Brian Goetz, Java Language Architect at Oracle, gave an interesting presentation “From Concurrent to Parallel” (available on InfoQ) on the subject back in 2009. Here are the most important points of his talk.

Java 8 introduces the Stream library (which was, in Brians words, developed as a showcase for the new Java 8 language features). Just calling “.parallel()” advices the library to process the stream “in parallel”, i.e. in multiple threads.

Do you really need it?

Parallel processing is an optimization and therefore the general questions regarding optimizations have to be answered:

  • Do you have any performance requirements at all (otherwise you are already fast enough)
  • Do you have any means to measure the performance?
  • Does the performance you measure violate the requirements?

Only if the answer to all these questions is “yes” you may take the time to investigate whether “.parallel()” will increase the performance.

Why the performance may not increase?

Compared to sequentionally processing a stream, processing it in parallel has overhead costs: splitting the data, managing the threads that will process the data and combining the results. So you may not see as much speed up as you might hope for.

Brian talks about several factors that undermine speedup.

NQ is insufficiently high

He talks about a simple model called “NQ model”. N*Q should be greater than 10.000. N is the number of data items. Q is factor that expresses how CPU-expensive the processing step is. “Summing numbers” or “finding the max of integers” is very inexpensive and Q would be near 1. More complex tasks would have higher values of Q. So if all you want is to add up numbers, you need a lot of numbers to see a performance gain.

Cache-miss ratio too high

If the data to be processed is saved next to each other in RAM, it will be transfered together in CPU caches. Accessing cache memory instead of main memory is very fast, so data in an array is processed much faster than data in a linked list that is spread all over the main memory. Also when the data in an array is just pointers data access will be slow. As a demonstration, Brian shows that summing up an array of (native) integers in parallel scales well with number of CPU cores. Summing up an array of Integer-objects scales very poorly because there are just pointers in the array.

The more indirections (pointers) you work with the more cache-misses will have the CPU wait for memory access.

The source is expensive to split

In order to process data in parallel, the source of the data has to be splitted in order to hand parts of the data to different CPUs. Splitting arrays is simple, splitting linked list is hard (Brian said: linked list are splitted into “first element, rest”).

Result combination cost is to high

When the result combination is the sum of numbers, that is easy to calculate. If the result is a set, and the result combination is to merge the resulting sets, this is expensive. It might be faster to sequentially add each result into one set.

Order-sensitive operations

Operations like “limit()”, “skip()” and “findFirst()” depend on the order of the data in the stream. This makes the pipelines using them “less exploitable” regarding parallelism. You can call “unordered()” if the order is not meaningful to you and the JVM will optimize those operations.

Mutating shared state is a big NO!

Of course you will lose performance if you have to mutate shared state and guarding access to it with locks etc.

Parallel streams are for CPU-heavy process steps

Brian mentions that parallel streams where built for CPU-heavy tasks. If you do mainly IO in the processing steps than use ThreadPools and Executors etc.

What degree of parallelism does the Stream API uses?

As stated in the Streams API documentation, the API uses one JVM wide Fork-Join-Pool with a number of threads that defaults to the number of processors on the computer system. Why? Because it is expected that a processing steps will utilize the CPU as much as possible. So there is no sense in having more threads then CPU cores.

In contrast, when having IO-heavy tasks you want to have many more threads than CPU cores because the threads will wait for IO most of the time and therefore not utilize the CPU.