[PROTON-1672] Large deliveries comprising many transfers are handled inefficiently

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: proton-j-0.23.0
• Fix Version/s: proton-j-0.27.0
• Component/s: proton-j
• Assignee: Timothy A. Bish
• Reporter: Keith Wall
• Created: 01/Nov/17 11:25
• Updated: 23/Apr/18 08:49
• Resolved: 11/Apr/18 20:22

Description

Running performance tests using Qpid Broker J and Qpid JMS Client (0.26.0) shows that receipt of large messages is very slow in comparison with the send of the same message. For instance, sending 300MiB bytes message takes 5 seconds on my laptop. The receipt takes 97 seconds.

Instrumenting the client stack shows an obvious hot-spot in Proton-J. org.apache.qpid.proton.engine.impl.TransportSession#handleTransfer re-allocates/array copies the entire delivery buffer for ever transfer that comprises it. This leads to a non-linear loss of performance. The stack include Proton-J 0.23.0, but is looks like this code is unchanged.