On 02/22/2017 03:59 PM, Pedro Ruivo wrote:
Hi team,
The 'default-jgroups-tcp.xml" has MFC protocol without the FRAG2/3
protocol. This is broken when we send a multicast message larger than
'max-credits'. It will block forever in MFC [1]. No timeouts since we
don't have the CompletableFuture at this point.
Possible solutions are:
#1 put back FRAG2/3
advantage: we have multicast flow control.
disadvantage: all messages are fragmented (unicast and multicast), that
probably requires more resources (more messages in NAKACK and UNICAST
tables?)
#2 remove MFC
advantage: probably low resources usages. TCP will handle any fragmentation.
disadvantage: we don't have multicast flow control.
#3 alternative?
Let MFC go one message beyond the limit? I don't know if that's
desirable, but it's about having less messages on the wire = in some
buffers. I've seen FRAG2 buffering a many parts of a big
StateResponseCommand. So it might have similar effect if we allow one
big message to be sent, rather than splitting it into parts and
buffering that anyway on the receiver side.
Just my 2c
Radim
Cheers,
Pedro
[1] actually, I need a thread dump to confirm it.
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev
--
Radim Vansa <rvansa(a)redhat.com>
JBoss Performance Team