Homework 4. Java shared memory performance races

Background

You're working for a startup company Ginormous Data Inc. (GDI) that specializes in finding patterns in large amounts of data. For example, a big retailer might give GDI all the web visits and purchases made and all the credit reports they've inspected and records of all the phone calls to them, and GDI will then find patterns in the data that suggest which toys will be hot this Christmas season. The programs that GDI writes are mostly written in Java. They aren't perfect; they're just heuristics, and they're operating on incomplete and sometimes-inaccurate information. They do need to be fast, though, as your clients are trying to find patterns faster than their competition can, and are willing to put up with a few errors even if the results aren't perfect, so long as they get good-enough results quickly.

The problem

GDI regularly uses multithreading to speed up its applications, and many of GDI's programs operate on shared-memory representations of the state of a simulation. These states are updated safely, using Java's synchronized keyword, and this is known to be a bottleneck in the code. Your boss asks you what will happen if you remove the synchronized keyword. You reply, "It'll break the simulations." She responds, "So what? If it's just a small amount of breakage, that might be good enough. Or maybe you can substitute some other synchronization strategy that's less heavyweight, and that'll be good enough." She suggests that you look into this by measuring how often GDI's programs are likely to break if they switch to inadequate-but-faster synchronization methods.

In some sense this assignment is the reverse of what software engineers traditionally do with multithreaded applications. Traditionally, they are worried about race conditions and insert enough synchronization so that the races become impossible. Here, though, you're deliberately trying to add races to the code in order to speed it up, and want to measure whether (and ideally, how badly) things will break if you do.

It should be noted that we are on thin ice here, so thin that some would argue we've gone over the edge. That's OK: we're experimenting! For more about the overall topic, please see: Boehm H-J, Adve SV. You don't know jack about shared variables or memory models. ACM Queue 2011 Dec;9(12):40. doi:10.1145/2076796.2088916.

The Java memory model

Java synchronization is based on the Java memory model (JMM), which defines how an application can safely avoid data races when accessing shared memory. The JMM lets Java implementations optimize accesses by allowing more behaviors than the intuitive semantics where there is a global clock and actions by threads interleave in a schedule that assumes sequential consistency. On modern hardware, these intuitive semantics are often incorrect: for example, intraprocessor cache communication might be faster than memory, which means that a cached read can return a new value on one processor before an uncached read returns an old value on another. To allow this kind of optimization, first, the JMM says that two accesses to the same location conflict if they come from different threads, at least one is a write, and the location is not declared to be volatile; and second, the JMM says that behavior is well-defined to be data-race free (DRF) unless two conflicting accesses occur without synchronization in between.

The details for proving that a program is DRF can be tricky, as is optimizing a Java implementation with data-race freedom in mind. Not only have serious memory-synchronization bugs been found in Java implementations, occasionally bugs have been found in the JMM itself, and sometimes people have even announced bugs only to find out later that they weren't bugs after all. For more details about this, please see: Lochbihler A. Making the Java memory model safe [PDF]. ACM TOPLAS 2013 Dec;35(4):12. doi:10.1145/2518191. You needn't read all this paper, just the first eight pages or so—through the end of §1.1.3.

How to break the JMM's rules

It's easy to write programs that violate the JMM's rules. To model this, you will use a simple prototype that manages a data structure that represents an array of integers. Each integer is in the range [0,maxval] where maxval is at most 127, so the integer can be represented by the Java type byte. A state transition, called a swap, consists of subtracting 1 from one of the positive integers in the array, and adding 1 to an integer that is less than maxval. The sum of all the integers should therefore remain constant; if it varies, that indicates that one or more transitions weren't done correctly. Also, the values in the array should always be in range. The converse is not true: if the sum remains constant and the values remain in range it's still possible that some state transitions were done incorrectly. Still, these tests are reasonable ways to check for errors in the simulation.

For an example of a simulation, see jmm.jar, a JAR file containing the simplified source code of a simulation. It contains the following interfaces and classes:

State
The API for a simulation state. The only way to change the state is to invoke swap(i,j), where i and j are indexes into the array. If the ith entry in the array is positive and the jth entry is less than the maximum value allowed, the swap succeeds, subtracting 1 from the ith entry and adding 1 to the jth entry, returning true. Otherwise the swap fails and does nothing, returning false.
Nullstate
An implementation of State that does nothing. Swapping has no effect. This is used for timing the scaffolding of the simulation.
SynchronizedState
An implementation of State that uses the Synchronized model so that it is safe but slow.
SwapTest
A Runnable class that tests a state implementation by performing a given number of successful swaps on it. It does not count failed swaps.
UnsafeMemory
A test harness, with a main method. Invoke it via a shell command like "java UnsafeMemory Synchronized 8 1000000 6 5 6 3 0 3". Here, Synchronized means to test the SynchronizedState implementation; 8 means to divide the work into 8 threads of roughly equal size; 1000000 means to do a million successful swap transitions total; 6 is maxval, an integer in the range [0,127] as described above; and the remaining five numbers are the initial values for the five entries in the state array. The shell command outputs a string like "Threads average 3318.01 ns/transition", giving the approximate average number of real-time nanoseconds that it took a thread to do a successful swap, including all the overhead. It also outputs an error diagnostic if a reliability test fails.

Assignment

Build and use a JMM-violating performance and reliability testing program, along the lines described below.

  1. Your program should operate under Java Standard Edition 8. There is no need to run on older Java versions.
  2. Your program should compile cleanly, without any warnings.
  3. Please keep your implementation as simple and short as possible, for the benefit of the reader.
  4. Use the SEASnet GNU/Linux servers, with Java version 1.8.0_31 or later, to do your performance and reliability measurements. On SEASnet your PATH should be set to a string starting with "/usr/local/cs/bin:".
  5. Do not use more than 32 threads at a time, to avoid overloading the servers.
  6. Gather and report statistics about your testing platform, so that others can reproduce your results if they have similar hardware. See the output of java -version, and see the files /proc/cpuinfo and /proc/meminfo.
  7. Run the test harness on the Null and Synchronized models, using various values for the number of threads, number of swap transitions, size of the state array, and sum of values in the state array, and characterize the performance of the two models. Both models should have 100% reliability, in the sense that they should pass all the tests (even though the Null model does not work); check this.
  8. Implement a new model Unsynchronized, which is implemented just like Synchronized except that it does not use the keyword synchronized in its implementation.
  9. Implement a new model GetNSet, which is halfway between unsynchronized and synchronized, in that it does not use synchronized code, but instead uses volatile accesses to array elements. Implement it with the get and set methods of java.util.concurrent.atomic.AtomicIntegerArray.
  10. Design and implement a new model BetterSafe of your choice, which achieves better performance than Synchronized while retaining 100% reliability.
  11. Explain why your BetterSafe implementation is faster than Synchronized and why it is still 100% reliable.
  12. Design and implement yet another new model BetterSorry of your choice, which achieves better performance than BetterSafe and better reliability than Unsynchronized. However, BetterSorry should not be DRF. BetterSorry must still be deadlock-free, though; as its goal is improved performance, and deadlock yields terrible performance. Try to make BetterSorry at least as good as GetNSet in all important respects, and better than GetNSet in some.
  13. Explain why your BetterSorry implementation is faster than BetterSafe. Explain why it is more reliable than Unsynchronized, by describing race conditions that Unsynchronized is liable to but BetterSorry avoids. Give a race condition that BetterSorry still suffers from. Can you write a test program that causes this race condition to occur with high probability and thus to make BetterSorry fail? If so, please supply it (and you can reuse it to answer questions below); if not, explain why not.
  14. Integrate all the models into a single program UnsafeMemory, which you should be able to compile with the command javac UnsafeMemory.java and to run using the same sort of shell command as the test harness.
  15. For each model Synchronized, Unsynchronized, GetNSet, BetterSafe, and BetterSorry, measure and characterize the model's performance and reliability. Discuss any problems you had to overcome to do your measurements properly. Explain whether and why the model is DRF; if it is not DRF give a reliability test (as a shell command "java UnsafeMemory model ...") that the model is extremely likely to fail on the SEASnet GNU/Linux servers.
  16. Compare the models' reliability and performance to each other. Does any model seem to be the best choice for GDI's applications?

You may want to look at java.util.concurrent, java.util.concurrent.atomic and java.util.concurrent.locks for implementation ideas.

Submit

Submit a JAR file jmmplus.jar containing your solution. It should contain a copy of the files in jmm.jar, possibly with modifications (though you should attempt to minimize these modifications). It should also contain the source code to your new models, and a PDF file report.pdf containing your explanations, discussions, and performance and reliability results. Please limit your source-code lines to 80 characters or less, and limit the report to at most two pages. See Resources for written reports and oral presentations for more information about what we're looking for in your report.


© 2014, 2015 Paul Eggert. See copying rules.
$Id: hw4.html,v 1.83 2015/02/11 13:13:32 eggert Exp $