parallel stream in java 8

1. Java 8 Parallel Streams
MASUD HASAN

2. What is stream
 A stream is not a collection , sequence , or a stream of objects
 A stream is an abstraction that holds zero or more values
 Not (necessarily) a collection : values might not be stored anywhere
 Not (necessarily) a sequence : order might not matter
 Values ,not objects : avoids mutation and side effects

3. Pipelines
 A stream source
 Zero of more intermediate operations
 A terminal operation
Collection.strteam()
.filter(…)
.map(…)
.collect(…);

4. Parallel Streams
 Sources start with stream(),parallelStream() or other stream factory
 Can be switched using parallel and sequential stream
 Parallel vs sequential is a property of the entire pipeline
 Can’t switch between parallel and sequential in the middle
 Last one wins
 Parallel makes it auto-magically go faster?
NO
collection().stream()
.filter(...)
.parallel()
.map(...)
.sequential()
.collect(...)
entire stream runs sequentially

5. Parallel stream considerations
 Parallel and sequential stream should give same result
 Parallelism leads to non determinism which is bad
 Encounter order vs processing order
 Stateful vs stateless : side effects
 Accumulation vs Reduction
 Reduction : Identity and associativity
 Explicit nondeterminism can speed things up
 Parallel has a overload , might also slow up things

6. Parallel vs sequential
Source op1 op2 terminal op
op1 op2
Source op1 op2
op1 op2 terminal op
op1 op2

7. Encounter order vs processing order
 The ordering of the source determines the ordering in the result
 Processing order is non deterministic

8. Accumulation vs Reduction
long sum = 0L;
for (long i = 0; i <= 1_000_000L; i++) {
sum += i;
}

9. A better way : reduction
1+2+3+4+5+6+7+8
{1+2}+ {3+4} + {5+6} + {7+8}
{3 + 7} + {11 + 15}
{10 + 26}
36

10. Identity Value
 The starting value of each partition in parallel stream
 Becomes the result if the stream is empty
 The values must be correct
 must really a VALUE(immutable)

11. Associativity
 Reduction operation must be associative in parallel stream

12. Where are threads
 Stream workload split and dispatched to the common-fork-join pool
 Control over concurrency is explicitly opaque in the api
 Common pool controlled by system properties

13. When go parallel
 Parallel stream has startup overload
 Typically 1000 misroseconds
 If you computation is shorter , do not even bother
 Consider parallel if N * Q >= 10,000
 N = number of elements
 Q = cost per element
 Assumptions
 Element processing is idependent and source is spliatble

14. END