redesign ofp inventory looses switches after few thousands of flows configured; often a problem to connect 25 switches

This job https://jenkins.opendaylight.org/releng/view/openflowplugin/job/openflowplugin-csit-1node-cds-config-performance-lithium-redesign-only-master runs a script which configures 10k flows on 25 switches, then tries to collect stats, removes flows and collect stats again.

Having a look into the results from May15 till now, we often have a problem to connect even 25 switches without any flows.

During the script run, we loose nodes from inventory
Output before adding flows:
Cycle 0:
Thread 3:
Adding 2000 flows on 25 nodes
Thread 4:
Adding 2000 flows on 25 nodes
Thread 1:
Adding 2000 flows on 25 nodes
Thread 0:
Adding 2000 flows on 25 nodes
Thread 2:
Adding 2000 flows on 25 nodes

Output before deleting flows:
Cycle 0:
Thread 2: Deleting 2000 flows on 4 nodes
Thread 3: Deleting 2000 flows on 4 nodes
Thread 0: Deleting 2000 flows on 4 nodes
Thread 1: Deleting 2000 flows on 4 nodes
Thread 4: Deleting 2000 flows on 4 nodes

Logs contains lots of WARN and ERROR messages (with stack traces) with different message content.

The first NOK behaviour is noticed in the job

https://jenkins.opendaylight.org/releng/view/openflowplugin/job/openflowplugin-csit-1node-cds-config-performance-lithium-redesign-only-master/45/

which uses:
http://nexus01.dfw.opendaylight.org:8081/nexus/content/repositories/opendaylight.snapshot/org/opendaylight/integration/distribution-karaf/0.3.0-SNAPSHOT/distribution-karaf-0.3.0-20150515.195941-1579.zip

The last OK behaviour was in
https://jenkins.opendaylight.org/releng/view/openflowplugin/job/openflowplugin-csit-1node-cds-config-performance-lithium-redesign-only-master/44/

with image
http://nexus01.dfw.opendaylight.org:8081/nexus/content/repositories/opendaylight.snapshot/org/opendaylight/integration/distribution-karaf/0.3.0-SNAPSHOT/distribution-karaf-0.3.0-20150515.192049-1576.zip