Project 1:
**** Search for Intelligent Puzzles ****
This project is designed to explore the bag-of-tasks model of parallel
programming. One commonly known example of this model is the SETI at
Home program where a SETI server send data samples to client machines
all over the world. The client processors perform data processing
operations on the data in order to determine if the signals have high
information content that may be indicative of intelligent life on
another planet. Since the amount of data involved in the transactions
is relatively small compared to the computational requirements, it is
possible to obtain extremely high performance using diverse and
distributed systems. Since the order of task execution is not
important for correct operation (this is the bag part of
bag-of-tasks), it is possible to get good processor utilization even
if workloads vary significantly. Since new tasks are assigned to
processors as they finish their currently assigned tasks, a
slow processor is naturally assigned fewer tasks to compensate.
In this problem, we will solve a related bag-of-tasks problem. In
this problem we are searching for intelligent puzzles. These puzzles
are in the general class of peg hopping games that are played on a 5x5
game board. In each position on the 5x5 grid may hold a peg, a hole
that a peg can fit in, or an unused space (no hole or peg). The game
is played by applying the following rule:
if a hole is position next to two pegs, then the peg on the opposite side can
jump over the middle peg to the hole. The middle peg is removed. For example,
if 0 is a hole and X is a peg, then this is illustrated as
0XX -> X00
This rule can be applied in either a horizontal or vertical direction.
A game is solved by multiple applications of the rule. A puzzle has a
solution if there are a sequence of moves such that all but one peg is
removed. A puzzle is determined to be intelligent if a solution exists
using the given rules. For example, below shows a proof of
intelligence for a puzzle. The puzzle is the configuration on the
leftmost side.
X 0 0 0 0 X X 0 0
XXX XX0 00X X0X XXX XX0 00X 000 00X
XX0->XXX->XXX->0XX->00X->000->000->00X->000
XXX XXX XXX 0XX 00X 00X 00X 00X 000
However, not all game configurations have a solution. We will be
given a selection of game configurations and it is our task to find
out how many of them and which ones have solutions. We input a game
configuration using a string of 25 characters. In each character we
have a "0" for a hole, a "1" for a peg, and a "2" for neither. Once
we have input the configurations from a file, we will send these
configurations to a number of clients for computations to determine if
the puzzle is intelligent or not. If the puzzle is not intelligent,
it will need to send a response to the server indicating that no
solution is found. If it is intelligent, it will need to send a
response to the server that includes the proof of intelligence (the
sequence of moves that demonstrates that the puzzle has a solution).
This directory includes the basic tools for solving this problem on a
single processor including a class that can convert a string of
characters into a configuration and a function that searches for a
possible solution using an exhaustive depth-first search algorithm.
Since some puzzles may have a larger search space than others, the
amount of time required to solve any given puzzle cannot be predicted
in advance. A bag-of-tasks model provides a perfect strategy for
solving this problem. The bag-of-tasks model is implemented as using
server-client strategy where the server implements a task queue.
Tasks (puzzles) are given to clients to work on. As clients finish
their work, new tasks are assigned to them from the queue of
unassigned tasks. Please refer to the notes discussed at the
beginning of the class for more in-depth discussion of the bag-of-tasks
parallelization strategy.
Since many processors are working on the problem at the same time, it
is possible to reduce the overall run-time to finish. However, there
are several overheads that are possible: load imbalance due to
different finishing times, time spent idle waiting for messages to
arrive, and bandwidth bottlenecks on the server process. The queue
scheduling strategy will work to eliminate the load imbalance due to
different finishing times. This is because clients will be assigned
work only when they become idle, a client that is assigned along
running task will not have the opportunity to request as much work as
clients that are assigned fast running tasks. Another overhead that
can be found occurs because the communication cost can be modeled as
t_comm = t_s + m*t_w, that is the communication time has a fixed cost,
t_s, and a per word bandwidth cost, t_w, that is how much additional
cost must be paid to send a larger message. We can reduce message
passing costs by combining several messages into one message, thus
saving multiple setup fixed costs t_s. In the bag-of-tasks model,
this saving can be implemented by sending a batch of tasks (known as a
chunk) to the client at one time. This can reduce the total number of
messages exchanged between the server and the clients, but may
increase load imbalance at the end since the last client to execute
may be assigned a chunk full of long running tasks. Finally, one can
reduce idle time of clients by having the clients working on two tasks
at a time, when the client finishes one task it can ask for another
and then begin working on the next task. In this way, the server will
deliver the next task to the client while it is working on the present
task.
In this assignment undergraduates have to implement the basic
bag-of-tasks model to parallelize this application. This means that
the server will farm tasks out to the client and collect their
results. Also, I expect the undergraduates to implement a chunking
strategy whereby tasks are sent in small clusters. This method can be
used to address performance issues related to a large startup time,
t_s. Note, the server will be responsible for writing the solutions
to the file, so once a solution is found, the results need to be sent
back to the server so that it can be written to a single output file.
For the report I expect a performance analysis using one of the
available models to anticipate how much parallel speedup should be
obtained. This analysis should include modeling the effects of
chunking on performance. Finally, run the parallel code on the
problem with the given data file and obtain speedups from 1 to 64
processors for a variety of chunk sizes and compare speedups with the
theoretical analysis. Write up the results in a short report that
includes a description of the implementation, analysis and
experimental results. Also include a final summary that includes
comparisons and contrasting statements regarding theoretical and
experimental findings.
In addition to the undergraduate assignment, graduate students are
expected to have the server process also do the work of clients using
MPI_Isend/MPI_Irecv and MPI_Test/MPI_Wait. Since in this case, the
server may be slower to respond to clients, use the method of having
the clients work on two tasks to hide latency of the server. If
chunking is used, then the processor that runs the server will test
between each task execution by the client. This should allow the
server process to do useful work as well as the clients. Your
theoretical analysis should discuss the effects of having the server
do client work. What are the benefits and costs of doing this?