CS 475, Spring 2019, Homework 4, Due April 15, 11am

Start by downloading the handout:

Download the handout

Please post questions on Piazza

Update log:

  • 3/31: Clarified: set should throw an IOException if write fails to be replicated.
  • 4/3: Clarified the textual description of innerWriteKey
  • 4/4: Clarified: setInTransaction does not acquire locks.

The purpose of this assignment is to introduce the concepts of replication and transactions. For this project, you’ll be enhancing your key value server so that keys and values are replicated to client-side caches. We’ll still have a single server, which will be in charge of maintaining locks and ensuring that new writes are replicated to all replicas. Our architecture will now look something like this:

Adding all of these replicas can increase the performance of our system when reading many large files, since each client will always have every key’s value cached. When clients read keys, they will not need to read them from the server: they can directly read them from their local cache. We will be maintaining our traditional notion of consistency: each client will always read the most recent version of each file. This means that the master will need to have some protocol to ensure that a write deemed “successful” at the server level has indeed succeeded on every active client.

Your primary interface for debugging your KVStore will be the same shell driver that we provided for HW3. We have provided stubs for all of the various commands you will need to implement, with a wrapper so that you can interact with VCFS interactively. When you run the compiled jar file, you’ll get a command prompt, like this (again, note that there are fewer commands than previously):

We have also provided a baseline suite of unit tests, which you can execute by calling  mvn test

Summary of changes compared to HW3:

  • All keys and values are now replicated to all clients with two-phase commit (read operations won’t require any network traffic)
  • Removed operations (hooray): getAll, remove, removeDirectory
  • While we’ll still use the notion of directories, you will no longer maintain any data structure to keep track of them. Instead, to implement listDirectory and putAll, your client will simply iterate over all of the keys, and find which keys match the same prefix (so these operations will appear to behave the same). The key here is that  set will now be guaranteed to touch only a single key at a time (in past HW, calling  set might result in creating parent directories too).

Here is a graphical summary showing how reads and writes will now work: when a client (replica) connects, the entire KVStore is copied to the client. Then, the client can read values directly (without going to the server). When a client wants to update a key, that update gets propagated to the master and to each replica.

General requirements:

You must use exclusively reentrant locks (e.g. synchronized or ReentrantReadWriteLock), with the single exception of lockKey and unlockKey, which must use StampedLocks.

Your KV Store will be compiled and tested using apache maven, which automatically downloads the various dependencies for VCFS and executes the provided JUnit tests. Please install Maven on your computer. Unfortunately, Maven is not installed on Zeus, however you can download the most recent version (e.g. apache-maven-3.5.2-bin.zip) and unzip it in your home directory. Then, to run maven, type ~/apache-maven-3.5.2/bin/mvn in the snippets below (instead of just mvn). Note that you can easily import maven projects into eclipse and IntelliJ.

To compile and run your shell, run mvn package  in the top level directory and then, in the server directory run  java -jar target/kvstore-server-2019.3.4-SNAPSHOT.jar portnumber to start the server, and to start the client, in the client directory run  java -jar target/kvstore-client-2019.3.4-SNAPSHOT.jar portnumber.  Your can specify any free port number on your computer over 1024; we have hardcoded the client to assume that the client runs on the same computer as the server. You’ll notice that the text-mode interface we’ve provided for you has a handy help command. To build the jar file without running the tests, run  mvn -DskipTests package.

Your KV store will be automatically graded for correctness (note that there will be a manual grading phase to check hard-to-automatically-catch concurrency issues). Included with your handout is all of the automated tests that we will use to test your assignment. Upon submitting your assignment, our server will automatically compile and test your assignment and provide you with test results. We will also for this assignment use a state-of-the-art race detector to check for races in your program – this will run automatically in Autolab. You can resubmit 50 times before the deadline without penalty. To run these tests, simply execute mvn test (of course, if you do this first off, you’ll see that they all fail!)

Note: Your code must compile and run on the autograder, under Java 8. It is unlikely that you will have any difficulties developing on a Mac, Linux or Windows, but please keep in mind the possibility of portability problems. When you feel satisfied with implementing one phase of the assignment, submit to AutoLab and verify that AutoLab agrees.

Academic honesty reminder: You may NOT share any of your code with anyone else. You may NOT post your code in a publicly viewable place (e.g. in a public GitHub repository).  You may face severe penalties for sharing your code, even “unintentionally.” Please review the course’s academic honesty policy.

General coding/grading requirements:

  1. You must use exclusively reentrant locks (e.g. synchronized or ReentrantReadWriteLock)
  2. You should feel free to add whatever additional classes you wish, or any additional methods to the existing  edu.gmu.cs475.KeyValueServer and  edu.gmu.cs475.KeyValueClient. You must not modify the  edu.gmu.cs475.IKeyValueServer  interface, the  edu.gmu.cs475.AbstractKeyValueClient, any of the tests, or any of the  internal classes.
  3. Your code should be thread-safe: concurrent calls to any of these methods (or any other method in IKeyValueServer) should not have any races. It should now be clearer how this can occur — you will potentially have multiple clients attempting to interact with the server simultaneously. It is OK for you to use coarse-grained synchronization here, and we will not test for concurrent operations on your KVStore (unlike in HW2, where this was the primary focus).
  4. You must not store any state in static fields
  5. All concurrency-related grading will account for a total 0f 10% of your grade (see Part 5). We will only consider concurrency-concerns for the first 4 parts to the extent that they are preventing your assignment from passing the given tests under normal circumstances.

Part 1: Client-side cache (35%)

For the very first part, you’ll configure the client and server code so that when they connect, the clients register themselves with the server, and the server provides the clients with a set of all of the keys and their values. This will form the initial cache on the client. Then, whenever the server receives a set request, it will forward that request to all registered clients, who will in turn update their caches. When a client is done, it will notify the server that it’s disconnecting, which will allow the server to stop sending updates to it.

Implement your server in the  server project, by implementing the empty methods in  edu.gmu.cs475.KeyValueServer, and your client in the  client project, by implementing the empty methods in  edu.gmu.cs475.KeyValueClient. You should feel free to reuse the code you had from HW3, or write something different (you’ll notice that the API changed slightly).

We’ve provided you a basic client that will automatically call the server’s registerClient method, and which will perform all read operations from its local cache. Our basic implementation of the server side registerClient will return the map to the client.

For part 1, your server’s  set should: (1) take out a write-lock on the file, (2) call   innerWriteKey on each cache client, passing transaction ID 0, and (3) then update the file locally. There is no need to implement heartbeats like in HW3 – lock and unlock can just use straightforward StampedLocks. Similarly, lockKey and unLockKey can assume that they are only ever used for write locks.

For part 1, in the server, you’ll implement:

For part 1 in the client, you’ll implement:

Hint: To prevent replicas from joining or departing during a write (but still allowing concurrent writes to different files), consider using a ReentrantReadWriteLock to guard your list of replicas. Code that is reading the list of replicas (e.g. the set method, and in part 3, setInTransaction) would need to acquire a read lock, while code that is changing the list of replicas (e.g. when registering or departing) would require a write lock.

When you are ready to check your work, you should run just the tests in the test class  edu.gmu.cs475.P1Tests. To do so from maven, you should run  mvn -Dtest=P1Tests test.

Precise grading breakdown:

  • Automated functional tests ( edu.gmu.cs475.P1Tests): 32 points
    • 8 JUnit tests, 4 points each
  • Manual feedback: 3 points

Part 2: Server-initiated transactions on write (35%)

Next, you will implement a simple two-phase-commit protocol. The motivation for this is that our implementation so far does not guarantee that each client will always see the most recent file. In particular: consider the case where clients C1, C2 and C3 are connected to the server. Client C1 updates a key Foo by telling the server. The server sends the update to C1 (OK) and C2 (OK). When it tries to update C3, it is unable to contact C3 (perhaps the network is being really slow temporarily). At this point: what should we do? If C3 is crashed, then this is probably OK: but if C3 might show up again later, then for the period of time that C3 is out of communication, it has the wrong version of file foo!

We’re going to play it safe: the server will first try to reach all of the clients and tell them that they should get ready to do the update. Then, after all clients says “Yes, I’m ready to update that key” (by returning the value true, from the server’s call to the client’s innerWriteKey), the server will send a commit message (calling the method commitTransaction on each replica), which tells each client that it should perform the commit. If any one client is not able to do the update (by voting no, responding false), then the server will abort the update, canceling the change and returning an error to the original client that wanted to perform the update. Note that although all of the clients that you write will always vote yes (returning true), your server must assume that some clients may vote no. Note also that clients might throw a RemoteException instead of a vote! (which would be a vote not to commit)

To implement this portion, you’ll need to (1) update  set on the server to generate a new transaction ID for each time that set is called (any number is fine as long as it doesn’t repeat) and pass that ID to each client’s innerWriteKey, (2) adapt  innerWriteKey so that it stores transaction writes into a separate cache, and (3) implement commit and abort on the client to apply or abort that transaction. If every  innerWriteKey successfully returns (no exception and returns true), your server should call commit on each client; if not, it should call abort on each client.

We will grade your client and server side implementation separately. We have provided a set of tests,  edu.gmu.cs475.P2Tests, which automatically mock a correctly functioning server — running your client against this fake server. To run this from maven, you should run  mvn -Dtest=P2Tests test.

Precise grading breakdown:

  • Automated functional tests ( edu.gmu.cs475.P2Tests): 32 points
    • 4 JUnit tests, 8 points each
  • Manual feedback: 3 points

Part 3: Client-initiated transactions on putAll (20%)

Finally, you’ll extend the notion of transactions, allowing clients to define themselves when a transaction will start. In particular, you’ll configure the client so that before a putAll starts, it creates a transaction. This way, even if writing to a single key on a single cache replica fails, the entire putAll can be aborted, and the invariant that putAll is atomic is preserved.

putAll should, in this order:

  1. Acquire locks on all keys being set
  2. Start a transaction ( startNewTransaction)
  3. Update each key by telling the server to, passing that transaction ID ( setInTransaction)
  4. If all writes succeeded, then commit the transaction ( issueCommitTransaction), else, abort it ( issueAbortTransaction)
  5. Release all locks (regardless of exceptions that may have occurred in the above)

The spec for all of the remaining server methods:

And, the spec for all of the remaining client methods:

Again, we will grade your client and server side implementation separately. We have provided a set of tests, edu.gmu.cs475.P3Tests, which automatically mock a correctly functioning client — running your server against this fake client and simulating error conditions. To run this from maven, you should run  mvn -Dtest=P3Tests test.

Precise grading breakdown:

  • Automated functional tests ( edu.gmu.cs475.P3Tests): 15 points
    • 5 JUnit tests, 3 points each
  • Manual feedback: 5 points

Part 4: Concurrency (10%)

To receive a top score on this assignment, you will also need to be sure that your code has no races. For this assignment, we will be using the tool, RV-Predict to detect races that may occur while running your tests. RV-Predict will give you precise feedback on the races that it detects, for instance:

AutoLab will automatically run RV-Predict on all of your submissions. You can run it on your own computer by downloading and installing it (it’s free for non-commercial use). When you run your tests with maven, use the command mvn -Drvpredict=/path/to/rv-predict.jar test (on Mac this would be  mvn -Drvpredict=/Applications/RV-Predict/Java/lib/rv-predict.jar test).

We will not give you a direct equation to correlate from # of reports from RV-Predict -> a grade on this section. We will manually award up to 10 points for concurrency correctness based on no apparent races and no over-synchronization (again, one way to avoid races could be to force every operation to be serial; this would not be ideal or correct). Moreover, note that RV-Predict will find many races, but will not find all races, which we might by hand! Note also that if you make wide use of stamped locks (instead of reentrant locks or synchronized blocks), RV-Predict will report that races are possible because it can not analyze stamped locks.

Grading

Your assignment will be graded on a series of functional tests, making sure that it implements the specification above.

In accordance with the “reasonable person principle,” we reserve the right to audit your code and correct any marks that are improperly assigned, for instance, due to your code incorrectly following the specification, but passing the test.  For instance: in HW1, there was a test that made sure that you allowed for very long argument lines (the requirement was that you allowed for arbitrarily long lines). The test had a very long line, of say, 9,000 characters. Hence, some students hard-coded their shells to accept input lines of 10,000 characters, which passes the test but does not meet the specification. We would encourage you to spend your time correctly implementing the assignment, and not trying to force it to pass the test.

Hand In Instructions

You must turn in your assignment using Autolab (You MUST be on the campus network, or connected to the GMU VPN to connect to Autolab). If you did not receive a confirmation email from Autolab to set a password, enter your @gmu.edu (NOT @masonlive) email, and click “forgot password” to get a new password.

Create a zip file of the root directory in your assignment (please:  .zip, not  .tgz or  .7z etc) — this is the root directory that includes the client, shared, server directories. When you upload your assignment, Autolab will automatically compile and test it. You should verify that the result that Autolab generates is what you expect. Your code is built and tested in a Linux VM. Assignments that do not compile using our build script will receive a maximum of 50%. Note that we have provided ample resources for you to verify that our view of your assignment is the same as your own: you will see the result of the compilation and test execution for your assignment when you submit it.

You can resubmit your assignment an unlimited number of times before the deadline. Note the course late-submission policy: assignments will be accepted up until 24 hours past the deadline at a penalty of 10%; after 24 hours, no late assignments will be accepted, no exceptions.

Note – You MUST be on the campus network, or connected to GMU VPN to connect to Autolab.

Decoding the output:

Note, AutoLab will run your code on the tests twice: once without RV-Predict (these are the scores used for parts 1-3), and once with RV-Predict (this is informational only). The outcomes should be the same with or without RV-Predict, but we wanted to make 100% sure that adding the tool doesn’t break your otherwise seemingly functioning code.

AutoLab scoreboard:

For this assignment we’ve enabled the scoreboard. It will show everyone’s scores across all of the parts of the assignment, anonymized based on either (1) the nickname that you set in AutoLab, or (2) a “random” name (drawn from the testdir random names; if your nickname was previously your name or email we changed it to a random name to make sure you understand that it will now be shown to all). Feel free to change your nickname to anything (be it your real name or a fake name), especially if it’s funny to you or others, but again, know that it will be visible to all of your classmates.

Questions

Please post questions on Piazza

Contact