Java Multithreading and CPU cores [closed]

Will be there a difference?

Yes.

How much difference there is will be highly dependent on your application.

In the unrealistic case where the application consists of N thread that do independent (i.e. no share data structures) CPU intensive computations with small memory working sets, no garbage generation and no I/O, then you can expect close to N-fold speedup with N real (i.e. not hyper-threaded) cores.

But:

• If there is shared data, then synchronization overheads may reduce the speedup.

• If the computation uses significant memory, then contention for physical memory bandwidth may cause slow down.

• If there is I/O, that will typically slow down the thread performing it ... to the effective speed of the device or network involved the data transfer.

• High rates of memory allocation may lead to slowdowns across the board. The precise effects are complicated, but adding active threads / cores could impact on throughput.

Note that these things are almost impossible to predict without detailed knowledge of the application, etcetera.

... and what does Java/OS do in backend with different threads?

Typically, the OS attempts to map each active thread to a different core. If there are too many threads for cores, then it will attempt to "time-slice" at a fairly coarse grain. Switching threads involves a "context switch" and that is another source of slowdown.

Also note that the OS has to share the cores between all processes in the system, not just your Java application's threads.

You commented:

I am sure there are people who already measured this stuff and can tell what the results are.

As I said, it is not practical to predict how an application is going to perform. Any measurements that other people have done (using artificial benchmarks) are unlikely to help you.

(I'm sure people have tried to use general models, etc to predict performance of specific application ... but it just doesn't work.)

Assuming that your four threads are capable of utilizing the core at 100%, your load is approximately equal on all threads, and there is no threading overhead, the application will run twice as fast on the dual core as the single core, and four times as fast on the quad core.

It never works out exactly that way, of course. How close your application comes to achieving that ideal outcome depends entirely on the design of your program: how well the threads utilize each processor core, and how much overhead is incurred maintaining and coordinating the threads, etc.