Arvados

A comprehensive performance and stress test for Keep would be useful.

A stress test should:

• Flood Keep with concurrent GET and PUT requests
• Include some invalid requests
• Record performance metrics including:
  • Number of successful requests
  • Number and breakdown of error results
  • Shortest, longest, and median response time
  • Any inconsistencies on Keep volumes after testing is done
• Configurable options could include:
  • Maximum concurrency for requests (e.g. --concurrency=50 means max 50 requests at a time)
  • SSL or non-SSL

A useful experiment would be to perform stress tests with the same parameters against new Keep and old Keep.

Benchmark tools that may prove useful:

• Apache Benchmark (http://httpd.apache.org/docs/2.2/programs/ab.html)
• wrk (https://github.com/wg/wrk)
• curl-loader (http://curl-loader.sourceforge.net/)
• httperf (http://www.hpl.hp.com/research/linux/httperf/)
• autobench (http://www.hpl.hp.com/research/linux/httperf/)