Pop quiz: Can this code deadlock?

A few weeks back I recorded a discussion with the infamous Erik Meijer and Charles from Channel9.

Perspectives on Concurrent Programming and Parallelism
http://channel9.msdn.com/shows/Going+Deep/Joe-Duffy-Perspectives-on-Concurrent-Programming-and-Parallelism/

In it, I show my cards a bit more than intuition says I should.  I'm not good at poker.

To summarize:

• Mostly functional (purity + immutability) is a great default.
• Safe, determinstic mutability (a la runST) is a must-have for cognitive familiarity.
• Isolation is key to achieve the former; type systems can help (a lot).
• Actors, agents, forkIO, <what have you> is a good model, but not the only one.  Isolation is (far) more general.
• Transactions can help around the edges.

I'm working on a few papers for public consumption this year where I espouse these ideas.  Keep watching for more detail.

I was very harsh in my previous post about reader/writer locks.

The results are clearly very hardware-specific.  And one can certainly argue that better implementations are possible.  (In fact, I will show one momentarily.)  But no matter which way you slice-and-dice it, a lock implies mutable shared state which implies contention.  Herb argued this point quite well, and rather thoroughly, in his recent Dr. Dobb’s article.  Interference due to contention means more time spent resolving memory conflicts and less time doing useful work.  A reader/writer lock can be infinitely clever, but there is still a consensus protocol that must be established: and that implies a loss of scalability.  Pretty simple.

It’s very tricky to develop a consensus protocol that is sufficiently lossy so as to relieve memory contention while at the same time being sufficiently precise that the lock works right.  In the case of a spinning reader/writer lock (which is, for what it’s worth, overly naïve an approach for most circumstances), you need to ensure that a writer knows for sure when there are 0 readers, and that each reader knows for sure whether there is 0 or 1 writer.  (For blocking reader/writer locks, there’s a whole lot more.)  One promising thing to note is that the writer only needs to know whether there are 0 or N readers, but not the specific value N; there’s a fair bit of research on scalable counters (like this) which exploit problems of this nature.  Unfortunately, it’s not completely relevant here.  You need to know exactly when the transition from N to 0 readers happens in order to let the writer through in a timely fashion; and in order to account for that transition, a consensus among readers is needed.  That's hard to do.

More scalable solutions are possible than the simple lock I showed previously.  Although writers need to know whether readers are present, the readers themselves could care less about other readers.  As a result, we can make the lock slightly more expensive for the writer, because it needs to accumulate the count of readers, but this allows us to make it it slightly cheaper for the readers to enter and exit.  Where cheaper means less contention.

Here’s one possible algorithm.  We’ll keep an array of read flags and a single write flag:

private volatile int m_writer;

private ReadEntry[] m_readers = new ReadEntry[Environment.ProcessorCount * 16];

A few things are noteworthy about the read flags.

First, it’s an array of ReadEntry values.  These are just simple structs that wrap a volatile int, but we also pad the struct so that it’s 128 bytes in total size.  That avoids the situation where multiple read flags just happen to end up sharing the same cache line (which are usually either 64 or 128 bytes in size), which leads to false sharing in the memory system (destroying our aim to reduce contention).

[StructLayout(LayoutKind.Sequential, Size = 128)]

struct ReadEntry {

    internal volatile int m_taken;

}

Second, we size the array to be 16-times the number of processors.  We hash into it based on the calling thread’s unique identifier, so to reduce (but not eliminate) the chance of hashing collisions, we’ll use a few times more buckets than the total number of concurrent threads.  Hashing collisions are expensive: they incur some amount of memory contention, and also demand that we use an atomic CAS increment instead of an ordinary ++.  (While a super-duper-cheap TLS solution might seem more ideal, there isn’t any good per-object TLS solution to use.  The array hashing approach is actually quite fast.)

Notice that we’re using an awful lot of space for a single lock.  This means the techniques I show here wouldn’t be readily applicable to a system that uses lots of fine-grained locks, like transactional memory.  But similar ideas can be extrapolated, e.g., by using shared lock tables.

Lastly, some invariants among these fields are self-evident.  When the writer flag is 0, no writers are waiting; when it is 1, either a writer is actively in the critical section, or there is a writer waiting for readers to exit.  When at least one reader flag entry is non-0, there is a reader either inside the lock or attempting to enter it.  Thus, no new writer is permitted while there’s a non-0 reader entry, and no new reader is permitted while there’s a non-0 writer flag.  This is sufficient to ensure the reader/writer lock properties hold.

Now let’s look at how the EnterReadLock and ExitReadLock methods work.

When a reader arrives, it spins until the writer flag is non-0.  It then hashes into the read flag array using its unique thread identifier, and then atomically increments the read counter.  It then needs to recheck that a writer didn’t arrive in the meantime.  (The CAS increment means we can safely do this without worry for reordering bugs, like the read of the writer flag passing the write to the reader flag.)  If a writer hasn’t arrived, the read lock has been successfully acquired and we’re done; if a writer has arrived, however, the reader needs to back out the change (since the writer might be waiting for the read flag to become 0) and then go back to spinning.  It will retry again once the writer exits.

private int ReadLockIndex {

    get { return Thread.CurrentThread.ManagedThreadId % m_readers.Length; }

}

 

public void EnterReadLock() {

    SPW sw = new SPW();

    int tid = ReadLockIndex;

 

    // Wait until there are no writers.

    while (true) {

        while (m_writer == 1) sw.SpinOnce();

 

        // Try to take the read lock.

        Interlocked.Increment(ref m_readers[tid].m_taken);

        if (m_writer == 0) {

            // Success, no writer, proceed.

            break;

        }

 

        // Back off, to let the writer go through.

        Interlocked.Decrement(ref m_readers[tid].m_taken);

    }

}

(Note that SPW is a little type to encapsulate the spin-wait logic, including some amount of backoff to reduce contention.  An example implementation at the bottom of this essay, along with the full reader/writer lock code.  .NET 4.0 includes a SpinWait type that provides this same functionality.)

Exiting the read lock is pretty simple.  We just need to decrement our counter.

public void ExitReadLock() {

    // Just note that the current reader has left the lock.

    Interlocked.Decrement(ref m_readers[ReadLockIndex].m_taken);

}

The writer lock is pretty straightforward.  It works the same way most spin-based mutually exclusive locks work, but using a CAS on the writer flag, but has an extra step after successfully acquiring the lock: a writer must walk the list of read flags, and wait for each one to become 0.  (This is similar to Peterson's mutual exclusion algorithm for N-threads.)  Because the write flag is set first (using a CAS), and because new readers won’t enter if the flag is set, we can be assured this works correctly without hokey memory reordering problems cropping up.

public void EnterWriteLock() {

    SPW sw = new SPW();

    while (true) {

        if (m_writer == 0 && Interlocked.Exchange(ref m_writer, 1) == 0) {

            // We now hold the write lock, and prevent new readers.

            // But we must ensure no readers exist before proceeding.

            for (int i = 0; i < m_readers.Length; i++)

                while (m_readers[i].m_taken != 0) sw.SpinOnce();

 

            break;

        }

 

        // We failed to take the write lock; wait a bit and retry.

        sw.SpinOnce();

    }

}

And exiting the write lock is even simpler than exiting the read lock.  We just set the writer flag to 0.

public void ExitWriteLock() {

    // No need for a CAS.

    m_writer = 0;

}

Given all of that, you might wonder how well this bad boy performs.  Well, single-threaded performance is a bit worse than the previous spin reader/writer lock: about 1.55x the cost of a monitor acquisition for the read lock instead of 0.95x, and about 5.52x for the write lock instead of 0.85X.  This makes sense.  There’s simply a whole lot more work going on in this new lock compared to the old, simple one.

But scalability is vastly improved.  Our hard work has apparently paid off.  Here’s a table much like the one in the previous post: scaling over the equivalent mutually exclusive monitor code, for various percentages of writers and various amounts of "work" (counts of function calls) inside the lock region.  (I have left out the legacy .NET ReaderWriterLock type because it is embarassingly terrible.)  Remember: 1.0x means it scales the new lock is the same as monitor, 0.5x means twice as fast, and 2.0x means twice as slow.  0.25x is ideal speedup (4x) since I am running the tests on a four way machine.

0% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.11x         2.01x         0.96x         0.32x

SpinRWL(old)  9.63x         7.04x         1.02x         0.26x

SpinRWL(new)  0.39x         0.36x         0.28x         0.25x

 

5% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.29x         2.36x         1.18x         0.61x

SpinRWL(old)  5.69x         5.59x         1.43x         0.94x

SpinRWL(new)  1.01x         0.96x         0.45x         0.38x

 

10% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.26x         2.04x         1.15x         1.00x

SpinRWL(old)  6.87x         5.03x         1.42x         1.34x

SpinRWL(new)  1.60x         1.51x         0.63x         0.53x

 

25% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.09x         2.10x         1.14x         1.00x

SpinRWL(old)  4.70x         4.20x         1.43x         1.69x

SpinRWL(new)  2.81x         2.29x         1.27x         0.73x

 

50% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.18x         1.95x         1.15x         0.95x

SpinRWL(old)  3.23x         3.73x         1.54x         1.39x

SpinRWL(new)  3.16x         2.76x         1.73x         1.10x

 

100% writers:

              0 calls       10 calls      100 calls     1000 calls

RWLSlim (3.5) 2.18x         1.95x         1.04x         0.92x

SpinRWL(old)  2.63x         2.04x         1.06x         0.87x

SpinRWL(new)  6.79x         3.96x         1.62x         1.06x

You can see there are now several more cases where the new reader/writer lock beats out both the .NET 3.5 ReaderWriterLockSlim type in addition to our previous attempt.  In fact, we now have a few new scenarios that scale, like 5% or 10% writers where the amount of work being done is at least 100 function calls.  (Unfortunately, doing 100 or more function calls inside a lock that uses spin-waiting is dangerous and considered a very bad practice: you should be able to count the number of instructions on your fingers (and toes).  But that’s somewhat beside the point.)  In summary, so long as there is a fair amount of work going on and the percentage of writers remains very low, we might see a benefit.

So was I overly harsh on reader/writer locks in my last post?  Sure, maybe a little.  While I am still very disappointed in the current .NET reader/writer locks (and, I imagine, the Vista SRWLock), the results I was able to get here are a bit more promising.

But the point I was trying to get across is the same: sharing is sharing is sharing.  Avoid it like the plague.

(Thanks to Tim Harris for sending me private email about my previous posts.  The brief discussion inspired me to pick this back up.)

Here’s the full code for the reader/writer lock.

using System;

using System.Threading;

using System.Collections.Concurrent;

using System.Collections.Generic;

using System.Runtime.InteropServices;

// We use plenty of interlocked operations on volatile fields below.  Safe.

#pragma warning disable 0420

/// <summary>

/// A very lightweight reader/writer lock.  It uses a single word of memory, and

/// only spins when contention arises (no events are necessary).

/// </summary>

public class ReaderWriterSpinLockPerProc {

    private volatile int m_writer;

    private volatile ReadEntry[] m_readers = new ReadEntry[Environment.ProcessorCount * 16];

    [StructLayout(LayoutKind.Sequential, Size = 128)]

    struct ReadEntry {

        internal volatile int m_taken;

    }

    private int ReadLockIndex {

        get { return Thread.CurrentThread.ManagedThreadId % m_readers.Length; }

    }

    public void EnterReadLock() {

        SPW sw = new SPW();

        int tid = ReadLockIndex;

        // Wait until there are no writers.

        while (true) {

            while (m_writer == 1) sw.SpinOnce();

            // Try to take the read lock.

            Interlocked.Increment(ref m_readers[tid].m_taken);

            if (m_writer == 0) {

                // Success, no writer, proceed.

                break;

            }

            // Back off, to let the writer go through.

            Interlocked.Decrement(ref m_readers[tid].m_taken);

        }

    }

    public void EnterWriteLock() {

        SPW sw = new SPW();

        while (true) {

            if (m_writer == 0 && Interlocked.Exchange(ref m_writer, 1) == 0) {

                // We now hold the write lock, and prevent new readers.

                // But we must ensure no readers exist before proceeding.

                for (int i = 0; i < m_readers.Length; i++)

                    while (m_readers[i].m_taken != 0) sw.SpinOnce();

                break;

            }

            // We failed to take the write lock; wait a bit and retry.

            sw.SpinOnce();

        }

    }

    public void ExitReadLock() {

        // Just note that the current reader has left the lock.

        Interlocked.Decrement(ref m_readers[ReadLockIndex].m_taken);

    }

    public void ExitWriteLock() {

        // No need for a CAS.

        m_writer = 0;

    }

}

struct SPW {

    private int m_count;

    internal void SpinOnce() {

        if (m_count++ > 32) {

            Thread.Sleep(0);

        } else if (m_count > 12) {

            Thread.Yield();

        } else {

            Thread.SpinWait(2 << m_count);

        }

    }

}

A couple weeks ago, I illustrated a very simple reader/writer lock that was comprised of a single word and used spinning instead of blocking under contention.  The reason you might use a lock with a read (aka shared) mode is fairly well known: by allowing multiple readers to enter the lock simultaneously, concurrency is improved and therefore so does scalability.  Or so the textbook theory goes.

As a purely theoretical illustration, imagine we’re on a heavily loaded 8-CPU server where a new request arrives every 0.25ms and runs for 1ms.  In an ideal world, we could service requests coming in at a rate of 1ms / 8-CPUs = 0.125ms without falling behind.  But imagine these requests need to access some shared state, and so there is a bit of serialization required.  In fact, let’s imagine each does 0.5ms’ worth of its work inside a lock.  If you were to use a mutually exclusive lock, then you’d have an immediate lock convoy on your hands.  Even with 8-CPUs you won’t be able to keep up.  You’ll start off gradually building up a debt, and eventually come to a crawl.  Let’s examine the initial timeline:

Req#          Arrival       Acquire       Release       Wait Time

1             0.0ms         0.0ms         0.5ms         0.0ms

2             0.25ms        0.5ms         1.0ms         0.25ms

3             0.5ms         1.0ms         1.5ms         0.5ms

4             0.75ms        1.5ms         2.0ms         0.75ms

5             1.0ms         2.0ms         2.5ms         1.0ms

6             1.25ms        2.5ms         3.0ms         1.25ms

7             1.5ms         3.0ms         3.5ms         1.5ms

8             1.75ms        3.5ms         4.0ms         1.75ms

Oh jeez, after only the first 8 requests, we’ve fallen way behind.

Each new request adds 0.25ms onto the amount of time the request must wait for the lock.  And it’s not going to get any better:

9             2.0ms         4.0ms         4.5ms         2ms

10            2.25ms        4.5ms         5.0ms         2.25ms

11            2.5ms         5.0ms         5.5ms         2.5ms

12            2.75ms        5.5ms         6.0ms         2.75ms

 

... and so on ...

By request #9, requests have to wait for twice as long as they run.  Eventually something has to give, or the server will come tumbling down.

Now, imagine we used a reader/writer lock instead.  Threads would never wait for each other, and we wouldn’t end up with this never-ending buildup of wait times.  In other words, the “Wait Time” column above would always be 0.0ms.  And because the arrival rate is less than our theoretical limit of one request per 0.125ms, our lock convoy is gone.  Right?

Unfortunately, probably not; this mental model is overly naïve.

Even when a read lock is acquired, there is mutual exclusion going on:

• Some reader/writer locks actually use mutually exclusive locks to protect their own internal state, like the list of current readers!  This can come as a surprise, but it’s true of the .NET reader/writer locks.  Vance’s example even does and, although it uses a spin lock in an attempt to reduce the overhead, there’s still no denying that it’s mutual exclusion.
• And even if they don’t use mutually exclusive locks, like the simple spin-based one from my previous blog post, there are CAS instructions.  And a CAS instruction actually amounts to a form of mutual exclusion at the hardware level, because the cache coherency machinery needs to ensure that no two processors try to acquire and modify the same cache line exclusively.
• In addition to all of that overhead, the cost in CPU cycles of acquiring the read-lock is nowhere near zero.  Because of the use of locks and/or CAS internally, and the resulting cache contention and line evictions that this will cause, throughput will suffer.  If there is contention, threads may end up blocked (if real locks are being used), spinning (if spin locks are being used), or simply optimistically retrying CAS’s due to line ping ponging.

The result?

Read locks are just as bad as mutually exclusive locks when lock hold times are short.  In fact, they can be worse, because reader/writer locks are more complicated and therefore cost more than simple mutually exclusive locks: many need to keep track of read lists in order to disallow recursive acquires, maintain multiple event handles so certain kinds of waiters can be awakened over others, and store various kinds of counters and flags.  Even my super simple single-word, spin-based reader/writer lock needed to worry about blocking out readers when a writer was waiting, properly incrementing and decrementing the reader count when readers are racing with one another (leading to more complicated CAS on the exit path than ordinary write locks), and so on.

That said, a reader/writer lock would in fact probably work in the situation above.  A hold time of 0.5ms is huge, and with only 8 concurrent threads and the arrival rate we’re talking about, the overheads are apt to be quite small in relation to the work being done.  Another similar setting in which reader/writer locks commonly make a noticeable difference is in the execution of large database transactions.

But the sad truth is that we tell programmers to keep lock hold times short, and most locks I see are comprised of two dozen instructions or less.  So we’re in the microsecond range at the very most, which is certainly not large enough for read locks to pan out.

To illustrate this point, I wrote a little benchmark program that benchmarks the legacy .NET ReaderWriterLock, the 3.5 ReaderWriterLockSlim type, and my little spin reader/writer lock.  All it does is spawn 4 threads on my dual-socket, dual-core (4-CPU) machine, and then loop around so many times acquiring and releasing a certain kind of lock.  I’ve written the test so that the amount of work done inside the lock is parameterized as a certain number of non-inlined function calls.  I also parameterize the percentage of acquires that will be write-locks.  Then I’ve run this a bunch of times, and compared the total time taken with the equivalent code using a CLR Monitor for mutual exclusion instead.

Here are some results, where each column represents the number of function calls.  The entries are the cost relative to Monitor: 1.00x means they are the same, 0.5x means the alternative lock is twice as fast, and 2.0x means the alternative lock is twice as slow.  Remember, the ideal situation would be 0.25x: that is, by allowing four threads to run completely concurrently, we run four times faster.

0% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         9.23x         6.46x         0.90x         0.49x

RWLSlim (3.5)        2.11x         2.01x         0.96x         0.32x

SpinRWL              9.63x         7.04x         1.02x         0.26x

 

5% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         10.55x        8.23x         1.71x         0.63x

RWLSlim (3.5)        2.29x         2.36x         1.18x         0.61x

SpinRWL              5.69x         5.59x         1.43x         0.94x

 

10% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         20.31x        10.39x        2.34x         0.99x

RWLSlim (3.5)        2.26x         2.04x         1.15x         1.00x

SpinRWL              6.87x         5.03x         1.42x         1.34x

 

25% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         74.49x        49.59x        9.18x         2.15x

RWLSlim (3.5)        2.09x         2.10x         1.14x         1.00x

SpinRWL              4.70x         4.20x         1.43x         1.69x

 

50% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         148.34x       98.46x        20.46x        3.63x

RWLSlim (3.5)        2.18x         1.95x         1.15x         0.95x

SpinRWL              3.23x         3.73x         1.54x         1.39x

 

100% writers:

                     0 calls       10 calls      100 calls     1000 calls

RWL (legacy)         170.59x       123.66x       24.04x        4.29x

RWLSlim (3.5)        2.18x         1.95x         1.04x         0.92x

SpinRWL              2.63x         2.04x         1.06x         0.87x

Clearly there are a number of anomalies in these numbers.  Why the legacy ReaderWriterLock balloons to 170X the cost of Monitor when we have 100% writers is a very interesting question indeed.  Why my simple spin reader/writer lock is 9.63X when we have pure reads and 0 calls, and yet the ReaderWriterLockSlim type is only 2.11X is also interesting.  And so on.  The numbers are very specific to the version of .NET I am using, and indeed the precise machine configuration, including the number and layout of cores and caches.

But if we look more generally at the numbers, ignoring some of the surprising ones, we can make one interesting and safe conclusion: You need a really low percentage of writers, and a really long amount of time inside the lock, for any scalability wins to show up as a result of using a reader/writer lock.  Our best case was the spin reader/writer lock when we had 0% writers and 1000 calls.  But clearly if you have no writers, i.e., state is immutable, there’s little point in using any locks whatsoever!  This is an extreme result, where threads are hammering on the lock constantly in a tight loop, but if you stop to think about it: When else would a reader/writer lock make a difference?  If threads are just getting in and out of the lock very quickly, and arrivals are infrequent, then there is no benefit to allowing multiple threads in at once anyway.

The moral of the story?  Besides suggesting that you seriously question whether a reader/writer lock is actually going to buy you anything, it's the same as the conclusion in my previous post on the matter:

Sharing is evil, fundamentally limits scalability, and is best avoided.

I frequently get asked about the C# compiler's warning CS0420 about taking byrefs to volatile fields.  For example, given a program

class P {
    static volatile int x;

    static void Main() {
        f(ref x);
    }

    static void f(ref int y) {
        while (y == 0) ;
    }
}

the C# compiler will complain

xx.cs(8,15): warning CS0420: 'P.x': a reference to a volatile field will not be treated as volatile

because of the line containing 'ref x'.  (The same applies to 'out' parameters too.)  The natural question is, of course, whether to worry about it.

In general, the answer is yes, you must worry.  In the above example, the use of the 'y' parameter inside 'f' will not be treated as volatile, as the warning says.  What does that mean in practice?  For one, the read of 'y' in 'f's while loop could be considered loop invariant by the JIT compiler and hoisted, and you'd possibly loop forever.  It also means that on IA64 platforms, such reads will be emitted as ordinary loads instead of the special load-acquire variant that is emitted for volatile loads.  This can lead to reordering bugs.  In other words, you lose the volatile-ness of the field as soon as you cast it away as an ordinary byref.  And unlike C++ where you can have a volatile pointer, there's no way to mark a .NET byref as volatile.

(You can use the Thread.VolatileRead and VolatileWrite methods to use a byref in a volatile manner.  Unfortunately they are far more costly than ordinary volatile loads and stores.)

There is one particularly annoying case in which this warning is complete noise: when passing a byref to an API that internally performs volatile (or stronger) loads and stores.  I.e., the Interlocked.*, Thread.VolatileRead, and VolatileWrite methods.  Because these APIs internally use explicit memory barriers and atomic hardware instructions, the byref will effectively be treated as volatile regardless of whether it was taken from a volatile field or not.  And therefore it is safe.

For instance, the compiler will warn you about the following code

volatile int x;

static void f() {
    Interlocked.Exchange(ref x, 1);
}

even though there is no problem.  You can suppress the warning with a "#pragma warning disable" just before the call

volatile int x;

static void f() {
#pragma warning disable 0420
    Interlocked.Exchange(ref x, 1);
#pragma warning restore 0420
}

and then restore it immediately afterwards.  (It's a good idea to restore the warning so that you catch other possibly-problematic instances from being missed.)

This comes up a whole lot.  Why?  Because many times you'll mark a field volatile, even though it is updated exclusively with CAS operations, because it's also used in other contexts: e.g., sequences where loads mustn't reorder or erroneously be considered loop invariant.  I personally have a habit of always marking these variables as such, mostly as a carryover from Win32 whose InterlockedXX family of APIs demand volatile pointers (i.e., volatile * LONG).

I'm told that this annoying case might be fixed in the next C# compiler, by the way.  Until then, I figured I'd throw this up for reference purposes.

Reader/writer locks are commonly used when significantly more time is spent reading shared state than writing it (as is often the case), with the aim of improving scalability.  The theoretical scalability wins come because the lock can be acquired in a special read-mode, which permits multiple readers to enter at once.  A write-mode is also available which offers typical mutual exclusion with respect to all readers and writers.  The idea is simple: if many readers can read simultaneously, the theory goes, concurrency improves.

(I’ll be posting an analysis of reader/writer lock scalability in an upcoming post.  For a variety of reasons--most related to my recent CAS post--they seldom make a dramatic impact in practice.)

In addition to showing up in libraries--such as Vista’s new SRWLock, .NET’s ReaderWriterLock, and .NET 3.5’s ReaderWriterLockSlim--they are used pervasively in relational databases, distributed transactions, and software transactional memory.

Vance Morrison demonstrated a lightweight reader/writer lock on his blog a couple years back.  Although quite small, you can get smaller.  Much like the new SpinLock type being made available in .NET 4.0, we can build a ReaderWriterSpinLock that offers several advantages:

1. It’s a struct, and so there is no object allocation or space for an object header necessary.
2. It’s a single word in size (i.e., 4 bytes).
3. No kernel events are ever allocated; we will spin instead.

For cases in which reads are extraordinarily frequent, and writes are extraordinarily rare, this approach can actually be useful.  Unfortunately, because one common case in which reader/writer locks scale very well is when hold times are lengthy, as will be shown in my upcoming post, even moderately common writes will result in chewing up a whole lot of wasted CPU time due to (3).  If there’s interest, I will look into implementing a variant of this type that uses events for waiting.  Clearly this would sacrifice (2).

Some design decisions have been made in the name of keeping this thing lightweight:

1. No thread affinity will be used.
2. And therefore no recursive acquires will be allowed.

The full code is below, at the bottom of this post.  But let’s review the details one-by-one.

First, all state is packed into a single field, m_state.  We’ll use the 32nd bit to represent whether the write lock is held, and we’ll use the 31st bit to represent whether a writer is attempting to acquire the lock.  As with most reader/writer locks, we will give writers priority over readers because they are supposed to be very infrequent.  In other words, once a writer arrives, no more read lock acquires will be permitted.  The remaining 30 bits will be used to store the reader count.  Some masks make this convenient:

private volatile int m_state;
private const int MASK_WRITER_BIT = unchecked((int)0x80000000);
private const int MASK_WRITER_WAITING_BIT = unchecked((int)0x40000000);
private const int MASK_WRITER_BITS = unchecked((int)(MASK_WRITER_BIT | MASK_WRITER_WAITING_BIT));
private const int MASK_READER_BITS = unchecked((int)~MASK_WRITER_BITS);

Now we can write the four methods: EnterWriteLock, ExitWriteLock, EnterReadLock, ExitReadLock.

Entering the write lock merely entails setting m_state to MASK_WRITER_BIT, provided that we see it available.  If it’s not available, we’ll just go ahead and try to set the MASK_WRITER_WAITING_BIT to prevent subsequent read locks from being acquired until we get in.  We then go ahead and spin until the lock is available using the new type SpinWait in .NET 4.0, checking the m_state field over and over again.  The lock is available if m_state is 0 or MASK_WRITER_WAITING_BIT:

public void EnterWriteLock()
{
    SpinWait sw = new SpinWait();
    do
    {
        // If there are no readers currently, grab the write lock.
        int state = m_state;
        if ((state == 0 || state == MASK_WRITER_WAITING_BIT) &&
             Interlocked.CompareExchange(ref m_state, MASK_WRITER_BIT, state) == state)
           return;

        // Otherwise, if the writer waiting bit is unset, set it.  We don't
        // care if we fail -- we'll have to try again the next time around.
        if ((state & MASK_WRITER_WAITING_BIT) == 0)
            Interlocked.CompareExchange(ref m_state, state | MASK_WRITER_WAITING_BIT, state);

        sw.SpinOnce();
    }
    while (true);
}

Leaving the write lock is actually quite simple.  We just set the m_state field to 0, preserving the MASK_WRITER_WAITING_BIT just in case another writer has arrived since we acquired the lock.  We use an Interlocked.Exchange (XCHG) operation for this, although we technically could have just done an ordinary write, provided doing so wouldn’t cause memory model or availability problems:

public void ExitWriteLock()
{
    // Exiting the write lock is simple: just set the state to 0.  We
    // try to keep the writer waiting bit to prevent readers from getting
    // in -- but don't want to resort to a CAS, so we may lose one.
    Interlocked.Exchange(ref m_state, 0 | (m_state & MASK_WRITER_WAITING_BIT));
}

Entering the read lock is even more straightforward.  The lock is available for readers when m_state & MASK_WRITER_BITS is 0.  In other words, no writer holds the lock and no writer is waiting for the lock.  Once we see the lock in such a state, we merely try to add one to the state value and CAS it in.  In this way, m_state & MASK_READER_BITS will be equal to the number of concurrent readers in the lock:

public void EnterReadLock()
{
    SpinWait sw = new SpinWait();
    do
    {
        int state = m_state;
        if ((state & MASK_WRITER_BITS) == 0)
        {
            if (Interlocked.CompareExchange(ref m_state, state + 1, state) == state)
                return;
        }

        sw.SpinOnce();
    }
    while (true);
}

Lastly, exiting the read lock is the most complicated operation of all.  It needs to decrement the reader count, while at the same time preserving the MASK_WRITER_WAITING_BIT:

public void ExitReadLock()
{
    SpinWait sw = new SpinWait();
    do
    {
        // Validate we hold a read lock.
        int state = m_state;
        if ((state & MASK_READER_BITS) == 0)
            throw new Exception("Cannot exit read lock when there are no readers");

        // Try to exit the read lock, preserving the writer waiting bit (if any).
        if (Interlocked.CompareExchange(
                ref m_state, ((state & MASK_READER_BITS) - 1) | (state & MASK_WRITER_WAITING_BIT), state) == state)
            return;

        sw.SpinOnce();
    }
    while (true);
}

And that’s it.

Here are some single-threaded performance numbers, comparing the relative costs of several locks out there.  These are taken from a large number of acquire/release pairs, i.e., ‘for (int i = 0; i < N; i++) { lock.Enter(); lock.Exit(); }’, for a very large value of N:

Monitor                         0004487479
RWL read lock (legacy)          0023042785      5.13491x
RWL write lock (legacy)         0023118085      5.15169x
SlimRWL read lock (3.5)         0009423579      2.099976x
SlimRWL write lock (3.5)        0008680855      1.934465x
Vance read lock                 0004923609      1.097193x
Vance write lock                0004802136      1.070123x
SpinRWL read lock               0004298525      0.9579604x
SpinRWL write lock              0003819024      0.8510431x

The Nx ratios compare the lock in question to Monitor as our baseline.  Smaller is better.  As you can see, we seem to be on pretty solid ground to start with.  But clearly the most interesting part of this whole thing is the scaling numbers--in particular whether read-mode helps with throughput--both for the existing reader/writer locks and our new one.  The results may surprise you.  That’s coming in the next post...

(Here is the full listing.)

using System;

// We use plenty of interlocked operations on volatile fields below.  Safe.
#pragma warning disable 0420

namespace System.Threading
{
    /// <summary>
    /// A very lightweight reader/writer lock.  It uses a single word of memory, and
    /// only spins when contention arises (no events are necessary).
    /// </summary>
    public struct ReaderWriterSpinLock
    {
        private volatile int m_state;
        private const int MASK_WRITER_BIT = unchecked((int)0x80000000);
        private const int MASK_WRITER_WAITING_BIT = unchecked((int)0x40000000);
        private const int MASK_WRITER_BITS = unchecked((int)(MASK_WRITER_BIT | MASK_WRITER_WAITING_BIT));
        private const int MASK_READER_BITS = unchecked((int)~MASK_WRITER_BITS);

        public void EnterWriteLock()
        {
            SpinWait sw = new SpinWait();
            do
            {
                // If there are no readers currently, grab the write lock.
                int state = m_state;
                if ((state == 0 || state == MASK_WRITER_WAITING_BIT) &&
                    Interlocked.CompareExchange(ref m_state, MASK_WRITER_BIT, state) == state)
                    return;

                // Otherwise, if the writer waiting bit is unset, set it.  We don't
                // care if we fail -- we'll have to try again the next time around.
                if ((state & MASK_WRITER_WAITING_BIT) == 0)
                    Interlocked.CompareExchange(ref m_state, state | MASK_WRITER_WAITING_BIT, state);

                sw.SpinOnce();
            }
            while (true);
        }

        public void ExitWriteLock()
        {
            // Exiting the write lock is simple: just set the state to 0.  We
            // try to keep the writer waiting bit to prevent readers from getting
            // in -- but don't want to resort to a CAS, so we may lose one.
            Interlocked.Exchange(ref m_state, 0 | (m_state & MASK_WRITER_WAITING_BIT));
        }

        public void EnterReadLock()
        {
            SpinWait sw = new SpinWait();
            do
            {
                int state = m_state;
                if ((state & MASK_WRITER_BITS) == 0)
                {
                    if (Interlocked.CompareExchange(ref m_state, state + 1, state) == state)
                        return;
                }

                sw.SpinOnce();
            }
            while (true);
        }

        public void ExitReadLock()
        {
            SpinWait sw = new SpinWait();
            do
            {
                // Validate we hold a read lock.
                int state = m_state;
                if ((state & MASK_READER_BITS) == 0)
                    throw new Exception("Cannot exit read lock when there are no readers");

                // Try to exit the read lock, preserving the writer waiting bit (if any).
                if (Interlocked.CompareExchange(
                    ref m_state, ((state & MASK_READER_BITS) - 1) | (state & MASK_WRITER_WAITING_BIT), state) == state)
                    return;

                sw.SpinOnce();
            }
            while (true);
        }
    }
}

I just uploaded a free sample chapter for my Concurrent Programming on Windows book:

2
Synchronization and Time

STATE IS AN important part of any computer system. This point seems so obvious that it sounds silly to say it explicitly. But state within even a single computer program is seldom a simple thing, and, in fact, is often scattered throughout the program, involving complex interrelationships and different components responsible for managing state transitions, persistence, and so on. Some of this state may reside inside a process’s memory—whether that means memory allocated dynamically in the heap (e.g., objects) or on thread stacks—as well as files on-disk, data stored remotely in database systems, spread across one or more remote systems accessed over a network, and so on. The relationships between related parts may be protected by transactions, handcrafted semitransactional systems, or nothing at all.

The broad problems associated with state management, such as keeping all sources of state in-synch, and architecting consistency and recoverability plans all grow in complexity as the system itself grows and are all traditionally very tricky problems. If one part of the system fails, either state must have been protected so as to avoid corruption entirely (which is generally not possible) or some means of recovering from a known safe point must be put into place.

While state management is primarily outside of the scope of this book, state “in-the-small” is fundamental to building concurrent programs. Most Windows systems are built with a strong dependency on shared memory due to the way in which many threads inside a process share access to the same virtual memory address space. The introduction of concurrent access to such state introduces some tough challenges. With concurrency, many parts of the program may simultaneously try to read or write to the same shared memory locations, which, if left uncontrolled, will quickly wreak havoc. This is due to a fundamental concurrency problem called a data race or often just race condition. Because such things manifest only during certain interactions between concurrent parts of the system, it’s all too easy to be given a false sense of security—that the possibility of havoc does not exist.

In this chapter, we’ll take a look at state and synchronization at a fairly high level. We’ll review the three general approaches to managing state in a concurrent system:

1. Isolation, ensuring each concurrent part of the system has its own copy of state.
2. Immutability, meaning that shared state is read-only and never modified, and
3. Synchronization, which ensures multiple concurrent parts that wish to access the same shared state simultaneously cooperate to do so in a safe way.

We won’t explore the real mechanisms offered by Windows and the .NET Framework yet. The aim is to understand the fundamental principles first, leaving many important details for subsequent chapters, though pseudo-code will be used often for illustration.

We also will look at the relationship between state, control flow, and the impact on coordination among concurrent threads in this chapter. This brings about a different kind of synchronization that helps to coordinate state dependencies between threads. This usually requires some form of waiting and notification. We use the term control synchronization to differentiate this from the kind of synchronization described above, which we will term data synchronization.

Read more here...

Related, I was recently interviewed by DZone about the book.  You can read my responses here.  Enjoy.

I received some feedback on my previous post, Some performance implications of CAS operations, indicating that a few clarifications are in order.  If I had to summarize the intended conclusion, it’d go something like this:

Sharing is evil, fundamentally limits scalability, and is best avoided.

I have to admit that the post was meant to focus more on concrete data, since I expected the meta-point about sharing to be implied.  I figured folks would pick up on the link: (i) Sharing memory requires concurrency control, (ii) Concurrency control requires CAS, (iii) CAS is expensive, therefore (iv) Sharing memory is expensive.  Many people simply don’t understand how crippling CAS can be when placed in a hot path, and I wanted to point out some (albeit extreme) examples of this point.

I did have a motivation for the post.  A lot of people point at lock-free techniques, software transactional memory, reader/writer locks, etc. as ways to improve scalability.  Sadly this seldom pans out.  Each involves CASs of some sort, and, assuming the lock-based equivalent is written properly (that is, to hold locks for very short periods of time), the alternative can in fact often fare worse.  I call this game “count the CASs.”  It’s the roundtrips back to shared memory, failed optmistic attempts, cache invalidations, and line ping ponging that kills you.

Some might accuse me of unfairly targeting CAS.  That’s hogwash.  I’ve been in the trenches for years writing and optimizing systems-level parallel code on Windows.  A parallel for loop can go from scaling perfectly to not scaling at all if you choose the wrong granularity for the loop counter increments.  And vice versa.  Why?  Because the frequency of CASs will bring the memory system to its knees.  You simply must consider these kinds of things when developing your data structures and algorithms; easing pressure on the cache hierarchy is the only way to scale beyond a handful of processors.

The sad truth is that only radical changes to the way we write software will allow fine-grained parallelism to scale to the numbers we expect in the 5 year time horizon.  Hiding more and more conveniently inserted CAS operations auto-magically for folks is not doing them any good.  Mostly functional combined with concurrency-safe mutation on guaranteed-isolated object graphs is, in my opinion, the only path forward.

Along with type systems, I'm casually interested in formal specifications and verification of software.  During lunch today, I watched an internal Microsoft Research talk given by Leslie Lamport.  The topic was TLA+ -- his formal verification system -- during which he blurted out a couple amusing quotes:

"Writing is nature's way of letting you know how sloppy your thinking is."
       --- Guindon (cartoon)

"Math is nature's way of letting you know how sloppy your writing is."
       --- Leslie Lamport (riffing on Guindon)

And related:

"Formal math is nature's way of letting you know how sloppy your math is."
       --- Leslie Lamport

They made me chuckle out loud, so I figured I'd share them.  Unfortunately the talk isn't available outside the company (as far as I can tell), but Lamport has written a book, Specifying Systems, available online, in addition to dozens of interesting papers, on the topic.

CAS operations kill scalability.

(“CAS” means compare-and-swap.  This is the term most commonly used in academic literature, but it is commonly referred to under many guises.  Windows has historically called it an “interlocked” operation and offers a bunch of such-named Win32 APIs; .NET does the same.  This set entails X86 instructions like XCHG, CMPXCHG, and certain instructions prefixed with LOCK, such as INC, ADD, and so on.)

My opening statement is a bit extreme, but it’s true enough.  There are several reasons:

0. CAS relies on support in the hardware to ensure atomicity.  Namely, most Intel and AMD architectures use a MOSEI cache coherency protocol to manage cache lines.  In such an architecture, CAS operations on uncontended lines that are owned exclusively (E) within a processor’s cache are relatively cheap.  But any contention – false or otherwise – leads to invalidations and bus traffic.  The more invalidations, the more saturated the bus, and the greater the latency for CAS completion.  Cache contention is a scalability killer for non-CAS memory operations too, but the need to acquire a line exclusively makes matters doubly worse when CAS is involved.

1. CAS costs more than ordinary memory operations, in CPU cycles.  This is due to the additional burden on the cache hierarchy, and also because of requirements around flushing write buffers, restrictions on speculation across the fences, and impact to a compiler’s ability to optimize around the CAS.

2. CAS is often used in optimistically concurrent operations.  That means a failed CAS will lead to a retry of some sort – typically with some kind of backoff – which is purely wasted work that isn’t present when there isn’t any contention.  And 0 and 1 both increase the risk of contention.

The most common occurrence of a CAS is upon lock entrance and exit.  Although a lock can be built with a single CAS operation, CLR monitors use two (one for Enter and another for Exit).  Lock-free algorithms often use CAS in place of locks, but due to memory reordering such algorithms often need explicit fences that are typically encoded as CAS instructions.  Although locks are evil, most good developers know to keep lock hold times small.  As a result, one of the nastiest impediments to performance and scaling has nothing to do with locks at all; it has to do with the number, frequency, and locality of CAS operations.

As a simple illustration, imagine we’d like to increment a counter 100,000,000 times.  There are a few ways we could do this.  If we’re just running on a single CPU, we can use ordinary memory operations:

Variant #0:
static volatile int s_counter = 0;
for (int i = 0; i < N; i++) s_counter++;

This clearly isn’t threadsafe, but provides a good baseline for the cost of incrementing a counter.  The first way we might make it threadsafe is by using a LOCK INC:

Variant #1:
static volatile int s_counter = 0;
for (int i = 0; i < N; i++) Interlocked.Increment(ref s_counter);

This is now threadsafe.  An alternative way of doing this – commonly needed if we must perform some kind of validation (like overflow prevention) – is to use a CMPXCHG:

Variant #2:
static volatile int s_counter = 0;
for (int i = 0; i < N; i++) {
    int tmp;
    do {
        tmp = s_counter;
    } while (Interlocked.CompareExchange(ref s_counter, tmp+1, tmp) != tmp);
}

An interesting question to ask now is: How much slower will each variant be when cache contention is introduced?  In other words, run a copy of each code on P separate processors, incrementing the same s_counter variable by N/P, and compare the running times for different values of P, including 1.  You might be surprised by the results.

For example, on one of my dual-processor/dual-core (that’s 4-way) Intel machines, the results are as follows.  I’ve run Variant #0 even though it’s not threadsafe, simply because it shows the effects of cache contention on ordinary memory loads and stores.

#0, P = 1: 1.00X
#1, P = 1: 4.73X
#2, P = 1: 5.38X
#0, P = 2: 2.11X
#1, P = 2: 10.74X
#2, P = 2: 16.70X
#0, P = 4: 3.87X
#1, P = 4: 7.57X
#2, P = 4: 73.35X

All numbers are normalized and compared to the ++ code on a single processor.  In other words, Variant #0 run on 2 processors is 2.11X the cost of Variant #0 run on 1 processor; similarly, Variant #0 run on 4 processors is 3.87X the cost of Variant #0 run on 1 processor.  Variant #1 gets even worse at 4.73X, 10.74X, and 7.57X, respectively.  And Variant #2 explodes in cost as more contention is added, going from 5.38X, to 16.70X, to a whopping 73.35X.  Adding more concurrency actually makes things substantially worse.

(The absolute numbers are not to be trusted, and there are anomalies undoubtedly introduced based on how threads are scheduled; I’ve not affinitized them, so they may end up sharing sockets at will.  A more scientific experiment needs to consider such things.) 

The CMPXCHG example (Variant #2) can be improved by strategic spinning when a CAS fails.  Part of what makes the numbers so bad – particularly the P = 4 case – is the amount of lost time due to livelock and the associated memory system interference.

This is an extreme example.  Few workloads sit in a loop modifying the same location in memory over and over and over again.  Even if they do – as in the case of a parallel for loop in which all threads fight to increment the shared “current index” variable – these accesses are ordinarily broken apart by sizeable delays during which useful work is done.  Augmenting the test to delay accessing the shared location by a certain number of function calls certainly relieves pressure.

For example, here are the numbers if we add a 2-function-call delay in between accesses:

#0, P = 1: 1.00X
#1, P = 1: 2.54X
#2, P = 1: 2.77X
#0, P = 2: 1.47X
#1, P = 2: 5.19X
#2, P = 2: 8.59X
#0, P = 4: 2.78X
#1, P = 4: 3.67X
#2, P = 4: 26.55X

And if we add a 64-function-call delay in between accesses, the micro-cost between the three variants doesn’t matter much.  But the contention behavior sure is different.  And we can even find some cases where the multithreaded variants run faster than the single-threaded counterpart:

#0, P = 1: 1.00X
#1, P = 1: 1.00X
#2, P = 1: 1.00X
#0, P = 2: 0.59X
#1, P = 2: 0.74X
#2, P = 2: 0.85X
#0, P = 4: 0.51X
#1, P = 4: 0.45X
#2, P = 4: 1.23X

This is the first time we have seen a number < 1.00X.  That's a speedup; remember, we are using parallelism after all.

As you might guess, in the region between 2 and 64 function calls the results gradually get better and better; and beyond 64, they get substantially better.  In fact, when we insert 128 function calls in between, we get very close to perfect, linear scaling for all 3 variants:

#0, P = 1: 1.00X
#1, P = 1: 1.00X
#2, P = 1: 1.00X
#0, P = 2: 0.50X
#1, P = 2: 0.52X
#2, P = 2: 0.52X
#0, P = 4: 0.30X
#1, P = 4: 0.29X
#2, P = 4: 0.27X

(As a reminder, 0.50X is a perfect speedup on a 2-CPU machine, and 0.25X is a perfect speedup on a 4-CPU machine.)

The moral of this story is that nothing is free, and CAS is certainly no exception.  You should be extremely stingy with adding them to your code, and conscious of the frequency at which threads will perform them.  The same is generally true of all memory access patterns when parallelism is in play, but particularly for expensive operations like CAS.

And even if you’re not using CAS’s directly in your code, you may be using them via some system service.  Parallel Extensions uses them in many ways.  For instance, when you’re doing a Parallel.For loop, we internally share a counter that is accessed by multiple threads.  So even if your algorithm is theoretically embarrassingly parallel, the internally counter management could get in your way.  We try to be intelligent by chunking up indices, but we aren’t perfect: if you have very small loop bodies the overhead of CAS could begin to impact scalability.  You can work around this by making loop bodies more chunky; one example of how is by doing your own partitioning on top of our library (like executing multiple loop iterations inside the body passed to Parallel.For).  Even things like allocating memory with the CLR’s workstation GC requires the occasional roundtrip to reserve a thread-local allocation context by issuing a CAS operation against a shared memory location.