Generalities & Details: Adventures in the High-tech Underbelly

Concurrency and exceptions

Wed, 24 Jun 2009 03:13:52 GMT

I wrote this memo over 2 1/2 years ago about what to do with concurrent exceptions in Parallel Extensions to .NET. Since Beta1 is now out, I thought posting it may provide some insight into our design decisions. (And yes, most design discussions start this way. Somebody develops a personal itch, dives deep into it, and emerges with a proposal for others to vote up, shoot down, or, as is typically the case, somewhere in the middle (provide constructive feedback, iterate, etc).) I've made only a few slight edits (like replacing code- and type-names), but it's mainly in original form. I still agree with much of what I wrote, although I'd definitely write it differently today. And in retrospect, I would have driven harder to get deeper runtime integration. Perhaps in the next release.

~~~

Concurrency and Exceptions
October, 2006

Exceptions raised inside of concurrent workers must be dealt with in a deliberate way. Failures can happen concurrently, and yet often the programmer is working with an API that appears to them as though it’s sequential. The basic question is, then, how do we communicate failure?

The problem

Fork/join concurrency, in which a single “master” thread forks and coordinates with N separate parallel workers, is an incredibly common instance of one of these sequential-looking concurrent operations. The same callback is run by many threads at once, and may fail zero, one, or multiple times. The exception propagation problem is inescapable here and comes with a lot of expectations, because the programmer is presented a traditional stack-based function calling interface papered on top of data or task parallelism underneath.

I am faced with the need for a solution to this problem for PLINQ right now and, while I could invent a one-off solution, we owe it to our customers to come up with a common platform-wide approach (or at least ManyCore-wide). Any solution should compose well across the stack, so that somebody invoking a PLINQ query from within their TPL task that was spawned from a thread pool thread yields the expected and consistent result. And I would like for us to reach consensus for both managed and native programming models.

Before moving on, there is one non-goal to call out. Long-running tasks not under the category of fork/join also deserve some attention, because of the ease with which stack traces can be destroyed and the corresponding impact to debugging, but I will ignore them for now. The problem is not new, exists with the IAsyncResult pattern, and PLINQ doesn’t use this sort of singular asynchronous concurrency. These cases can typically be trivially solved using existing mechanisms, like standard exception marshaling.

No errors, one error, many errors

To understand the core of the issue, imagine we have an API ‘void ForAll<T>(Action<T> a, T[] d)’. It takes a delegate and an array, and for every element ‘e’ in ‘d’ invokes the delegate, passing the element, i.e. ‘a(e)’. If multiple processors are available, the implementation of ForAll may use some heuristic to distribute work among several OS threads, for instance by partitioning the array, probably running one partition on the caller’s thread, and finally joining with these threads before returning so that the caller knows that all of the work is complete when the API returns.

ForAll is not fictitious, and is similar to a number of PLINQ APIs: Where, Select, Join, Sort, etc. It is also exposed directly by the TPL runtime’s Parallel class which intelligently forks and joins with workers.

‘a’ is a user-specified delegate and can do just about anything. That includes, of course, throwing an exception. What’s worse, because ‘a’ is run in several threads concurrently, there may be more than one exception thrown. In fact, there are three distinct possibilities:

No errors: No invocations of ‘a’ throw an exception.
One error: A single invocation of ‘a’ throws an exception.
Many errors: Concurrent invocations of ‘a’ on separate threads throw exceptions.

Clearly letting an exception crash whichever thread the problematic ‘a(e)’ happened to be run on is problematic and confusing. If for no other reason than the IAsyncResult pattern has established precedent. But realistically, the developer would be forced to devise his or her own scheme to marshal the failure back to the calling thread in order for any sort of chance at recovery. They would get it wrong and it would lead to incompatible and poorly composing silos over time. A Byzantine model that fully prohibits exceptions passing fork/join barriers goes against the simple, familiar, and understandable (albeit often deceptively so) model of exceptions.

(That said, marshaling leads to a crappy debugging experience. An already attached debugger will get a break-on-throw notification at the exception on the origin thread, but since we catch, marshal, and (presumably) rethrow, the first and second chances for unhandled exceptions won’t happen until after the exception been marshaled. This breaks the first pass, and by the time the debugger breaks in, or a crash dump is taken, the stack associated with the origin thread is apt to have gone away, been reused for another task (in the case of the thread pool), etc. We generally try to avoid breaking the first pass in the .NET Framework, but do it in plenty of places: the BCL today already contains tons of try { … } catch { /*cleanup */ throw; }-style exception handlers, for example. For this reason I’m not terribly distraught over the implications of doing it ourselves. And sans deeper integration with the exception subsystem – something we ought to consider – there aren’t many reasonable alternatives.)

What makes this problem really bad is that ForAll appears as though it’s synchronous:

void f() {

    // do some stuff

    ForAll(…, …);

    // do some more stuff, ‘ForAll’ is completely done

}

The method call to ForAll itself is synchronous, but of course its internal execution is not. But still, to the developer, the call to this function represents one task, one logical piece of work, regardless of the fact that the implementation uses multiple threads for execution. As higher level APIs are built atop things like ForAll, the low level parallel infrastructure problem becomes a higher level library or application problem. A Sort that is internally parallel must now decide what exception(s) it will tell callers it may throw.

Nondeterministic exception ordering

We assume the ForAll API stops calling ‘a(e)’ on any given thread when it first encounters an exception. That is, each thread just does something like this:

for (int i = start_idx; i < end_idx; i++) {

a(d[i]);

}

The for loop terminates when any single iteration throws an exception. Imagine our array contains 2048 elements and that ForAll smears the data across 8 threads, partitioning the array into 256-element sized chunks of contiguous elements. So partition 0 gets elements [0…256), partition 1 gets [256…512), …, and partition 7 gets [1792…2048). Now imagine that ‘a’ throws an exception whenever fed a null element, and that every 256th element in ‘d’, starting at element 10, is null. What can a developer reasonably expect to happen?

On one hand, if we’re trying to preserve the illusion of sequential execution, we would only want to surface the exception from the 10th element. With a sequential loop, this would have prevented the 266th, 522nd, and so on, elements from even being passed to ‘a’. So we might simply say that the “left most” exception (based on ordinal index) is the one that gets propagated. The obvious problem with this is there are races involved: subsequent iterations indeed may have actually run. Alternatively, we might consider only letting the “first” propagate. Unfortunately, that doesn’t work either, because we unfortunately can’t necessarily determine, for a set of concurrent exceptions, which got thrown first. Even if they have timestamps, they could occur in parallel at indistinguishably close times. Nor does this really matter, because it feels fundamentally wrong.

The reason is that we can’t simply throw away failures without true recoverability in the system, a la STM. The execution of code leading up to the exception did actually happen, after all, and there could be residual effects. We might be masking a terrible problem by throwing failures away, possibly leading to (more) state corruption and (prolonged, perhaps unrecoverable) damage. What if the 10th element was a simple ArgumentNullException that the caller chooses to tolerate, but the 266th element’s exception was in response to a catastrophic error from which the application can’t recover? We can’t choose to propagate the 10th but swallow the 266th. Broadly accepted exceptions best practices suggest that app and library devs never catch and swallow exceptions they cannot reasonably handle. We should do our best to follow the spirit of this guidance too.

Re-propagation

We could employ an approach similar to the IAsyncResult pattern, with some slight tweaks.

If each concurrent copy of ForAll caught any unhandled exceptions and marshaled them to the forking thread, including any exceptions that happen on the forking thread itself, we could then propagate all of them together after the join completes. The question is then: what exactly do we propagate?

If there is just a single exception, it’s tempting to just rethrow it. But I don’t believe this is a good approach for two primary reasons:

This will destroy the stack trace of the original exception. This means no information about the actual source of the error inside ‘a’ is available. With some help from the CLR team, we might be able to get a special type of ‘rethrow’ that copied the original stack trace before recreating a new one. This is already done for remoted exceptions, and the Exception base class will prefix the original remoted stack trace to the new stack trace.
This doesn’t scale to handle multiple exceptions. If we could solve #1, it might be attractive because it appears as-if things happened sequentially, but we can’t escape #2, no matter what we do. We could have different behavior in these two cases, but I believe it’s better to remain consistent instead. Otherwise, developers will need to write their exception handles two ways: one way to handle singular cases, and the other way to handle multiple cases, where the same API may do either nondeterministically.

Given that we need to propagate multiple exceptions, we should wrap them in an aggregate exception object, and propagate that instead. At least this way, the original exceptions will be preserved, stack trace and all. Of course the original exceptions themselves might be other aggregates, handling arbitrary composition.

For sake of discussion, call this aggregate exception System.AggregateException, which of course derives from System.Exception. It exposes the raw Exception objects thrown by the threads, via an ‘Exception[] InnerExceptions’ property, and additional meta-data about each exception: from which thread it was thrown, and any API specific information about the concurrent operation itself. This last part is just to help debuggability. For instance, we might tell the developer that the ArgumentNullException was thrown from a thread pool thread with ID 1011, and that it occurred while invoking the 266th element ‘e’ of array ‘d’. We might also guarantee the exceptions will be stored in the order in which they were marshaled back to the forking thread, just to help the developer (as much as we can) piece together the sequence of events leading to failures.

(Editor’s note: we decided against storing this meta-data information for various reasons.)

Now the dev can do whatever he or she wishes in response to the exception. Previously they might have written:

try {

ForEach(a, d);

} catch (FileNotFoundException fex) {

// Handler(fex);

}

And now they would have to instead write:

try {

    ForAll(a, d);

} catch (AggregateException pex) {

    List<Exception> unhandled = new List<Exception>();

    foreach (Exception e in pex.InnerExceptions) {

        FileNotFoundException fex = e as FileNotFoundException;

        if (fex == null) {

            unhandled.Add(fex);

        } else {

            // Handler(fex);

        }

    }

    if (unhandled.Count > 0)

        throw new AggregateException(unhandled);

}

In other words, they would catch the AggregateException, enumerate over the inner exceptions, and react to any FileNotFoundExceptions as they would have normally. (Taking into consideration that there might have been multiple.) At the end, if there are any non-FileNotFoundExceptions left over, we propagate a new AggregateException with the handled FileNotFoundExceptions removed. If there was only one remaining, we could, I suppose, try to rethrow just that, but this has the same nondeterminism problems mentioned above.

Very few people will write this code. But one of the most vocal arguments against it is: just throw one singular exception, such as ForAllException, and let it crash, because no developer will handle it. Well, that scheme is no better than throwing the AggregateException. At least the aggregation model lets people write backout and recovery code if they have the patience to deal with the reality that multiple exceptions occurred.

To make this slightly easier, we could expose an API, ‘void Handle(Func<Exception, bool> a) where T : Exception’, that effectively encapsulates the same logic as shown above, repropagating the exception at the end if all the exceptions weren’t handled (i.e. some weren’t of type T):

try {

    ForAll(a, d);

} catch (AggregateException pex) {

    pex.Handle(delegate(Exception ex) {

        FileNotFoundException fex = ex as FileNotFoundException;

        if (fex != null) {

            // Handle(fex);

            return true;

        }

        return false;

    });

}

(One problem with this approach is that the ‘throw’ inside of Handle will destroy the original stack trace for ‘pex’. An alternative might be for Handle to modify the AggregateException in place, keeping the stack trace intact, returning a bool that the caller switches on and does a ‘throw’ if it returns false; this is unattractive because it’s error prone and could lead to accidentally swallowing, but in the end might help debuggability.)

If we cared about eliminating unnecessary catch/rethrows, we could use 1st pass filters instead, but this would only be available to VB and C++/CLI programmers, as C# doesn’t expose filters. For example, in pseudo-code:

try {

ForAll(a, d);

} catch (fex.InnerExceptions.Contains<FileNotFoundException>()) {

// Handle …

}

Although interesting, we’re trying to move away from our two pass model. So let’s forget about this for now.

This approach suffers when composing with non-aggregate exception aware code. For it to work well, everybody on the call stack needs to be looking inside the aggregate for “their” exception, handling it, and possibly repropagating. If we want existing BCL APIs to start using data parallelism internally, we would have to be careful here, not to break AppCompat because we start throwing AggregateExceptions instead of the originals.

This is probably where there’s an opportunity for better CLR and tool integration. For instance, you could imagine a world where the CLR automatically unravels the parallel failures, matching and running handlers for specific exceptions inside the aggregate as it goes, but repropagating if all exceptions weren’t handled. This is very hand-wavy and fundamentally changes the way exceptions work, so it would require a lot more thought. A catch block that swallows an exception (today) is just about guaranteed—asynchronous exceptions aside—that the IP will soon reach the next instruction after the try/catch block. This is a pretty basic invariant. With this proposal, that wouldn’t be the case, and would be bound to break large swaths of code. Sticking with the library approach (with all its imperfections) seems like the best plan of attack for now.

Waiting for the “join” to finish

There was something implicit in the design mentioned above. The ForAll API, and others like it, wouldn’t actually propagate exceptions until the fate of all threads was known.

Imagine we have the scenario described earlier (2048 elements, 8 threads), but slightly different: the 0th element causes an exception, but no other. It turns out this is probably a common case, i.e. that only a subset of the partitions will yield an exception. In this case, we would still have to wait for 7*256 = 1,792 elements to be run through ‘a’ before this exception is propagated. Imagine a slightly different case. The 0th element throws a catastrophic exception, and the application is going to terminate as soon as it propagates. ‘a’ simply can’t be run any more, and will keep reporting back this same exception. But it will take 8 of these exceptions to actually stop the application, i.e. by calling ‘a’ on the 0th, 256th, 512th, etc. elements, if we wait for all tasks to complete. If each exception corresponds to some failed attempt at forward progress, one that possibly corrupts state, then the damage is O(N) times “worse” (for some measurement) than in the sequential program, where N is the number of concurrent tasks.

Instead of waiting helplessly, we could try to aggressively shut down these concurrent workers.

At first, you might be tempted to employ CLR asynchronous thread aborts, but this is fraught with peril. Almost all .NET Framework code today is taught that thread abort == AppDomain unload, and reacts accordingly. State corruption stemming from libraries as fundamental as the BCL would be just about guaranteed. Changing this state of mind and the state of our software would be quite the undertaking.

Instead, we can have the concurrent API itself periodically check an ‘abort’ flag shared among all workers. The first thread to propagate an exception would set this flag. And whenever a worker has seen that it has been set, it voluntarily returns instead of finishing processing data:

for (int i = start_idx; i < end_idx && !aborted; i++) {

a(d[i]);

}

This increases the responsiveness of exception propagation, but clearly isn’t foolproof. There will still be a delay for long-running callbacks. Thankfully, with PLINQ, TPL, and I hope most of our parallel libraries, the units of work will be individually fine-grained, and therefore this technique should suffice.

If a concurrent worker is blocked, there’s not a whole lot we can do. Much like thread aborts, you might be tempted to use Thread.Interrupt to remove it from the wait condition. Unfortunately this will leave state corruption in its wake, because plenty of code does things like WaitHandle.WaitOne(Timeout.Infinite) without checking the return value or expecting a ThreadInterruptionException. The same argument applies to, say, user-mode APCs. Eventually you might also be tempted to use IO cancellation in Windows Vista to cancel errant, runaway network or disk IO requests. This would be great. But this also generally has the same problem as interruption, so until we find a general solution to that, we can’t do any of this.

(Editor’s note: We eventually solved this problem by coming up with a unified cancellation framework.)

One last note

This path forward seems best for now, but it leaves me wanting more.

In the end, this feels like a more fundamental problem. An API like ForAll gives the illusion of an ordinary, old sequential caller/callee relationship. But the callee doesn’t use a stack-based calling approach: instead, it distributes work among many concurrent workers, turning the linear stack into a sort of dynamically unfolding cactus stack (or tree). And SEH exceptions are fundamentally linear stack-based creatures.

In this world, it’s just a simple fact that data all over the place can become corrupt simultaneously. Many things can fail at once because many things are happening at once. It’s inescapable. Recovery is disastrously difficult, so most failures will end in crashes. STM’s promise for automatic recovery offers a glimmer of hope, but without it, I worry that papering a sequential “feel” on top of data/task parallelism is a dangerous game to play.

Exploring memory models

Wed, 17 Jun 2009 06:53:26 GMT

One of my many focuses lately has been developing a memory ordering model for our project here at Microsoft. There are four main questions to answer when defining such a model:

What are the ordering guarantees for ordinary loads and stores?
What are the ordering guarantees for volatile loads and stores?
What kinds of explicit fences are allowed?
Where are fences used automatically, e.g. to preserve type safety and security?

These tend to be the differentiation points for any model. Everything else is mostly commodity. Not that there is much else, mind you, but respecting data dependence, not speculating ahead such that exceptions occur that wouldn't have occurred in a sequential execution, and so forth are all must haves, for instance. Most interesting permutations of answers for these questions have already been explored, and industry consensus is being reached, so it would be better to say I've been picking a model rather than defining one.

What's interesting is that memory model designers are often colored by their favorite architecture du jour. If somebody cares primarily about X86, they are apt to choose something very strong. If somebody cares primarily about ARM, however, they are apt to choose something very weak. There is a classic tradeoff here. Stronger means easier to program, while weaker means better performance. For some reason, many of the projects I've worked on have had an abundance of strong hardware (like X86) and a scarcity of weak hardware (like ARM and IA64). The reality sinks in: most developers on the team code to X86, and then when it comes time to getting more serious about the other platforms, code starts breaking all over the place. This is why the CLR went so strong in 2.0, even though IA64 was an important platform to support.

Let's look at some common answers to the above questions.

For #1:

C++, Visual C++, ECMA 1.0, Java Memory Model, and Prism: no ordering guarantees.
CLR 2.0: ordered stores, no ordering for loads.

For #2:

C++: prevents compiler-only code motion, but explicit fences are needed for processor ordering.
Visual C++, ECMA 1.0, and CLR 2.0: loads are acquire, stores are release ordered.
Java Memory Model: loads and stores are fully ordered (sequentially consistent).

For #3:

C++: implementation-specific.
Visual C++: intrinsics and Win32 APIs.
ECMA 1.0 and CLR 2.0: locks, and mostly Win32-style interlocked APIs.
Java Memory Model: locks, compare-and-swap, atomics, etc.

For #4:

Managed environments like the CLR and JVM need to ensure type safety, even if ordinary loads and stores are unordered. This is nontrivial, because the boundary around type safety is blurred. Certainly we must ensure garbage v-table pointers are not seen. But is a thread allowed to read non-zeroed memory behind an object reference? And can it contain garbage (e.g. "values out of thin air")? What about writes done by mutator threads, including write barriers, while a concurrent collector is tracing objects in the heap? Are array lengths part of the set of protected fields that mustn't be read out of order? Strings, since they are commonly used for security checking? And so on.

It is mainly the deep questions around #4, and also some simple compatibility struggles (around things like double checked locking), that caused the stronger answers for #1 in the CLR 2.0.

In any case, I'm advocating a very different approach than the traditional models.

We pick completely weak ordering for ordinary loads and stores, to enable efficient execution on weaker platforms like ARM, PowerPC, IA64, etc. That part isn't new. But here's the clincher. No volatiles. There are special variables that are used to communicate between threads (call them volatile if you'd like), but using them implies no kind of special automatic fencing. Instead, whenever accessing such a variable, at the site of usage, the kind of fence desired must be used (compiler-enforced): full-fence (sequentially consistent), acquire-fence, release-fence, no-fence, or compiler-only-fence (for things like ensuring loads don't get hoisted as loop invariant). Of course, certain kinds of fences are sprinkled throughout the system to guarantee type safety in all of the aforementioned places (and more), but these are implementation details.

(This approach is rather like Herb Sutter's Prism and C++0x atomics. See http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2664.htm.)

Particularly after managing teams who developed a plethora of lock free code, I love this approach. I can review code and immediately understand what ordering invariants the developer assumed when writing the code. This doesn't really make writing lock free code any simpler, except that it forces you to pause and think about things a bit more carefully than you may have otherwise. But it certainly makes code easier to understand and maintain, and makes it clear to people that sprinkling volatile all over the place isn't going to save your butt: the only thing that will do that is careful thinking and engineering.

A scalable reader/writer scheme with optimistic retry

Fri, 05 Jun 2009 00:36:23 GMT

An interesting alternative to reader/writer locks is to combine pessimistic writing with optimistic reading. This borrows some ideas from transactional memory, although of course the ideas have existed long before. I was reminded of this trick by a colleague on my new team just a couple days ago.

The basic idea is to read a bunch of state optimistically, without taking a lock of any sort, and then prior to using it for meaningful work (which may depend on the state being consistent and correct), a validation step must take place. This validation uses version numbers which writers are responsible for maintaining. Specifically, we'll use two version counters, version1 and version2: the writer increments version1, performs the writes, and then increments version2; and the reader reads version2, performs its reads, and then verifies that version 1 is equal to the version2 that it saw at the start. If this verification fails, we'll ordinarily just do a little spinning and then go back around the loop again.

Stop for a moment and ponder something very critical to this algorithm. The writer increments variables in the opposite order of the reader's reads. To see why this works, imagine we start with version1 == version2 == 0. There are two hazards to worry about. (1) A reader begins reading, and writes occur before it has finished. And (2) a reader begins reading while a write is in progress. These are simple to detect, and in fact boil down to the same thing. A reader sees version2 == 0, and the first thing a writer does is version1++. So when the reader attempts to validate, it will notice the version2 it saw != version1 any longer. If the writer has already begun by the time the reader arrives, it is possible for the reader to know it is doomed even before it has started doing any of its reads.

Here is the code in its full glory:

using System;

using System.Threading;

public class OptimisticSynchronizer

{

private volatile int m_version1;

private volatile int m_version2;

public void BeforeWrite() {

++m_version1;

}

public void AfterWrite() {

++m_version2;

}

public ReadMark GetReadMark() {

return new ReadMark(this, m_version2);

}

public struct ReadMark

{

private OptimisticSynchronizer m_sync;

private int m_version;

internal ReadMark(OptimisticSynchronizer sync, int version) {

m_sync = sync;

m_version = version;

}

public bool IsValid {

get { return m_sync.m_version1 == m_version; }

}

public void DoWrite(Action writer) {

BeforeWrite();

try {

writer();

} finally {

AfterWrite();

}

public T DoRead<T>(Func<T> reader) {

T value = default(T);

SpinWait sw = new SpinWait();

while (true) {

ReadMark mark = GetReadMark();

value = reader();

if (mark.IsValid) {

break;

}

sw.SpinOnce();

}

return value;

}

We leave it to the caller of this class to acquire locks as appropriate to synchronize writers. Typically this will just mean wrapping a Monitor.Enter/Exit around calls to things like BeforeWrite, AfterWrite, and DoWrite. But readers explicitly do not need this same protection. DoRead exemplifies the safe reading pattern, although it can be done by hand via the ReadMark APIs.

It's also worth considering what kinds of fences are truly required for this to work. Logically speaking, we need to ensure the entrance to a protected block (either read or write) is an acquire fence, and exit from the block is a release fence. This is similar to the ordering semenaitcs necessary for a lock block. So long as we use volatile modifiers for the version counters, and for the variables read within the protected block, this will work fine. Even on weak models like IA64. The beautiful thing is that we don't need full fences, even on models like X86 that make use of store buffer forwarding The classic store buffering case we may worry about (on a single-threaded execution) would be something like this, in pseudo-code:

version1++;

X = 42;

Y = 99;

version2++;

tmp = version2;

r0 = X;

r1 = Y;

success = (tmp == version1);

We'd be worried about satisfying some loads out of the store buffer, while satisfying others out of the memory system. But this is safe: if the load of X or Y sees a different processor's writes, then the subsequent load of version1 necessarily must witness the new value written by the other processor too. And therefore the validation will fail as we would expect and hope.

Here is a quick performance benchmark I whipped together, much in the same spirit as my previous reader/writer lock examples. I've measured varying numbers of writers (0%, 5%, 10%, 25%, 50%, and 100%), and each thread spends a certain amount of time inside the "lock region" doing some nonsense busy work. The certain amount of time is measured in terms of number of function calls (0, 10, 100, and 1000), and the work doesn't vary at all depending on whether a thread is reading or writing. I've measured four things: (1) Monitor.Enter/Exit as the baseline (where both readers and writers just acquire the mutually exclusive lock), (2) ReaderWriterLockSlim, (3) the spin-based lock that I showed previously, and (4) the new OptimisticSynchronizer class with optimistic retry. The values are the ratio compared to the baseline (1), so that >1.0x means the particular entry is slower, while <1.0x is faster. I did these measurements on an 8-way machine -- unlike the previous study which was on a 4-way machine -- which means that 0.125x would be a linear speedup compared to the serialized Monitor version:

0% writers: