Points on Safepoints

I've been coming across safepoints for over a year without really knowing what they are (you can see safepoints mentioned in the native code JIT produced in an old post). Despite a lot of Googling, I've not been able to find much.

When Gil Tene, CTO and co-founder of Azul Systems, was in town over summer, he gave a lecture on garbage collection and mentioned safepoints. The notes I took were:

• All commercial GCs use safepoints.
• The GC reigns in all threads at safepoints. This is when it has exact knowledge of things touched by the threads.
• They can also be used for non-GC activity like optimization.
• A thread at a safepoint is not necessarily idle but it often is.
• Safepoint opportunities should be frequent. 
• All threads need to reach a global safepoint typically every dozen or so instructions (for example, at the end of loops).

Note that safepoints can stop all threads.

These rough notes jotted down are unsatisfactory so I went looking at the code for safepoints in OpenJDK. I'm far from the best C/C++ programmer in the world (caveat emptor), but these are the interesting things I discovered:

From safepoint.cpp:

  // Begin the process of bringing the system to a safepoint.
  // Java threads can be in several different states and are
  // stopped by different mechanisms:
  //
  //  1. Running interpreted
  //     The interpeter dispatch table is changed to force it to
  //     check for a safepoint condition between bytecodes.
  //  2. Running in native code
  //     When returning from the native code, a Java thread must check
  //     the safepoint _state to see if we must block.  If the
  //     VM thread sees a Java thread in native, it does
  //     not wait for this thread to block.  The order of the memory
  //     writes and reads of both the safepoint state and the Java
  //     threads state is critical.  In order to guarantee that the
  //     memory writes are serialized with respect to each other,
  //     the VM thread issues a memory barrier instruction
  //     (on MP systems).  In order to avoid the overhead of issuing
  //     a memory barrier for each Java thread making native calls, each Java
  //     thread performs a write to a single memory page after changing
  //     the thread state.  The VM thread performs a sequence of
  //     mprotect OS calls which forces all previous writes from all
  //     Java threads to be serialized.  This is done in the
  //     os::serialize_thread_states() call.  This has proven to be
  //     much more efficient than executing a membar instruction
  //     on every call to native code.
  //  3. Running compiled Code
  //     Compiled code reads a global (Safepoint Polling) page that
  //     is set to fault if we are trying to get to a safepoint.
  //  4. Blocked
  //     A thread which is blocked will not be allowed to return from the
  //     block condition until the safepoint operation is complete.
  //  5. In VM or Transitioning between states
  //     If a Java thread is currently running in the VM or transitioning
  //     between states, the safepointing code will wait for the thread to
  //     block itself when it attempts transitions to a new state.
  //

This point number 2 is particularly clever. The GC thread can protect some memory to which all threads in the process can write (using the mprotect system call) so they no longer can. Upon accessing this temporarily forbidden memory, a signal handler kicks in (see this previous post). Example code for this can be found on the man pages where it says:

"If the calling process tries to access memory in a manner that violates the protection, then the kernel generates a SIGSEGV signal for the process." 

From what I can tell, this corresponds to the signal handling code in hotspot/src/os_cpu/linux_x86/vm/os_linux_x86.cpp:


    // Check to see if we caught the safepoint code in the
    // process of write protecting the memory serialization page.
    // It write enables the page immediately after protecting it
    // so we can just return to retry the write.
    if ((sig == SIGSEGV) &&
        os::is_memory_serialize_page(thread, (address) info->si_addr)) {
      // Block current thread until the memory serialize page permission restored.
      os::block_on_serialize_page_trap();
      return true;
    }

More safepoint notes to come.

SOA Far, SOA Good

My quibble with SOA patterns is that they tend to be common sense and hardly worth documenting.

A colleague was talking about the Federation Pattern and when I asked what it was, it turns out that it's something we've been happily doing for years.

Thomas Erl describes it in SOA Principles of Service Design as:

"A federated IT environment is one where resources and application are united while maintaing their individual autonomy and self-governance. SOA aims to increase a federated perspective of an enterprise to whatever extent it is applied. It accomplishes this through the widespread deployment of standardized and composable services each of which encapsulates a segment of the enterprise and expresses it in a consistent manner.

"In support of increasing federation, standardization becomes part of the up-front attention each service receives at design time. Ultimately this leads to an environment where enterprise-wide solution logic becomes naturally harmonized, regardless of the nature of its underlying implementation."

Erl hints that there are greater costs in the attention each federated service demands at design time. Most of the development time in an over-federated suite of applications is the tedious work of writing yet more mapping code and config that allows the interoperability.

What this threshold must be assessed on a case-by-case basis. As ever with architecture, your mileage my vary.

System Properties and the KeyManager

Setting fields using System.setProperty(...) is nasty as it's a global variable by any other name. But sometimes you have to tolerate it when dealing with other people's libraries.

One such library is Java's security classes where you can define your key store file by setting system property javax.net.ssl.keyStore. But be warned: you can only use it once. After that, it is set for the lifetime of the JVM.

If you take a peek into DefaultSSLContextImpl.getDefaultKeyManager(...), you'll see the static defaultKeyManagers being set (from a KeyManagerFactory initiated with a KeyStore that was loaded from the file in the system property). Subsequent calls to this method get the static reference.

This monkeyed around with our test suite no end as it was not immediately obviously where this was being set.

As it happened, an earlier test set it incorrectly but didn't exercise that path of the code. Later in the suite, we saw NoSuchAlgorithmException being thrown by an unrelated test. This is a misleading error.

"NoSuchAlgorithmExceptions are often cause by other underlying exceptions (file not found, wrong password, wrong keystore type, ...)." - from Bruno on StackOverflow.

And so it proved to be. The path the key store value was pointing to didn't exist.

[Miscellaneous: pictures of the SSL Handshake and the Wikipedia description are worth seeing.]

Server Keys

The KeyStore for a server can have many entries. The JavaDocs say:

Each entry in a keystore is identified by an "alias" string. In the case of private keys and their associated certificate chains, these strings distinguish among the different ways in which the entity may authenticate itself. For example, the entity may authenticate itself using different certificate authorities, or using different public key algorithms.

But which entry is passed back to the client? By stepping through the OpenJDK code, I came to this point:



Thread [main] (Suspended)
SunX509KeyManagerImpl.chooseServerAlias(String, Principal[], Socket) line: 271
ServerHandshaker.setupPrivateKeyAndChain(String) line: 1011
ServerHandshaker.trySetCipherSuite(CipherSuite) line: 879
SSLServerSocketImpl.checkEnabledSuites() line: 313
SSLServerSocketImpl.accept() line: 272
.
.

Where I saw this:


        if ((aliases != null) && (aliases.length > 0)) {
            return aliases[0];
        }


The first alias is given to the client no matter what! And thus the private key on the ServerHandshaker object is set.

So, beware if you have other keys in your KeyStore as you might not be giving your client the right one.

Callooh callay! I have OpenJDK

I'm using OpenJDK because I want access to the JCE code to see why my system is behaving the way it is. Since the official JDK Reference Implementation for Java 7 is entirely OpenJDK [1], this seems reasonable.

Remembering I have OpenJDK on my Fedora VM (albeit 1.6), it was just a matter of downloading the source code like any other package. After doing so, I get this:


[phenry@localhost Code]$ rpm -qa | grep openjdk
java-1.6.0-openjdk-devel-1.6.0.0-65.1.10.8.fc15.i686
java-1.6.0-openjdk-1.6.0.0-65.1.10.8.fc15.i686
java-1.6.0-openjdk-src-1.6.0.0-65.1.10.8.fc15.i686

Hoorah! There it is. But where is the actual source file?


[phenry@localhost Code]$ rpm -ql java-1.6.0-openjdk-src-1.6.0.0-65.1.10.8.fc15.i686
/usr/lib/jvm/java-1.6.0-openjdk-1.6.0.0/src.zip
/usr/share/doc/java-1.6.0-openjdk-src-1.6.0.0
/usr/share/doc/java-1.6.0-openjdk-src-1.6.0.0/README.src


(Posting this because I keep forgetting all that RPM goodness [2]).

[1] Java Platform, Standard Edition 7 Reference Implementation
[2] A good crib sheet on RPM Commands.

When buffering does not improve performance

We wrote a simple server earlier this month and were surprised how inefficient it was when under heavy load.

Following the Agile mantra of the simplistic thing that works, we used standard Java classes such as java.io.BufferedReader. To our surprise, this proved a very poor performer when compared to reading the raw InputStream. To illustrate, let's create an interface thus:

package com.henryp.io.stream;

import java.io.IOException;

public interface StreamReader {

    public void readStream() throws IOException;

}

and have two implementing classes. The first is a simple implementation of BufferedReader:


package com.henryp.io.stream;

import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStream;
import java.io.InputStreamReader;
import java.io.Reader;

public class BufferedStreamReader implements StreamReader {

    private final BufferedReader bufferedReader;

    BufferedStreamReader(InputStream inputStream) {
        Reader reader = new InputStreamReader(inputStream);
        bufferedReader = new BufferedReader(reader);
    }

    public void readStream() throws IOException {
        String line = bufferedReader.readLine();
        while (line != null) {
            line = bufferedReader.readLine();
        }
    }

}


The second is a little more sophisticated. It doesn't allow the creation of Strings but instead uses byte arrays. What's more, it caches these byte arrays so it can reuse them:


package com.henryp.io.stream;

import java.io.IOException;
import java.io.InputStream;

public class FixedBufferStreamReader implements StreamReader {
    
    private static final int BUFFER_SIZE = 8192;
    
    private final ThreadLocal buffers = new ThreadLocal();

    private final InputStream inputStream;

    FixedBufferStreamReader(InputStream inputStream) {
        super();
        this.inputStream = inputStream;
    }

    @Override
    public void readStream() throws IOException {
        byte[] buffer = getBufferForThread();
        int read = inputStream.read(buffer);
        while (read != -1) {
            read = inputStream.read(buffer);
        }
    }

    private byte[] getBufferForThread() {
        byte[] buffer = buffers.get();
        if (buffer == null) {
            buffer = new byte[BUFFER_SIZE];
            buffers.set(buffer);
        }
        return buffer;
    }

}


Note, neither implementation actually parses the data. Obviously, using raw byte arrays will be harder but if you want better performance, you'll have to take that hit.

The test harness looks like this:


package com.henryp.io.stream;

import java.io.ByteArrayInputStream;
import java.io.IOException;
import java.io.InputStream;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.atomic.AtomicInteger;

public class StreamReaderStressTest {
    
    private static final int CONTENT_SIZE = 10000;

    private final int numThreads = 100;
    
    private final int numIterations = 200;
    
    private final AtomicInteger readerIndex = new AtomicInteger(-1);

    private final StreamReader[] streamReaders;
    
    private final CountDownLatch countDownLatch = new CountDownLatch(numThreads);

    public static void main(String[] args) throws InterruptedException {
        long initialMemory = memory();
        System.out.println("Initializing...");
        StreamReaderStressTest app = new StreamReaderStressTest();
        long memoryBeforeRun = memory();
        System.out.println("Sleeping...");
        Thread.sleep(4000L);
        app.runTest();
        long memoryAfterRun = memory();
        System.out.println("Memory consumed during setup: " + (initialMemory - memoryBeforeRun) + " bytes"
                + "\nChange in memory during test: " + (memoryBeforeRun - memoryAfterRun) + " bytes");
    }
    
    private static long memory() {
        System.gc();
        long freeMemory = Runtime.getRuntime().freeMemory();
        System.out.println("Free memory: " + (freeMemory / (1024L*1024L)) + " MiB");
        return freeMemory;
    }

    StreamReaderStressTest() {
        InputStream[] inputStreams = initialize();
        streamReaders = new StreamReader[inputStreams.length];
        for (int i = 0 ; i < inputStreams.length ; i++) {
            streamReaders[i] = 
                //new FixedBufferStreamReader(inputStreams[i]); // 13ms
                new BufferedStreamReader(inputStreams[i]); // 367ms
        }
    }

    private void runTest() {
        System.out.println("About to start test...");
        StreamReadingThread[]   readers = new StreamReadingThread[numThreads];
        Thread[]                threads = new Thread[numThreads];
        for (int i = 0 ; i < numThreads ; i++) {
            readers[i] = new StreamReadingThread();
            threads[i] = new Thread(readers[i]);
            threads[i].start();
        }
        waitForThreadsToEnd(threads);
        outputStats(readers);
    }
    
    private void outputStats(StreamReadingThread[] readers) {
        long totalTime = 0;
        for (int i = 0 ; i < readers.length ; i++) {
            totalTime += readers[i].getDuration();
        }
        System.out.println("\nAverage time per thred " + (totalTime / (numThreads)) + "ms\n");
    }

    private void waitForThreadsToEnd(Thread[] threads) {
        for (int i = 0 ; i < threads.length ; i++) {
            try {
                threads[i].join();
            } catch (InterruptedException e) {
                e.printStackTrace();
            }
        }
    }

    class StreamReadingThread implements Runnable {
        
        private volatile long duration = 0;

        @Override
        public void run() {
            getReady();
            long start = System.currentTimeMillis();
            try {
                for (int i = 0 ; i < numIterations ; i++) {
                    streamReaders[readerIndex.incrementAndGet()].readStream();
                }
                duration = System.currentTimeMillis() - start;
            } catch (IOException x) {
                x.printStackTrace();
            }
        }

        private void getReady() {
            countDownLatch.countDown();
            try {
                countDownLatch.await();
            } catch (InterruptedException e) {
                e.printStackTrace();
            }
        }
        
        long getDuration() {
            return duration;
        }
    }

    private InputStream[] initialize() {
        int numStreams = numThreads * numIterations;
        InputStream[] inputStreams = new InputStream[numStreams];
        String stringForStream = createArbitraryString(CONTENT_SIZE);
        for (int i = 0 ; i < numStreams ; i++) {
            inputStreams[i] = new ByteArrayInputStream(stringForStream.getBytes());
        }
        return inputStreams;
    }

    private String createArbitraryString(int length) {
        StringBuffer szb = new StringBuffer();
        for (int i = 0 ; i < length ; i++) {
            szb.append("A");
        }
        return szb.toString();
    }

}


Note:

• All initialization is done up-front.
• A 4 second pause is implemented before the test so there is no interference from biased locking being turned off at start up (see here for a full description).
• Run the JVM with the arguments -Xmx1024m -Xmn1024m so the memory is fully initialized to begin with.

Over 5 runs each, the use of BufferedReader yielded a mean time of 1813ms whereas using the raw InputStream took a mean time of 380ms.

A clue for why this is the case is in the source code for BufferedReader. There is a liberal sprinkling of System.arraycopy().

Furthermore, the amount of memory used by the BufferedReaders was about 4 times greater than the InputStreams (roughly 700MB versus 185MB). This is not surprising when you look at the code for BufferedReader. It creates byte arrays on instantiation.

Greater efficiency using the raw InputStream was achieved by using a smaller size for the byte array. Reducing it to 1024 bytes gave a mean time of 217ms (although with a greater standard deviation).

Looking at the source code for GlassFish's Grizzly project, you can see ByteBuffers being cached to read the data from a SocketChannel (see the method org.glassfish.grizzly.nio.transport.TCPNIOUtils.readSimpleByteBuffer(...) for more information. The test org.glassfish.grizzly.http.server.HttpInputStreamsTest.testCharacter010() executes this code).

[Aside: to make Grizzly into an Eclipse project, you need to run:


mvn org.apache.maven.plugins:maven-eclipse-plugin:2.6:eclipse

See here for why.]

Garbage Collection Rubbish

More memory leaks recently and again from various byte-code manipulators. A memory dump showed huge swathes of memory taken by classes whose names indicate that they were dynamically generated by CGLIB.

The problem manifested itself during integration test runs where we constantly start and stop Jetty servers. Since we start and stop Jetty programatically (see a previous post) and the way we used Spring seemed to dynamically generate many classes, these classes ultimately consumed all the memory because they were then be stored in static references - that is, they are stored with the classes. And these classes were not garbage collected since Spring's classloader was the same as that running the test.

"Types loaded through the bootstrap class loader will always be reachable and will never be unloaded. Only types loaded through user-defined class loaders can become unreachable and can be unloaded by the virtual machine." - Inside the Java 2 Virtual Machine, Bill Venners, p266.

If we'd deployed a WAR file presumably all generated classes would have been garbage collected when the web app was unloaded in the test's tear down method.

To demonstrate our problem, let's:

1. write our own class loader
2. use it to load a class
3. instantiate an instance of that class and let it consume lots of memory by storing objects in one of its static references.
4. allow the classloader to be garbage collected
5. see how the memory changed during all these steps

OK, so here's the class loader:



package com.henryp.classloader;

import java.io.IOException;
import java.io.InputStream;
import java.net.URL;

public class MyClassLoader extends ClassLoader {

    private final String baseDir;

    MyClassLoader(String baseDir) {
        super();
        this.baseDir = baseDir;
    }

    @Override
    protected Class findClass(String filename) throws ClassNotFoundException {
        filename                        = filename.replace(".", "/") + ".class";
        InputStream resourceAsStream    = null;

        try {
            URL     url         = new URL(baseDir + filename);
            System.out.println("URL = " + url);
            resourceAsStream    = url.openStream();
            System.out.println("Resource null? " + (resourceAsStream == null));
            byte[]  buffer      = new byte[4096];
            byte[]  bytes       = new byte[4096];
            int     read        = resourceAsStream.read(buffer);
            int     total       = 0;

            while (read != -1) { // this could be a lot more efficient but it does the job
                System.out.println("Read " + read + " bytes");
                byte[] dest = new byte[total + read];
                System.arraycopy(bytes,     0, dest, 0, total);
                System.arraycopy(buffer,    0, dest, total, read);
                bytes = dest;
                total += read;
                read = resourceAsStream.read(buffer);
            }
            return defineClass(null, bytes, 0, total);
        } catch (IOException e) {
            e.printStackTrace();
            throw new ClassNotFoundException(filename);
        } finally {
            try {
                resourceAsStream.close();
            } catch (Throwable e) {
                e.printStackTrace();
            }
        }
    }
}


Not awfully pretty but it doesn't need to be terribly efficient.

The main class that uses our class loader looks like this (notice that we measure the memory before and after our class loader is eligible for garbage collection):

package com.henryp.classloader;


public class Main {
    
    public static void main(String[] args) {
        Main app = new Main();
        app.doClassloading();
    }

    private void doClassloading() {
        MyClassLoader   classLoader = new MyClassLoader("file:/Users/phenry/Documents/workspace/TestOther/bin/");
        consumeMemory(classLoader);
        long            beforeNull  = gcThenLogMemory();
        System.out.println("Making classloader eligible for garbage collection");
        classLoader                 = null;
        long            afterNull   = gcThenLogMemory();
        System.out.println("Memory usage change: " + (afterNull - beforeNull) + " bytes liberated");
    }

    private void consumeMemory(MyClassLoader classLoader) {
        try {
            // filename from a project that depends on this but this project does not depend on that
            Class clazz = classLoader.findClass("com.henryp.classloader.OtherStaticReferenceBloater");
            consumeMemory(clazz);
        } catch (Exception e) {
            e.printStackTrace();
            System.exit(-1);
        }
    }

    private static long gcThenLogMemory() {
        System.gc();
        long freeMemory = Runtime.getRuntime().freeMemory();
        System.out.println("After garbage collection, free memory = " + freeMemory);
        return freeMemory;
    }

    private void consumeMemory(Class clazz) throws Exception {
        LargeMemoryConsumable   memoryConsumer  = instantiateMemoryConsumer(clazz);
        long                    initialMemory   = Runtime.getRuntime().freeMemory();
        doCallOnForeignObject(memoryConsumer);
        System.gc();
        long                    finalMemory     = Runtime.getRuntime().freeMemory();
        System.out.println("Initial memory = " + initialMemory);
        System.out.println("Final   memory = " + finalMemory);
        System.out.println("Total consumed = " + (initialMemory - finalMemory));
    }

    private LargeMemoryConsumable instantiateMemoryConsumer(Class clazz)
            throws InstantiationException, IllegalAccessException {
        System.out.println("Class name: " + clazz.getName());
        gcThenLogMemory();
        LargeMemoryConsumable memoryConsumer = (LargeMemoryConsumable) clazz.newInstance();
        checkDifferentClassLoader(clazz);
        return memoryConsumer;
    }

    private void doCallOnForeignObject(LargeMemoryConsumable memoryConsumer) throws Exception {
        for (int i = 0 ; i < 500000 ; i++) {
            memoryConsumer.consumeMemory();
        }
    }

    private void checkDifferentClassLoader(Class clazz) {
        if (clazz.getClassLoader() == this.getClass().getClassLoader()) {
            System.out.println("Experiment useless if the classloaders are the same");
            System.exit(-1);
        }
    }
}


And the memory consumer we refer to is a simple interface:

package com.henryp.classloader;

public interface LargeMemoryConsumable {
    public void consumeMemory() throws Exception;
}


Now, this is where it gets a little more interesting. In another project, we have an object that just takes up memory. But note that although this new project depends on the classes in our first project, the first project does not reference this new project. It does, however, use a URL that points at the directory that happens to be the directory into which the classes of the new project are compiled.

(I've added a toString method to stop the compiler optimizing away the aString member.)

package com.henryp.classloader;

public class MyOtherFatObject {

    public static Object create() {
        return new MyOtherFatObject("" + System.currentTimeMillis());
    }
    
    private final String aString;

    private MyOtherFatObject(String aString) {
        super();
        this.aString = aString;
    }

    @Override
    public String toString() {
        return "MyOtherFatObject [aString=" + aString + "]";
    }
    
}


In the same project, we have a class that emulates what we saw Spring doing, that is having a static reference to a collection of memory-consuming objects.

package com.henryp.classloader;

import java.util.ArrayList;
import java.util.Collection;

public class OtherStaticReferenceBloater implements LargeMemoryConsumable {
    
    private static final Collection

    public void consumeMemory() throws Exception {
        if (this.getClass().getClassLoader() != MyOtherFatObject.class.getClassLoader()) {
            throw new Exception("Different classloaders");
        }
        objects.add(MyOtherFatObject.create());
    }

    public void addObject(Object object) {
        objects.add(object);
    }

}



Running with these JVM args:



-verbose:gc -Xmx64m -Xmn64m

the output looks like:

URL = file:/Users/phenry/Documents/workspace/TestOther/bin/com/henryp/classloader/OtherStaticReferenceBloater.class
Resource null? false
Read 1300 bytes
Class name: com.henryp.classloader.OtherStaticReferenceBloater
After garbage collection, free memory = 63333768
URL = file:/Users/phenry/Documents/workspace/TestOther/bin/com/henryp/classloader/MyOtherFatObject.class
Resource null? false
Read 896 bytes
Initial memory = 63333768
Final   memory = 15130184
Total consumed = 48203584
After garbage collection, free memory = 15130200
Making classloader eligible for garbage collection
After garbage collection, free memory = 66266816
Memory usage change: 51136616 bytes liberated


(Notice how we explicitly load one class but it implicitly pulls in another using our class loader, the same class loader that loaded it - see the lines beginning with URL = ...).

Noting the usual caveats that System.gc() is a suggestion to the JVM and not a guarantee that the garbage collector should run, making the reference to the class loader null seems to liberate all the objects stored in the static reference of the class it loaded.

What we finally did was make all references we had in our integration tests to Spring's context loaders non-static. As our tests were garbage collected, memory was freed.