On the Locality of Java 8 Streams in Real-Time Big Data

Report
On the Locality of Java 8 Streams in RealTime Big Data Applications
Yu Chan
Ian Gray
Andy Wellings
Neil Audsley
Real-Time Systems Group, Computer Science
University of York, UK
Outline






Context of the work
Focus of the current paper
Previous work on Stored Collections
Java 8: Streams and Pipelines and their relationship to Fork
and Join framework
Explore the impact of ccNUMA and locality on the Java 8
model
Conclusions
Java 8 implementation of Streams and pipelines is very complex
2 - 22
Context I

The JUNIPER EU project is currently investigating
how the Java 8 platform augmented by the RTSJ
can be used for real-time Big Data applications
3 - 22
Context II

JUNIPER is interested in both Big Data applications
on clusters of servers and on supercomputers


Here were are concerned with the cluster environment
JUNIPER wants to use Java 8 streams to provide
the underlying programming model for the
individual programs executing on the server
computers
4 - 22
Context III

The Java support is targeted at the server
computers contained within the clusters


it is not an alternative to, for example, the Hadoop
framework whose main concern is the distribution of the
data
Current work is considering how to extend the
Java stream support to a distributed environment
5 - 22
Context IV


A JUNIPER application consists of a set of Java 8
programs (augmented with the RTSJ) that are
mapped to a distributed computing cluster, such as
an internet-based cloud service
Performance is critical for big data applications


We need to understand the impact of using Java
streams and pipelines
Currently aicas are updating Jamaica for Java 8
and to support locality
6 - 22
Focus of the Paper
To evaluate the JVM server-level support

Java is architectural neutral: the programming
model essentially assumes SMP support


But, servers nowadays tend to have a ccNUMA
architecture
The JVM has the responsibility of optimizing
performance

But, we are also interested in the potential to have
FPGA accelerators
7 - 22
Previous Work I
Java's built-in stream sources have a number of
drawbacks for use in Big Data processing
1.
the in-memory sources (e.g. arrays and collections)
store all their data in heap memory


2.
this implies populating the collection before any operations can be
performed, resulting in a potentially long delay while it takes place
heap memory is small compared to disk space, so for Big Data
computations, there may not be enough heap memory to load the entire
dataset from disk
the file-based sources (e.g. BufferedReader.lines)
produce sequential streams, making parallel
execution of the pipeline impossible
8 - 22
Previous Work II

To overcome these limitations, we have introduced
in the idea of a Stored Collection



reads its data from a file on-demand, thus eliminating
the initial population step
generates a parallel stream to take advantage of multicore hardware
Stored Collection programs are up to 1.44 times
faster and their heap usage is 2.35%- 84.1% of
those for in-memory collection programs
9 - 22
Streams and Pipelines
List<Integer> transactionsIds =
transactions.stream() .
filter(t -> t.getType() == Transaction.GROCERY) .
sorted(comparing(Transaction::getValue).reversed()) .
map(Transaction::getId) .
collect(toList());
Lazy evaluation: the data is pulled through the stream not pushed
10 - 22
Streams and Pipelines
class InputData {
private long sensorReading;
// ...
public long getSensorReading() {
return sensorReading;
}
}
class OutputData {
private byte[] hashedSensorReading;
// ...
public void setHashedSensorReading(byte[] hash) {
hashedSensorReading = hash;
}
}
11 - 22
Streams and Pipelines
class ProcessData {
public void run() {
Collection<InputData> inputs = ...;
inputs.parallelStream().map(data -> {…}).
forEach(outData -> { ... });
}
}
Operation
Input
Stream
Operation
…
Output
Stream
12 - 22
Operation
Terminal
Operation
Streams and Pipelines
class ProcessData {
public void run() {
Collection<InputData> inputs = ...;
inputs.parallelStream().map(data -> {
long value = data.getSensorReading();
byte[] hash = new byte[32];
SHA256 sha256 = new SHA256();
for (int shift = 0; shift < 64; shift += 8)
sha256.hash((byte) (value >> shift));
sha256.digest(hash);
OutputData out = new OutputData();
out.setHashedSensorReading(hash);
// ...
return out;
}).forEach(outData -> { ... });
}
}
13 - 22
Streams and Fork-Join Framework



Each parallel stream source can provide a spliterator which partitions
the stream
Internally in the Java 8 stream support, the spliterator is called to
generate sub streams
Each sub stream is then processed by a task submitted to the default
fork and join pool
14 - 22
Incore Stream Sources and Locality




Here the memory used to hold the partitioned stream source spans two
ccNUMA nodes
Hence threads executing the tasks may be accessing remote memory
In our experimental set-up, remote access is 18% slower than local
access
Setting thread affinities does not necessarily help
15 - 22
Experimental Setup

2 GHz AMD Opteron 8350 running Ubuntu 13.04






16 cores, 4 cores per NUMA node
2MB L2 cache: 512KB per node
2 MB of L3 shared cache
16 GB of main memory: 4GB per node
Swap disabled
Java SE 8u5


14 GB initial and maximum heap memory
GC avoided by reusing objects
16 - 22
Experiment

Measure the main processing time of computing
the SHA-256 cryptographic hash function on
consecutive long integers starting from 1





Without thread affinity
Binding one thread to one core
Binding not more than 4 threads to each NUMA node
Use array-backed stream and stored collectionbacked stream
For the stored collection: the data is created when
needed rather than reading from disk
17 - 22
Performance of Array-backed Streams
200 runs graph shows cumulative histograms
2 26 long integers
2 28 long integers
18 - 22
Performance of Stored Collection backed Streams
2 26 long integers
2 28 long integers
19 - 22
Experiment

Measure the execution time of computing the SHA256 cryptographic hash function on consecutive
long integers starting from 1





Without thread affinity
Binding one thread to one core
Binding not more than 4 threads to each NUMA node
Use array-backed stream and stored collectionbacked stream
This stream source is on disk: hence more similar
to a big data application
20 - 22
Array-based versus Stored Collections
Array
Stored Collection
2 28 long integers
21 - 22
Conclusions

The goal of this work has been (in the context of
Java 8 streams and pipelines) to



understand what impact a ccNUMA architecture will
have on the ability of a JVM to optimize performance
without programmer help
If we just use thread affinity, we may undermine
any attempt made by the JVM to optimize
Stored collections, a partitioned heaped (or
physical scoped memory area) should allow the
programmer more control and enforce locality of
access
22 - 22

similar documents