fabiankessler / importbench Goto Github PK

Tests behavior and performance when importing data into the titan graph database http://thinkaurelius.github.io/titan/
The data source is from Java objects (not external text files).

• Adapt the Config.java to your environment. Changing the BERKELEY_PATH should be enough.
• Run the App.java.

• Titan graph 0.3.1
• Embedded berkeley
• No external indexer active (such as ElasticSearch)
• One simple vertex with 4 fields (string, string, int, bool) one of them indexed
• Single thread for importing
• One transaction does 10k vertices
• 10 mio vertices in total
• Before each insert, I perform a string-based (dummy) lookup to see if a vertex with that name
  already exists. It never does… it’s just to be closer to my real app.

I’ve noticed non-linear execution time when batch-importing data into titan graph with berkeley.
The problem is that it slows down so dramatically in my application that the importer won’t run
to the end in reasonable time (days).

This project tries to reproduce it, and succeeds on small scale, I believe.

Here are the numbers for how it behaves “on my machine”.
My machine is a Windows 7 64bit workstation, the data folder is on a secondary ssd (no other work).
The project starts with an empty db (it creates it).

With an empty db, the first 100k vertices took 17.639 seconds.
After 10 mio vertices the transaction commit time increased to 22.283 seconds.
Then stopping the app, and continuing to insert into the same db brings the same
execution times as before. The longer the importer runs, the slower it gets.
DB size seems irrelevant.

Why? What can I do about it?