Details
-
Enhancement
-
Resolution: Unresolved
-
Major
-
None
-
11.0.3.Final
-
None
Description
UDP and IP multicast in particular are not reliable in some environments. E.g.
- In our test environment datagrams bigger than 9KB are sometimes dropped, causing poor performance.
- In some IPv6 environments datagrams bigger than the MTU are dropped instead of being fragmented when receiving a ICMP 4 "The datagram is too big. Packet fragmentation is required but the 'don't fragment' (DF) flag is on." packet.
- Multicast groups sometimes disappear with IGMP snooping enabled.
Users can diagnose these problems by using JGroups' McastSenderTest and McastReceiverTest on all the nodes, but it's a manual process, and it requires an investigation to know that the network might have a problem. This kind of issue first appears in the log as a generic timeout error, e.g.
org.infinispan.commons.CacheException: Initial state transfer timed out for cache org.infinispan.CONFIG on Node at org.infinispan.statetransfer.StateTransferManagerImpl.waitForInitialStateTransferToComplete(StateTransferManagerImpl.java:246)
We should try to help the user by send a big multicast message at the beginning and failing if we don't get responses from all other members.
An alternative would be to expose a multicast test as a Console/CLI operation and invoking it automatically when a diagnostic report is generated.
Attachments
Issue Links
- is related to
-
ISPN-12044 CLI performance test tool
- Closed