When we deployed the iSchool instance using deploy-tools (minimal instance with one master and 2 compute nodes, GFS on the compute nodes); starting Postgres on the GFS mount was very slow.  This problem also exists on our current workbench instance at SDSC.

To repeat:

• Send Craig your public key to add to authorized keys on our ops node
• ssh to ops node:  
  • ssh core@132.249.238.166
• Connect to deploy-tools container
  • docker exec -it f5405cc793db bash
• ssh to node with Gluster client/mount
  • ssh workbench-node1
• Start postgres container on local disk
  • mkdir /tmp/pgtest
  • docker run -v /tmp/pgtest:/var/lib/postgresql/data postgres:9.3
• Start postgres container on Gluster mount
  • mkdir /var/glfs/global/pgtest
  • docker run -v /var/glfs/global/pgtest:/var/lib/postgresql/data postgres:9.3
• Note difference in startup time.

Postgres is creating ~1000+ files during initialization, and Gluster is known to have some performance problems with lots of small files.  The slow startup causes problems with some services that depend on Postgres.  We've solved this by changing the services to wait for Postgres to be ready, but are wondering if there's a way to improve performance for this scenario.