SparkOscope: Enabling Apache Spark Optimization through Cross Stack Monitoring with Yiannis Gkoufas
2 likes726 views
The document discusses Sparkoscope, an enhancement to Apache Spark's monitoring framework that adds OS-level metrics and improves the Spark web UI for better visualization of these metrics. It includes various modules for data export and real-time monitoring, addressing limitations in current monitoring architectures. The presentation also outlines installation instructions and future roadmap plans for expanding the project.
SparkOscope: Enabling Apache Spark Optimization through Cross Stack Monitoring with Yiannis Gkoufas
- 1. SparkOscope: Enabling Apache Spark Optimization Through Cross- Stack Monitoring and Visualization Yiannis Gkoufas IBM Research Dublin,Ireland High Performance Systems
- 2. whoami • Research Software Engineer in IBM Research, Ireland since 2012 • Work on Analytics Foundations Middleware – Distributed Frameworks, Anything Java/Scala based, Web-based POCs • High Performance Systems Group: Kostas, Andrea, Dimitris, Khalid, Michael, Michele, Mustafa, Pierre, Sri
- 3. Spark Experience • We love developing in Spark our analytical workloads and fully embraced it since the early 1.0.x versions • Last few years, used it to run jobs on large volume of energy-related sensor data
- 4. Jobs on Daily Basis • Once we managed to develop the needed jobs, they were executed in a recurring fashion • We were receiving a new batch of data every day
- 5. Fighting Bugs • When there was a bug on our code, it was very easy to discover it the Spark Web UI • We could easily retrieve information about the job, stage and line number in our source code
- 6. Fighting bottlenecks • However we couldn’t easily spot which jobs and stages were causing a slow down • What was the part of our code that was the bottleneck?
- 7. Ganglia Extension • We had the option to use the Ganglia Extension to export the metrics but: – We need to maintain/configure yet another external system – There is no association with the Spark jobs/stages/source code
- 8. Spark Monitoring Framework • We could use the built-in Spark Monitoring Framework but: – Collecting CSVs from the worker nodes and aggregating them seems cumbersome – Again we couldn’t easily extract associations with our source code of the job
- 9. Current Monitoring Architecture Spark Worker1 Spark Worker2 Executor1 Executor2 Executor3 Executor4 Executor5 Executor6 Executor Source Executor Source Executor Source Executor Source Executor Source Executor Source CSV CSV CSV CSV CSV CSV Job Execution Monitoring Framework Local Filesystem
- 11. SparkOscope Overview • Extension to enrich Spark’s Monitoring Framework with OS-level Metrics • Enhancement of the Web UI to plot all the available metrics + the newly developed OS- level metrics
- 12. SparkOscope Modules • SigarSource: Attached to the executor, leveraging Hyperic Sigar library to get OS-Level Metrics • HDFSSink: Exports all available metrics to an HDFS directory • MQTTSink: Publishes all available metrics on an MQTT Topic • Modified Web UI: Modified Spark Web UI to plot historical and realtime plots, generated from the modules
- 13. SparkOscope Flavors • Historical Plots: View metrics on the UI after the job has finished • Realtime Plots: View metrics on the UI in realtime as the job is being executed • Headless: Use SigarSource, HDFSSink, MQTTSink without viewing the plots on the UI – https://github.com/ibm-research-ireland/sparkoscope-headless
- 14. SparkOscope High-level Architecture - Historical plots Spark Worker1 Spark Worker2 Executor1 Executor2 Executor3 Executor4 Executor5 Executor6 Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Job Execution Monitoring Framework HDFS /custom-metrics/app-xxxxxxx /executor1 /executor2 /executor3 /executor4 /executor5 /executor6 Spark Web UI
- 15. SparkOscope High-level Architecture - Realtime plots Spark Worker1 Spark Worker2 Executor1 Executor2 Executor3 Executor4 Executor5 Executor6 Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Executor Sigar Source Job Execution Monitoring Framework Master /custom-metrics/app-xxxxxxx Spark Web UI MQTT Broker
- 16. SparkOscope Basic Installation • Clone the git repo: https://github.com/ibm-research- ireland/sparkoscope • Build Spark • Modify the configuration files: metrics.properties spark-defaults.conf
- 17. SparkOscope OS-level Metrics • Download the Hyperic Sigar library to all the slave nodes • Extract it anywhere in the system • Modify the configuration files metrics.properties spark-env.sh
- 18. SparkOscope Realtime Plots • Modify the configuration files metrics.properties spark-defaults.conf • Make sure that no service is currently running on ports specified on the Master • Make sure that executor.sink.mqtt.port is the same as spark.moquette.conf
- 19. SparkOscope Headless Installation • Clone the git repo: https://github.com/ibm-research- ireland/sparkoscope-headless • Build the maven project • Modify the configuration files as described for SigarSource, HDFSSink, MQTTSink • Additionally you need to append to spark.executor.extraClassPath the paths of the created jars • No need to have the patched Spark version, since the metrics are not displayed in the UI
- 20. Demo!
- 21. Roadmap • Expand the range of available Sinks and Sources • Smart recommendations on infrastructure needs derived from patterns of resource utilization of jobs • Work with the opensource ecosystem to improve it and target more use cases