This repository and the benchmark suite is derived from HiBench. It is extended to benchmark Thrill and Apache Flink, and contains scripts to run the benchmarks on Amazon's EC2 Cloud and on SLURM-operated HPC systems.
To run benchmarks on Amazon's EC2 Cloud we pull up the following test cluster using scripts in setup-ec2/.
+-------------+ +-----------+
| Control Box |---+--->| Compute 1 |
+-------------+ | +-----------+
|
| +-----------+
+--->| Compute 2 |
| +-----------+
|
| ... more ...
The control box hosts a NFS server, which is exported to the compute nodes as /home/. The NFS volume contains all sources, report logs, etc, but no compute data.
Additionally, we setup a ceph distributed file system using the compute machines' ephemeral local disks to store "Big Data". The control box contains only the central ceph services like the metadata server, while the compute nodes are the ODS nodes. The ceph DFS is mounted to /efs/ on all boxes.
We select a c3.8xlarge on-demand EC2 instance for the control box, and r3.8xlarge spot instances for the compute nodes.
Launch a control box. We start with a current Ubuntu LTS image: ubuntu/images/hvm-ssd/ubuntu-xenial-16.04-amd64-server-20171121.1 - ami-aa2ea6d0
aws --region us-east-1 ec2 run-instances --image-id ami-aa2ea6d0 \
--key-name rsa.tb2 --instance-type c3.4xlarge \
--security-groups default \
--placement "AvailabilityZone=us-east-1a,GroupName=cluster-1a" \
--enable-api-termination \
--block-device-mappings '{ "DeviceName": "/dev/sda1", "Ebs": { "VolumeSize": 60, "DeleteOnTermination": true, "VolumeType": "gp2" } }' \
--ebs-optimized
Log into the box and run the following setup script:
wget https://raw.githubusercontent.com/thrill/fst-bench/master/setup-ec2/setup-control.sh
chmod +x setup-control.sh
./setup-control.sh
The script will first install lots of additional useful packages, then setup the control box as an NFS server and as ceph Monitor and MDS. If the script breaks at some point, please check and continue by manually copy and pasting. The amount of things that can go wrong due to some small change in the dependencies is large, often it is just a single additional keystroke needed.
After running the script, reboot the box to load the newest kernel.
Launch one or more compute boxes. Again we start with the current standard Ubuntu LTS image. (On eu-west-1 use ami-0ae77879)
aws --region us-east-1 ec2 request-spot-instances \
--spot-price "2.00" --instance-count 1 \
--type "one-time" \
--launch-specification \
'{"ImageId": "ami-aa2ea6d0","InstanceType": "r3.8xlarge", "KeyName": "rsa.tb2", "SecurityGroups": ["default"], "Placement": {"AvailabilityZone": "us-east-1a", "GroupName": "cluster-1a"}, "EbsOptimized": false }'
For each compute box, run the following setup on the control box. Replace $BOXIP with the IP of the compute box in the internal VPC network.
cd ~/fst-bench/setup-ec2/
./setup-compute.sh $BOXIP
-
Current version: 5.0
-
Release date: 2015-10-12
-
To ask a question or report an issue, please use Github Issues
-
Contents:
- Overview
- Getting Started
- Advanced Configuration
- Possible Issues
This benchmark suite contains 10 typical micro workloads. This benchmark suite also has options for users to enable input/output compression for most workloads with default compression codec (zlib). Some initial work based on this benchmark suite please refer to the included ICDE workshop paper (i.e., WISS10_conf_full_011.pdf).
Note:
- Since HiBench-2.2, the input data of benchmarks are all automatically generated by their corresponding prepare scripts.
- Since HiBench-3.0, it introduces Yarn support
- Since HiBench-4.0, it consists of more workload implementations on both Hadoop MR and Spark. For Spark, three different APIs including Scala, Java, Python are supportive.
- Since HiBench-5.0, it introduces Streaming related workloads including 4 frameworks: SparkStreaming, Storm, Storm-Trident and Samza.
Job based Micro benchmarks:
For job based tasks (in the contrast to streaming based tasks), HiBench provide following workloads:
-
Sort (sort)
This workload sorts its text input data, which is generated using RandomTextWriter.
-
WordCount (wordcount)
This workload counts the occurrence of each word in the input data, which are generated using RandomTextWriter. It is representative of another typical class of real world MapReduce jobs - extracting a small amount of interesting data from large data set.
-
TeraSort (terasort)
TeraSort is a standard benchmark created by Jim Gray. Its input data is generated by Hadoop TeraGen example program.
-
Sleep (sleep)
This workload sleep an amount of seconds in each task to test framework scheduler.
SQL:
-
Scan (scan), Join(join), Aggregate(aggregation)
This workload is developed based on SIGMOD 09 paper "A Comparison of Approaches to Large-Scale Data Analysis" and HIVE-396. It contains Hive queries (Aggregation and Join) performing the typical OLAP queries described in the paper. Its input is also automatically generated Web data with hyperlinks following the Zipfian distribution.
Web Search Benchmarks:
-
PageRank (pagerank)
This workload benchmarks PageRank algorithm implemented in Spark-MLLib/Hadoop (a search engine ranking benchmark included in pegasus 2.0) examples. The data source is generated from Web data whose hyperlinks follow the Zipfian distribution.
-
Nutch indexing (nutchindexing)
Large-scale search indexing is one of the most significant uses of MapReduce. This workload tests the indexing sub-system in Nutch, a popular open source (Apache project) search engine. The workload uses the automatically generated Web data whose hyperlinks and words both follow the Zipfian distribution with corresponding parameters. The dict used to generate the Web page texts is the default linux dict file /usr/share/dict/linux.words.
Machine Learning:
-
Bayesian Classification (bayes)
This workload benchmarks NaiveBayesian Classification implemented in Spark-MLLib/Mahout examples.
Large-scale machine learning is another important use of MapReduce. This workload tests the Naive Bayesian (a popular classification algorithm for knowledge discovery and data mining) trainer in Mahout 0.7, which is an open source (Apache project) machine learning library. The workload uses the automatically generated documents whose words follow the zipfian distribution. The dict used for text generation is also from the default linux file /usr/share/dict/linux.words.
-
K-means clustering (kmeans)
This workload tests the K-means (a well-known clustering algorithm for knowledge discovery and data mining) clustering in Mahout 0.7/Spark-MLlib. The input data set is generated by GenKMeansDataset based on Uniform Distribution and Guassian Distribution.
HDFS Benchmarks:
-
enhanced DFSIO (dfsioe)
Enhanced DFSIO tests the HDFS throughput of the Hadoop cluster by generating a large number of tasks performing writes and reads simultaneously. It measures the average I/O rate of each map task, the average throughput of each map task, and the aggregated throughput of HDFS cluster. Note: this benchmark doesn't have Spark corresponding implementation.
Streaming based Micro benchmarks:
- Streaming (streamingbench)
Starting from HiBench 5.0, we provide following streaming workloads for SparkStreaming, Storm, Storm-Trident and Samza:
| Benchmark | Data type | Complexity | Store state involvment |
|---|---|---|---|
| Identity | Text | Single Step | Not Involved |
| Sample | Text | Single Step | Not Involved |
| Project | Text | Single Step | Not Involved |
| Grep | Text | Single Step | Not Involved |
| Wordcount | Text | Multi Step | Involved |
| Distinctcount | Text | Multi Step | Involved |
| Statistics | Numeric | Multi Step | Involved |
a) Data type
Big data benchmarks can be roughly classified into two types, Textual and Numeric. Text data is converted from the data source in SQL related benchmarks, which is user visit logs generated by Zipfian distribution. Numeric data is converted from vectors in Kmeans data samples.
b) Complexity
Some basic opertions are essential for understanding in any data stream computation frameworks, including identity, sample, project and grep. And multi-step operations like wordcount, distinctcount and statistics are considered for sophisticated applications.
c) Store state
One feature of stream processing is integrating stored and streaming data, which may require referring to historical information and may result in updating global status either in disk or in memory. wordcount, distinctcount and statistics are provied for a demonstration for such reqirement.
Supported hadoop/spark/storm/samza release:
- Apache release of Hadoop 1.x and Hadoop 2.x
- CDH4/CDH5 release of MR1 and MR2.
- Spark1.2 - 1.6
- Storm 0.9.3
- Samza 0.8.0