haoqili/SparseMatrix_Benchmark

Sparse Matrix Benchmark where each thread runs on a different phone

Android Parallel Sparse Matrix Multiplication Benchmark

This is a benchmark for doing sparse matrix on multiple phones over ad-hoc Wifi. There is 1 leader phone and multiple slave phones (the leader also has a slave component).

1. Leader sends out Sparse Matrix task to all the slaves
2. Slaves compute their portions
3. Each slave reply to leader their results
4. Leader gathers the results and outputs a final answer

Depending on using UDP or TCP, the active end of the communication (leader-end or slave-end) is different.

• Each thread runs on a different phone
• Choice of using TCP (more reliable) or UDP (a bit faster)
• Sparse Matrix code taken from Java Grande Forum multi-threaded benchmarks

This app runs Java Grande Forum's sparse matrix multiplication on different phones, where each phone completes a different thread of computation and sends its results back to the preselected master. UDP communication is used and that might account for 5% of packet disappearance (slave didn't get a packet or the master didn't get the response from slave). Later we added TCP and no packets were lost. TCP only takes about 200ms longer to run than UDP.

This is a model for building distributed apps that can be used to run tasks other than Sparse Matrix Multiplication.

Current Status

• Master runs n threads on itself and n-1 slaves.
• Hard code IP

How to Run

1. Start an ad-hoc wifi on n phones

2. Install this app on n phones

3. Look at Globals.java to change parameters, especially the MASTER_IP_ADDRESS and SLAVE_ADDRS

    In shell:
    `su`
    `ifconfig eth0 192.168.5.xxx`
    `netcfg` and you should see the changed IP
   

4. Based on whether you want to use UDP or TCP, press its appropriate buttons in the order displayed on the screen

   E.G. if you want to use UDP, you see: UDP 1. Start Slaves and UDP 2. Start Master. So you press the 1st button on the slave phone (we only have one slave as of now) and then press the 2nd button on the master phone.

5. Observe ******* end of benchmark ********* note the "ms" times right before that line

6. Get ready for the next run by:

    Manage apps -> select this app -> Force stop
   

N.B.

• See example results in results directory.

• If Globals.SPARSE_NUM_ITER is too big, SparseRunner will take a long time to run!

• Using UDP might result in packet losses, so TCP is strongly recommended (a few hundred millisecs of overhead)

Troubleshooting

• Most errors can be resolved by force stopping the app on all phones

• Error something like "Connection refused ... java.net.ConnectException ... PlainSocketImpl", that means the ports are in use. To fix the problem change in Globals.java TCP_MASTER_PORT and TCP_SLAVE_PORTS to some other numbers