Add notes for 2020-02-07

content/posts/2020-02.md

@@ -303,4 +303,40 @@ $ ./run.sh -s http://localhost:8983/solr/statistics -a import -o ~/Downloads/sta
 
 - OK that imported! I wonder if it works... maybe I'll try another day
 
+## 2020-02-07
+
+- I did some investigation into DSpace indexing performance using flame graphs
+  - Excellent introduction: http://www.brendangregg.com/flamegraphs.html
+  - Using flame graphs with java: https://netflixtechblog.com/java-in-flames-e763b3d32166
+  - Fantastic wrapper scripts for doing perf on Java processes: https://github.com/jvm-profiling-tools/perf-map-agent
+
+```
+$ cd ~/src/git/perf-map-agent
+$ cmake     .
+$ make
+$ ./bin/create-links-in ~/.local/bin
+$ export FLAMEGRAPH_DIR=/home/aorth/src/git/FlameGraph
+$ export JAVA_HOME=/usr/lib/jvm/java-8-openjdk
+$ export JAVA_OPTS="-XX:+PreserveFramePointer"
+$ ~/dspace63/bin/dspace index-discovery -b &
+# pid of tomcat java process
+$ perf-java-flames 4478
+# pid of java indexing process
+$ perf-java-flames 11359
+```
+
+- All Java processes need to have `-XX:+PreserveFramePointer` if you want to trace their methods
+- I did the same tests against DSpace 5.8 and 6.4-SNAPSHOT's CLI indexing process and Tomcat process
+  - For what it's worth, it appears all the Hibernate stuff is in the CLI processes, so we don't need to trace the Tomcat process
+- Here is the flame graph for DSpace 5.8's `dspace index-discovery -b` java process:
+
+![DSpace 5.8 index-discovery flame graph](/cgspace-notes/2020/02/flamegraph-java-cli-dspace58.svg)
+
+- Here is the flame graph for DSpace 6.4-SNAPSHOT's `dspace index-discovery -b` java process:
+
+![DSpace 6.4-SNAPSHOT index-discovery flame graph](/cgspace-notes/2020/02/flamegraph-java-cli-dspace64-snapshot.svg)
+
+- If the width of the stacks indicates time, then it's clear that Hibernate takes longer...
+- Apparently there is a "flame diff" tool, I wonder if we can use that to compare!
+
 <!-- vim: set sw=2 ts=2: -->