Skip to content

Latest commit

 

History

History
171 lines (127 loc) · 5.46 KB

README.md

File metadata and controls

171 lines (127 loc) · 5.46 KB
Raw

Skiis: Streaming + Parallel collection for Scala

Think "Parallel Skiis"

Skiis are streaming / iterator-like collections that support both "stream fusion" and parallel operations.

Skiis address different needs (and some deficiencies depending on your point of view) than the standard Scala {serial, parallel} collections,

  • All operations from Skiis[T] => Skiis[U] are lazy -- processing happens only on as-needed basis. This applies to both serial (e.g. map) and parallel operations (e.g., parMap)

  • Serial/parallel processing are explicit by name (e.g., map vs parMap) -- no guessing how each operation processes the data.

  • Streaming-oriented framework is memory-safe -- won't cause bad surprises since none of the operations will suddenly materialize a huge data collection in memory or lead to memory retention as is sometimes the case when handling Streams (i.e., keeping reference to the head). Skiis have fewer operations than Scala collections but those offered are amenable to stream- and/or parallel-processing. You can easily and explicitly convert Skiis[T] to an Iterator[T] if you want to use Scala collections methods -- caveat emptor!

  • You can mix serial and parallel processing operations together to tailor the level of parallelism at each stage of your processing pipeline for SEDA-style pull-based streaming architecture.

  • Pluggable execution context -- you can easily plug your own Executor (thread pool) and/or use a different Executor for different parallel operations. (Note this deficiency only applies to Scala 2.9.x since 2.10.0 introduced pluggable contexts).

  • Beyond pluging-in your own Executor, you can also control the level of

    1. parallelism -- number of workers simultaneously submitted to Executor,
    2. queuing -- number of work-in-progress elements between parallel operations to balance memory usage against parallelism and 3) batch size -- to balance efficiency/contention against sharing your executor(s)/thread-pool(s) with other tasks (whether they use Skiis or not).

  • Skiis[T] exposes a Control trait that supports cancellation. (This feature is currently experimental.)

See CHANGELOG.md for evolution details.

Performance

Skiis' performance is generally comparable to Scala Parallel Collections -- sometimes better, sometimes worse. It depends on your workload (types of operations), the thread pool you use, the allowable queue depth, worker batch size, and so on.

I have not yet tested Skiis on machines with > 16 CPU cores.

Examples

Launch your Scala REPL,

# launch Scala REPL
buildr shell

and you can then interactively try the Skiis[T] collections,

Welcome to Scala version 2.9.1.final (Java HotSpot(TM) 64-Bit Server VM, Java 1.7.0_04).
Type in expressions to have them evaluated.
Type :help for more information.

scala> import skiis2.Skiis

scala> import skiis2.Skiis.DefaultContext

// Define a timing function
scala> def ptime[A](f: => A) = {
     |   val start = System.nanoTime
     |   val result = f
     |   printf("Elapsed: %.3f sec\n", (System.nanoTime - start) * 1e-9)
     |   result
     | }
ptime: [A](f: => A)A

// Stream fusion

scala> ptime {
     |   Skiis(1 to 20)
     |     .map (_ * 3)
     |     .filter (_ % 2 == 0)
     |     to(List)
     | }
Elapsed: 0.001 sec
res1: List[Int] = List(6, 12, 18, 24, 30, 36, 42, 48, 54, 60)

// Parallel operations
// (note unordered nature of the output)

scala> Skiis(1 to 20)
         .parMap (_ * 3) (DefaultContext)
         .parFilter (_ % 2 == 0) (DefaultContext)
         to(List)
res3: List[Int] = List(24, 12, 6, 18, 30, 36, 48, 42, 60, 54)

// You can mix & match non-parallel and parallel operations
// Here the last parallel operation (reduce) "pulls" elements from
// previous operations in parallel.

scala> ptime {
 |   Skiis(1 to 20)
 |     .map (_ * 3)
 |     .parFilter (_ % 2 == 0) (DefaultContext)
 |    to(List)
 | }
 Elapsed: 0.003 sec
 res4: List[Int] = List(18, 12, 6, 24, 36, 30, 42, 48, 54, 60)


// compared to Scala parallel collection (we're in the same ballpark)

scala> ptime {
     | Skiis(1 to 100000)
     |   .map (_ * 3)
     |   .filter (_ % 21 == 0)
     |   .parReduce (_ + _)(DefaultContext)
     | }
Elapsed: 0.014 sec
res5: Int = 2142792855

scala> ptime {
     | (1 to 100000).par
     |   .map (_ * 3)
     |   .filter (_ % 21 == 0)
     |   .reduce (_ + _)
     | }
Elapsed: 0.015 sec
res5: Int = 2142792855

// compared to Scala "serial" collections
// (all parallel collections have some overhead)

scala> ptime {
     | (1 to 100000)
     |   .map (_ * 3)
     |   .filter (_ % 21 == 0)
     |   .reduce (_ + _)
     | }
Elapsed: 0.012 sec
res71: Int = 2142792855

Caveats

  • Similarly to Iterator, Skiis do not implement content-based equals() or hashCode(). If you want to compare Skiis, first convert them to a specific collection that supports equality, such as Seq.

Building

You need Apache Buildr 1.4.x or higher.

# compile, test and package .jars
buildr package

Target platform

  • Scala 2.10+
  • JVM 1.6+

License

Skiis is is licensed under the terms of the Apache Software License v2.0. http://www.apache.org/licenses/LICENSE-2.0.html