This is not the current version of the class.

Problem set 6X: Network Pong

This optional part of Problem Set 6 teaches you to support network robustness by handling loss, delay, congestion, and unexpected inputs.

We are not covering networking in lecture this year, but lecture notes for networking are available, and working on networking may be fun for you. Out of fairness concerns, we will not evaluate network pong for extra credit.

The ./pong61 program tests network robustness (and some synchronization).

Once you have merged, edit the pset6/serverinfo.h file so that PONG_USER is defined to a unique string that only you know.

Note. We currently do not plan to implement authentication on the pong server. This means that, if you learn another person’s string, you can mess with their heads. Please do not abuse this system. Aggressive abuse will be reported to The Authorities.

It will print out a URL. Visit that URL in your browser. You should see a bouncing ball in a rectangular field:

Pong sample

pong61 works by sending HTTP messages to a Web server we run. HTTP is the application protocol on which the Web is built. Read more about HTTP before you continue. When it starts up, pong61 makes a single request to a URL like this:

http://cs61.seas.harvard.edu:6168/PONG_USER/reset

This tells the server to reset the pong board’s state. The server clears the board and returns a simple response containing the board’s width and height:

18 23

Then pong61 makes many requests to URLs like this:

http://cs61.seas.harvard.edu:6168/PONG_USER/move?x=XPOS&y=YPOS&style=on

This request causes a new ball to appear at position (XPOS,YPOS). The server responds with a numeric code and explanation. If everything goes well, it says:

0 OK

If there’s a problem, the numeric code is negative, as in:

-1 x and y parameters must be numbers

After each request, pong61 waits 0.1 seconds before making the next request.

Our handout code runs each HTTP request in its own thread. The main thread spins to wait for each thread to complete before going on to the next. This works just fine on Phase 0. To do the problem set, you must change the code so it works on the other phases too, and introduce synchronization along the way. Use the web page’s phase buttons to change phases.

Phase 1: Loss

In Phase 1, the server starts to lose messages. It will occasionally go offline for a short period. During that time, every move request is rejected by closing down the connection. The http_connection::receive_response_headers() function sets conn->cstate to cstate_broken and conn->status_code to -1 when this happens, but the pong thread ignores this problem and continues as if everything was fine. That position in the pong trail never gets filled in. Our server shows this mistake by drawing black marks in the spaces.

Your job in this phase is to detect lost messages and retry. When the server drops a connection, your code should close that connection (to free its resources) and make a new connection attempt at the same position. It shouldn’t move to the next position until the server responds for the current position.

However, you must be careful not to bombard the server while it is offline. The server will notice this and explode. Instead, you must implement a form of exponential backoff. This is a simple, powerful concept.

Exponential backoff is awesome because it responds to short outages quickly, but imposes only logarithmic overhead (i.e., the number of messages sent during the outage is logarithmic in the length of the outage). It’s ubiquitous: Ethernet is built on it, and the next time your Gmail goes offline, check out the numbers that appear after “Not connected. Trying again in...”.

Hint: Implement one phase at a time, always thinking how you could accomplish the task in the simplest correct way. Avoid overengineering! Our solution set implements all phases, without race conditions, in less than 60 lines of code.

Slow networks: The pong61 server may report spurious errors if you are doing the problem set remotely or over a slow network. Use the grading server to check for errors if you think your errors are spurious.

Phase 2: Delay

In Phase 2, the server delays its responses. It will send you the full header for its response, but delay the body. Since the handout loop waits for the body before sending the next request, the pong ball will move extremely slowly in Phase 2. Too slowly, in fact: the server enforces a minimum ball speed, and when your code is slower than that speed, you’ll see some black marks on the display.

You might think solving this problem would be easy: just close the connection before the response arrives. But the server is too clever for this. If you close a connection prematurely, it explodes.

To support delay, your pong61 must handle multiple concurrent connections to the server. Now, the main thread may need to spawn a new thread before the response arrives! But watch out. If you leak connections, the server will explode.

Your Phase 2 code must also work in Phase 1. We suggest you make Phase 2 work first on its own, then go back and make Phase 1 work again.

Phase 3: Utilization and Synchronization

So far, your pong61 client opens a new network connection for every ball. This is wasteful and slow and in Phase 3 the server will not allow it. You should instead reuse valid HTTP connections for new ball positions.

An http_connection object is available for reuse if and only if conn->cstate == cstate_idle. This means that the server sent a complete response and is waiting for another request.

Reusing connections would be really easy—except that in Phase 3 the server also drops some connections (as in Phase 1) and delays some connections (as in Phase 2). Your pong61 client must handle it all, and you must use synchronization primitives correctly.

The key function you’ll need to add is a connection table of available connections. This can be a linked list, an array, a std::list or std::deque, or whatever you’d like. When a connection reaches state cstate_idle, add it to the table. When you need to call a new RPC, check the table first, and use that connection if one exists.

Make sure that you protect your connection table from concurrent access! There should be no race condition bugs. Use synchronization objects to handle this. You should simultaneously remove the unsafe global move_done variable and replace it with a synchronization object.

Phase 4: Congestion

In Phase 4, the server sometimes behaves as if it is congested. A congested server responds to a request not with 0 OK, but with a positive number, such as this:

+1948 STOP

This means that the server is overloaded. pong61 is not allowed to send any more requests for (in this case) 1948 milliseconds—not even from concurrent threads. This should give the server enough time to cool down. The display will show a stop sign during the cool-down period, and if pong61 ignores the stop request and sends a message anyway, the server will explode. But after the cool-down period, the client should go right back to sending requests. A congestion response can sometimes be delayed, as in Phase 2; your client threads should proceed during the delay, entering the cool-down period only once the whole response is available.

Phases 1 through 3 are still active in Phase 4. Phase 4 may catch some race conditions in your code from Phase 3.

The best solutions will avoid all race conditions for non-delayed responses, meaning that the main thread will remain blocked during an immediate congestion signal (i.e., a complete congestion signal included in the server’s initial response). This may require changes to your Phase 2 code.

Phase 5: Evil

Phase 5 is a mystery, but run it and you’ll figure out the problem soon enough.

Hints and advice

For full credit, your code must not suffer from race condition bugs. You’ll need to think this through carefully, as race conditions may not show up during normal testing. ./pong61 should run cleanly with thread sanitizers. The ./pong61 -f flag, which runs pong61 faster than normal, might be useful as you look for race conditions.

We are only concerned with race conditions inside your client (i.e., between different client threads). We are not concerned with rare issues with scheduling between the server and the client, such as network reordering. It is impossible to avoid all client–server race conditions in this pset. But as usual, your code should never execute undefined behavior.

If you have extra time, implement something fun. For example, two teams could get together and implement Space Invaders (one team programming the monsters, and one team programming the spaceship)! Here are some RPCs the server implements that might be useful.