Tuesday, June 2, 2009

How to implement “lock” with timeout?

Anybody who did any multithreaded application probably used the “lock” keyword. This is actually a good thing.

“lock” is the most optimized way to lock a resource. First let’s take a look at what lock really do. Here is the simple Account class we will use for this post.

using System.Threading;

namespace Banking
{
    public class Account
    {
        private readonly string _accountNumber;
        private double _balance;

        private Account(string accountNumber, double amount)
        {
            _accountNumber = accountNumber;
            Balance = amount;
        }

        public string AccountNumber
        {
            get { return _accountNumber; }
        }

        public double Balance
        {
            get { return _balance; }
            set { _balance = value; }
        }

        public static Account OpenNew(string accountNumber, double amount)
        {
            return new Account(accountNumber, amount);
        }

        public void Deposit(double amount)
        {
            Interlocked.Exchange(ref _balance, _balance + amount);
        }

        public void Withdraw(double amount)
        {
            Interlocked.Exchange(ref _balance, _balance - amount);
        }

        public static void Transfer(double amount, Account fromAccount, Account toAccount)
        {
            // Bad code here. Potential deadlock.
            lock (fromAccount)
            lock (toAccount)
            {
                fromAccount.Withdraw(amount);
                toAccount.Deposit(amount);
            }
        }

        public override string ToString()
        {
            return string.Format("{0} (Balance = {1})", _accountNumber, _balance);
        }
    }
}

Here are some important things to point out about this class:

  • The constructor is private to limit the creators of this class.
  • The public OpenNew method is the only way to create an instance. This ensure that every Account starts with a name and a balance.
  • Deposit and Withdraw methods are thread safe. They both uses Interlocked class which is has low level methods to modify values.
  • Transfer is not thread safe even though it uses locks. There is a potential deadlock if two thread transfer funds using the same accounts at the same time.

Most of the time the transfer method will work without any problem but the is a slight chance of deadlock. Of course this is a fairly simple method and in fact we can use another private object field to lock on like in this sample.

private static object _syncLock = new object();

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    // Bad code here. Potential deadlock.
    lock (_syncLock)
    {
        fromAccount.Withdraw(amount);
        toAccount.Deposit(amount);
    }
}

That mean you will need to use that _syncLock object all the time to be consistent, even if you need to lock only one of the two accounts. And because Transfer is a static method we need to make the _syncLock object static too. That mean that any other call to Transfer will have to wait until this call finish. That is a huge performance issue.

What we really need is to be able to lock actual Account objects and recover from any deadlock. The best way to do this is to use timeouts on the locking process. The caller can catch timeouts and handle it properly instead of waiting forever. Here is a better Transfer method.

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    bool fromLock = Monitor.TryEnter(fromAccount, 1000);
    bool toLock = Monitor.TryEnter(toAccount, 1000);
    try
    {
        if (fromLock && toLock)
        {
            fromAccount.Withdraw(amount);
            toAccount.Deposit(amount);
        }
    }
    finally
    {
        if (fromLock)
            Monitor.Exit(fromLock);

        if (toLock)
            Monitor.Exit(toLock);
    }
}

This is a lot more code to write. Actually this is not far from what the lock keyword would do. Because if you look at your code with Reflector you will see that lock does translate to Monitor.Enter and Monitor.Exit (your have to look in IL not in C#). So a simple lock statement like this.

lock(obj)
{
    // do somtehing
    Console.WriteLine("Locked");
}

Would be translated to.

Monitor.Enter(obj);
try
{
    // do somtehing
    Console.WriteLine("Locked");
}
finally
{
    Monitor.Exit(obj);
}

You won’t see it in C# (with Reflector) but if you look at IL code you will see exactly the same sequence twice.

So to solve our problem, would it be nice to have something like this?

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    Safe.Lock(new [] {fromAccount, toAccount}, 1000, () =>
    {
        fromAccount.Withdraw(amount);
        toAccount.Deposit(amount);
    });
}

This look pretty much like the well known lock keyword, isn’t it? Now how can we do this? That’s better. The Safe static call has a Lock static method that take care of everything. The fist argument can be a single object or an array of objects. This allow locking multiple object at once. The Lock method use the Monitor.TryEnter to acquire lock on objects. If this can be done before the timeout occur the Action argument is executed.

Of course this code can and should be surrounded with a try-catch block. here is how to do so.

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    var retries = 10;

    while (retries-- > 0)
    {
        try
        {
            Safe.Lock(new[] { fromAccount, toAccount }, 1000, () =>
            {
                fromAccount.Withdraw(amount);
                toAccount.Deposit(amount);
            });
            break;
        }
        catch (TimeoutException e)
        {
            if (retries == 0)
                throw;
            Thread.Sleep(100);
        }
    }
}

In this code snippet it will retry at most 10 time to lock toAccount and fromAccount before giving up. Notice that only the TimeoutException is handled, so any other exception will be thrown immediately and stop the process. The break at the end of the try block let us out as soon as it works.

This is a good timeout pattern and easy to implement in your own code. Give it a try and let me know if it works.

Here is the full code for the Safe class.

using System;
using System.Linq;
using System.Threading;

namespace MultithreadHelper
{
    public class Safe : IDisposable
    {
        private readonly object[] _padlocks;
        private readonly bool[] _securedFlags;

        private Safe(object padlock, int milliSecondTimeout)
        {
            _padlocks = new[] {padlock};
            _securedFlags = new[] {Monitor.TryEnter(padlock, milliSecondTimeout)};
        }

        private Safe(object[] padlocks, int milliSecondTimeout)
        {
            _padlocks = padlocks;
            _securedFlags = new bool[_padlocks.Length];
            for (int i = 0; i < _padlocks.Length; i++)
                _securedFlags[i] = Monitor.TryEnter(padlocks[i], milliSecondTimeout);
        }

        public bool Secured
        {
            get { return _securedFlags.All(s => s); }
        }

        public static void Lock(object[] padlocks, int millisecondTimeout, Action codeToRun)
        {
            using (var bolt = new Safe(padlocks, millisecondTimeout))
                if (bolt.Secured)
                    codeToRun();
                else
                    throw new TimeoutException(string.Format("Safe.Lock wasn't able to acquire a lock in {0}ms",
                                                             millisecondTimeout));
        }

        public static void Lock(object padlock, int millisecondTimeout, Action codeToRun)
        {
            using (var bolt = new Safe(padlock, millisecondTimeout))
                if (bolt.Secured)
                    codeToRun();
                else
                    throw new TimeoutException(string.Format("Safe.Lock wasn't able to acquire a lock in {0}ms",
                                                             millisecondTimeout));
        }

        #region Implementation of IDisposable

        public void Dispose()
        {
            for (int i = 0; i < _securedFlags.Length; i++)
                if (_securedFlags[i])
                {
                    Monitor.Exit(_padlocks[i]);
                    _securedFlags[i] = false;
                }
        }

        #endregion
    }
}

Starting a new user group in Montreal area

Maxime Rouiller and I are starting a new user group in the Montreal area.

It’s called Alt.Net Montreal.

Our goal with this user group is to raise the level of expertise of everyone interested. We already did a Coding Dojo couple of weeks ago and it went well.

We need your input to plan the future. Get involve and together we will build something cool.

All suggestions are welcomed.

Monday, June 1, 2009

Code Camp 2009 Montreal, a real Success

CodeCampLogo175.jpg

Code camp 2009 at Montreal was a real success.

There was about 300 people attending and 24 sessions available.

My session on Threading, UI and Data Model, was more than full. There was people standing in the back of the room. I had a lot present but I think I did it well, I hope.

For those of you who were there (and those who were not) you can find all the material on my web site (here).

Monday, May 18, 2009

Visual Studio 2010 beta 1 is available

Finally Visual Studio 2010 Beta 1 is available. I was waiting for that release for so long. I’m eager to test my Parallel Extensions demo on it.

http://msdn.microsoft.com/en-ca/subscriptions/downloads/default.aspx?pv=18:370

Stay tuned I will tell you how it goes.

Wednesday, May 13, 2009

How to test your multi-threaded code (part 3 of 3)?

In the last post we learned how to find and fix a simple multi-thread problem. Now we will see a more complex scenario and see how CHESS wil find the solution.

To do that we will add a ne method to our Account type.

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    lock (fromAccount)
    {
        lock(toAccount)
        {
            fromAccount.Withdraw(amount);
            toAccount.Deposit(amount);
        }
    }
}

Because we want to be sure that the transfer works we lock both the “from” and the “to” account.

Now we can easily wrtie this test to see that this is working fine in signle threaded scenario.

[TestMethod]
public void TransferTest()
{
    Account a1 = Account.OpenNew(10000);
    Account a2 = Account.OpenNew(10000);

    Account.Transfer(100, a1, a2);
    Account.Transfer(100, a2, a1);

    Assert.AreEqual(10000, a1.Balance);
    Assert.AreEqual(10000, a2.Balance);
}

As we did before we will convert this single-thread method to a multi-thread one.

[TestMethod]
[HostType("Chess")]
public void TransferMultiThreadTest()
{
    Account a1 = Account.OpenNew(10000);
    Account a2 = Account.OpenNew(10000);

    Thread thread = 
new Thread(o => Account.Transfer(100, ((Account[]) o)[0], ((Account[]) o)[1])); thread.Start(new[] {a1, a2}); Account.Transfer(100, a2, a1); thread.Join(); Assert.AreEqual(10000, a1.Balance); Assert.AreEqual(10000, a2.Balance); }

Now if we run this CHESS will detect a deadlock scenario. If you’ve done some SQL queries you know you should always try to lock all your resources always in the same order. But why doesn’t it working here. We have only one method that lock resources they should be locked in the same order every time, and that’s true. The deadlock occurs because in some cases a thread start to lock the fromAccount (or maybe the toAccount too) and get interrupt by another thread trying to do the same. Then both thread are waiting for each other to complete. Databases engine use timeouts to get out of these situation, but the lock keyword doesn’t support timeout, eventough it uses Monitor.Enter which support it. In another post I will show you how to build your own lock implementation that support timeout, but now we need to find a way to make our code thread safe.

We have to go back to our account class and do some changes.

public static object _locObject = new object();

public static void Transfer(double amount, Account fromAccount, Account toAccount)
{
    lock (_locObject)
    {
        fromAccount.Withdraw(amount);
        toAccount.Deposit(amount);
    }
}

We have to create a static lock object we can use to lock on. Because we are doing the lock in one signle operation our test will now run without any problem.

This is only a small overview of what CHESS can do.

Monday, May 11, 2009

How to test your multi-threaded code (part 2 of 3)?

Previously we saw how to build a test to find a multi-thread bug our your code. Now we will look at how to reproduce debug and fix it.

Remember our bank account class:

public class Account
{
   public double Balance { get; set; }

   private Account(double amount)
   {
       Balance = amount;
   }

   public static Account OpenNew(double amount)
   {
       return new Account(amount);
   }

   public void Deposit(double amount)
   {
       double tempAmount = Balance;
       // potential problem
       lock (this)
       {
           Balance = tempAmount + amount;
       }
   }

   public void Withdraw(double amount)
   {
       double tempAmount = Balance;
       // potential problem
       lock (this)
       {
           Balance = tempAmount - amount;
       }
   }
}

And our test:

[TestMethod()]
[HostType("Chess")]
public void BalanceMutiThreadTest()
{
   Account account = Account.OpenNew(10000);

   Thread thread = new Thread(a => ((Account) a).Withdraw(100));
   thread.Start(account);
   account.Deposit(100);
   thread.Join();

   Assert.AreEqual(10000, account.Balance);
}

Now if we look carefully at the test result detail, there is an explanation on how to reproduce this particular schedule. All you have to do is to copy/paste the code provided just before your test method call.

[TestMethod()]
[HostType("Chess")]
[TestProperty("ChessMode", "Repro")]
[TestProperty("ChessBreak", "BeforePreemption")]
#region ChessScheduleString (not human readable)
[TestProperty("ChessScheduleString", @"bpilaiaaaaaaaaaaaeaaonlnahgabmejjgcfcgcpgnmkhlhpekpfeknhoahekbaiiagabdcenijaeabaommbiimnogjcombngjehcdcjklckibmkgffggffnggbgeammonjnlmphnohloplnphnohloplnphlkdljneochphnpppdpfmgggeabgmpgmoeknkmjjocbiakkmibpdphohmbpdpcchomnfpodnhpidfpappfpfhhpponplpkgpmdelpppbgpkplkpoflfmbopdpkgdppbppfpephpocllfpedhppkopjkppjlohpppohpaaldcaoojfhfaaaaaa")]
#endregion
public void BalanceMutiThreadTest()
{
   Account account = Account.OpenNew(10000);

   Thread thread = new Thread(a => ((Account) a).Withdraw(100));
   thread.Start(account);
   account.Deposit(100);
   thread.Join();

   Assert.AreEqual(10000, account.Balance);
}

(your code may be different)

With that code in place if you run your test again (without the debugger) you will see that only one schedule was evaluated and you got the same result as the previous test. From there, as a tester, you job is done. You check in the code and hand it to the development team. If you are a member of a small team, as I usually am, you may be the tester and the developer so you’ll have to debug the code yourself.

Now run this test again but this time with the debugger. The execution should stop just before one of the lock statement. If you step once you will be able to inspect tempAmount and Balance values and see the problem. Between the line where the debugger stop and the previous line another thread changed the Balance value. Now you can see the don’t match.

Of course in this case the solution is easy, we just have to put the tempAmount assignation inside of the lock block but in the next post you will see a case where the solution is not so obvious.

Tuesday, May 5, 2009

How to test your multi-threaded code (part 1 of 3)?

CHESS is the answer. At least this is what we have best right now.

In multi-threaded application, bug are hard to almost impossible to find. For the last years the only true way to detect threading problems was to run load test until the system crash. Once it does, every once in thousands of iterations, the tools to reproduce and debug the problem were inexistent.

The RiSE (Research in Software Engineering) team at Microsoft have been working for a long time on a product called CHESS. When run with CHESS, you unit tests will try every possible combination of thread interleave to find a case where you application crash or worst doesn’t give you the result you expect.

Here is a simple demo to show you the power of CHESS. Let’s start with a banking account management system.

public class Account
{
    public double Balance { get; set; }

    private Account(double amount)
    {
        Balance = amount;
    }

    public static Account OpenNew(double amount)
    {
        return new Account(amount);
    }

    public void Deposit(double amount)
    {
        double tempAmount = Balance;
        // potential problem
        lock (this)
        {
            Balance = tempAmount + amount;
        }
    }

    public void Withdraw(double amount)
    {
        double tempAmount = Balance;
        // potential problem
        lock(this)
        {
            Balance = tempAmount - amount;
        }
    }
}

Of course I voluntarily introduce “potential problems” to show how CHESS will get them. With that class you can write this test:

[TestMethod()]
public void BalanceTest()
{
    Account account = Account.OpenNew(10000);

    account.Withdraw(100);
    account.Deposit(100);

    Assert.AreEqual(10000, account.Balance);
}

This will always run fine as it is single threaded. Now what if we change it a little to make it multi-threaded:

[TestMethod()]
public void BalanceMutiThreadTest()
{
    Account account = Account.OpenNew(10000);

    Thread thread = new Thread(a => ((Account) a).Withdraw(100));
    thread.Start(account);
    account.Deposit(100);
    thread.Join();

    Assert.AreEqual(10000, account.Balance);
}

As you can see here we put the withdraw part in a new thread. But even then we can run this test again and again without any problem. You can put it in a loop if you want and never being able to make it return an invalid balance.

Once you have CHESS installed on your system the only thing you have to do is to add a HostType attribute to your method.

[TestMethod()]
[HostType("Chess")]
public void BalanceMutiThreadTest()
{
    Account account = Account.OpenNew(10000);

    Thread thread = new Thread(a => ((Account) a).Withdraw(100));
    thread.Start(account);
    account.Deposit(100);
    thread.Join();

    Assert.AreEqual(10000, account.Balance);
}

Now when you run this test you should get something like:

Assert.AreEqual failed. Expected:<10000>. Actual:<10100>.

You should also have noticed that it took a little longer to run the test. This is because the CHESS host scans your code to build an execution schedule for every possible thread interleave that CHESS can detect. If you double click on the test result you will see how many schedules that were tried before finding the bug, in my case it is 3.

Next time we will see how to reproduce and debug that code.

Friday, April 24, 2009

Parallel Extension Talk at Vermont .NET User Group

I want to thank Julie Lerman for letting me talk about Parallel Extensions at the Vermont .NET User Group.

The power point slides are available as well as the code samples.

http://www.decarufel.net/parallelextensionsvermont2009

Friday, March 27, 2009

Developers are artists AND craftsman not ONLY craftsman.

Let me take some time to quote and comment a friend of mine:

Forget about the “creation” part of the job. As you are developing your first application, you are not creating something. You are building something. Mind you, it’s not like building a bridge as in engineering or a plane.

from Software development is not an art. It’s a craft. by Maxime Rouiller

Developer are not just craftsman they are also artist. It’s even more true for architect. Architecture is not just about building something by assembling parts and make sure its functional. Most building architect will spent almost as much time on the look of building as on its internal structure. If that wasn’t true, our cities would have been only an ugly pile of grey steel and concrete. But when we look around we can clearly see things that are as beautiful as they are useless like sculpted arches, round towers or multicolour spotlights.

This is also true in sofware design. You just have to compare Vista to the first version of Windows (Windows 3.1 in my case) to see how much more beautiful  it is today than back then. Does that make it more useful? Does that make it stronger? Does all that is really necessary? No, no and no. In fact, some will tell that this is so useless that Microsoft shouldn’t have spent a dime on all those things. But now the trend is to talk about “user experience” and how it is important to feel great when you use a product as they said at MIX09 opening keynote. In some very successful product advertisement you don’t even see the product itself but people happy just because they own that product.

Building an application, even today, is not just putting parts together to make it work. Software development is not mature enough to be considered an exact science. You just have to think about how difficult it is to commit youself when you have to give an estimate on how long it would take you to build something. Why? Baecause every case is unique. In my life (tank god) I rarely had to build exactly the same thing twice. That would be boring for me. Every time I start something new I have to think about what would be the best solutiion to the problem in hand. Even with the same problem, solution may change based on the customer priority and constraints.

I think that develeoper and architect alike are not only craftsman, they are also artist. Let them express themselves, you will be surprised. Where are now on the eve of new era of tools that will open new possibilities of expression, tools like the Expression Suite. The reign of ugly and boring application is over. Over the next year we will witness all kind of new designs and user experiences.

Time will tell us which will survive and which will not.