Java Parallel Programming Performance suggestion

Question

I am building option valuation (binomial tree) using java . My binomial tree with normal sequential code with 1000 steps takes 0.1 second . But when i try to run it using parallel it takes about 65 seconds . Any suggestion and input will be appreciated..

Binomial Thread is my class which does the calultion ThreadPoolExecutor threadExecutor = new ThreadPoolExecutor(10000, 10000, 500, TimeUnit.MILLISECONDS, new ArrayBlockingQueue<Runnable>(10), new ThreadPoolExecutor.CallerRunsPolicy()); for (int i = 0; i < 10000; i++) { Runnable thread = new BinomialThread(); threadExecutor.execute(thread); }

try with 2 threads, then 5 then 10, and check how the timing is. — nafas
– nafas, Commented Mar 9, 2015 at 11:25

Davide Lorenzo MARINO · Accepted Answer · 2015-03-09 11:22:11Z

If the number of threads is greater than the number of processors and there is no waiting time (for example waiting from network data, reading from a file and so on) the performance degrades instead of increasing for the time needed to switch from one process to the other. You have 10.000 threads, that's why your program take more time in a multithreading environment than a single thread environment.

The fork/join framework would sort this out:docs.oracle.com/javase/tutorial/essential/concurrency/…

javaHunter · Accepted Answer · 2015-03-09 13:38:04Z

For an N cpu processor. The optimal number of threads should is N+1.(Check Java Concurency in Practice Chapter 8). Once you create threads beyond this number(for compute intensive tasks), performance starts to decrease because of the context switches involved. Thus this explains the numbers you were getting.

To determine the number of cpus:

int N_CPUS = Runtime.getRuntime().availableProcessors();

You then should configure the executor to N+1 threads.

Collectives™ on Stack Overflow

Java Parallel Programming Performance suggestion

2 Answers 2

1 Comment

Comments

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

1 Comment

Comments

Related