a thread pool) will be higher than just doing the job in the first place.
Sometimes you can break your app into a couple of seprate parts which run in their own threads. For example in games the updates/physics etc may be one thread, while grahpics are another, sound/music is a third, and networking is another. The problem here is you really have to think about how these parts will interact or else you may have worse proformance, bugs that happen seemingly "randomly", or it may even deadlock.
I'll second Fire Lancer's answer - creating your own threads is an excellent way to process big tasks or to handle a task that would otherwise be "blocking" to the rest of synchronous app, but you have to have a clear understanding of the problem that you must solve and develope in a way that clearly defines the task of a thread, and limits the scope of what it does.
For an example I recently worked on - a Java console app runs periodically to capture data by essentially screen-scraping urls, parsing the document with DOM, extracting data and storing it in a database.
As a single threaded application, it, as you would expect, took an age, averaging around 1 url a second for a 50kb page. Not too bad, but when you scale out to needing to processes thousands of urls in a batch, it's no good.
Profiling the app showed that most of the time the active thread was idle - it was waiting for I/O operations - opening of a socket to the remote URL, opening a connection to the database etc. It's this sort of situation that can easily be improved with multithreading. Rewriting to be multi-threaded and with just 5 threads instead of one, even on a single core cpu, gave an increase in throughput of over 20 times.
In this example, each "worker" thread was explicitly limited to what it did - open the remote a remote url, parse the data, store it in the db. All the "high level" processing - generating the list of urls to parse, working out which next, handling errors, all remained with the control of the main thread.
The use of threads makes you think more about the way your application needs threading and can in the long run make it easier to improve / control your performance.
Async methods are faster to use but they are a bit magic - a lot of things happen to make them possible - so it's probable that at some point you will need something that they can't give you. Then you can try and roll some custom threading code.
It all depends on your needs.
The answer is "it depends".
It depends on what you're trying to achieve. I'm going to assume that you're aiming for more performance.
The simplest solution is to find another way to improve your performance. Run a profiler. Look for hot spots. Reduce unnecessary IO.
The next solution is to break your program into multiple processes, each of which can run in their own address space. This is easiest because there is no chance of the individual processes messing each other up.
The next solution is to use threads. At this point you're opening a major can of worms, so start small, and only multi-thread the critical path of the code.
The next solution is to use asynch IO. Generally only recommended for people writing some of very heavily loaded server, and even then I would rather re-use one of the existing frameworks that abstract away the details e.g. the C++ framework ICE, or an EJB server under java.
Note that each of these solutions has multiple sub-solutions - there are different breeds of threads and different kinds of asynch IO, each with slightly different performance characteristics, but again, it's generally best to let the framework handle it for you.