• Stars
    star
    120
  • Rank 288,903 (Top 6 %)
  • Language
    Scala
  • License
    Apache License 2.0
  • Created over 7 years ago
  • Updated almost 6 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Drizzle integration with Apache Spark

Drizzle: Low Latency Execution for Apache Spark

Drizzle is a low latency execution engine for Apache Spark that is targeted at stream processing and iterative workloads. Currently, Spark uses a BSP computation model, and notifies the scheduler at the end of each task. Invoking the scheduler at the end of each task adds overheads and results in decreased throughput and increased latency.

In Drizzle, we introduce group scheduling, where multiple batches (or a group) of computation are scheduled at once. This helps decouple the granularity of task execution from scheduling and amortize the costs of task serialization and launch.

Drizzle Example

The current Drizzle prototype exposes a low level API using the runJobs method in SparkContext. This method takes in a Seq of RDDs and corresponding functions to execute on these RDDs. Examples of using this API can be seen in DrizzleSingleStageExample and DrizzleRunningSum.

To try out Drizzle locally, we first build Spark based on existing instructions. For example, using SBT we can run

  ./build/sbt package

We can run then run the DrizzleRunningSum example with 4 cores for 10 iterations with group size 10. Note that this example requires at least 4GB of memory on your machine.

  ./bin/run-example --master "local-cluster[4,1,1024]" org.apache.spark.examples.DrizzleRunningSum 10 10

To compare this with existing Spark, we can run the same 10 iterations but now with a group size of 1

  ./bin/run-example --master "local-cluster[4,1,1024]" org.apache.spark.examples.DrizzleRunningSum 10 1

The benefit from using Drizzle is more apparent on large clusters. Results from running the single stage benchmark for 100 iterations on a Amazon EC2 cluster of 128 machines is shown below.

Status

The source code in this repository is a research prototype and only implements the scheduling techniques described in our paper. The existing Spark unit tests pass with our changes and we are actively working on adding more tests for Drizzle. We are also working towards a Spark JIRA to discuss integrating Drizzle with the Apache Spark project.

Finally we would like to note that extensions to integrate Structured Streaming and Spark ML will be implemented separately.

For more details

For more details about the architecture of Drizzle please see our Spark Summit 2015 Talk and our Technical Report

Acknowledgements

This is joint work with Aurojit Panda, Kay Ousterhout, Mike Franklin, Ali Ghodsi, Ben Recht and Ion Stoica from the AMPLab at UC Berkeley.

More Repositories

1

shark

Development in Shark has been ended.
Scala
994
star
2

SparkNet

Distributed Neural Networks for Spark
Scala
605
star
3

keystone

Simplifying robust end-to-end machine learning on Apache Spark.
Scala
468
star
4

spark-ec2

Scripts used to setup a Spark cluster on EC2
Python
390
star
5

graphx

Former GraphX development repository. GraphX has been merged into Apache Spark; please submit pull requests there.
Scala
353
star
6

snap

Scalable Nucleotide Alignment Program -- a fast and accurate read aligner for high-throughput sequencing data
C++
283
star
7

succinct

Enabling queries on compressed data.
Java
278
star
8

docker-scripts

Dockerfiles and scripts for Spark and Shark Docker images
Shell
259
star
9

spark-indexedrdd

An efficient updatable key-value store for Apache Spark
Scala
249
star
10

datascience-sp14

Repository for data science course Spring 14
Shell
182
star
11

MLI

An API for Distributed Machine Learning
Scala
154
star
12

training

Training materials for Strata, AMP Camp, etc
Scala
150
star
13

carat

Carat: Collaborative Energy Debugging
Java
114
star
14

velox-modelserver

Scala
110
star
15

benchmark

Large scale query engine benchmark
Python
99
star
16

ml-matrix

Distributed Matrix Library
Scala
70
star
17

ampcrowd

A RESTful web service that runs microtasks across multiple crowds, provides quality control techniques, and is easily extensible.
Python
51
star
18

smash

Benchmarking toolkit for variant calling
Python
46
star
19

training-scripts

Scripts to launch cluster used for Strata
Python
33
star
20

ernest

Code for Ernest
Python
32
star
21

cyclades

Cyclades
C++
28
star
22

succinct-cpp

Succinct C++
C++
24
star
23

ampcamp

scripts used for ampcamp
Python
16
star
24

zipg

A Memory-efficient Graph Store for Interactive Queries
Java
12
star
25

orchestra

Fine-Grained Distributed Computing
Python
11
star
26

iolap

Scala
11
star
27

cs262a-fall2016

HTML
9
star
28

sprint

Sprint Transformations for RegEx queries
C++
8
star
29

keystone-example

A example skeleton for an application built on top of KeystoneML
Shell
8
star
30

mlsys

An open source survey of the emerging systems for large-scale machine learning.
CSS
4
star
31

clipper-v0

Rust
3
star
32

ray-core

Experiments for the Ray backend
C++
3
star
33

Buggypedia

Objective-C
3
star
34

sparse-covariance-inverse

1
star
35

siren-release

Public version of the SiRen project
Scala
1
star
36

keystone-integration-tests

Integration Tests for KeystoneML
Shell
1
star
37

ampcrowd-client-py

A python client for using the AMPCrowd service
Python
1
star