Hi all,
I have followed the discussion, and think Ying's concerns have been well
responded to.
That switch is annoying, but I cannot see how we can get rid of it
easily, at least for now.
My hunch is that retransmission by unicast will work well in smaller
clusters, while broadcast may work better in larger clusters where there
are more likely to be multiple losses. I.e., when one node reports a
loss it is highly likely that even other nodes have experienced the same
loss, so a broadcast retransmission may be more optimal. But we have no
way of knowing this except from empirical data, and the optimal cluster
size where to switch may vary from system to system and even depending
on the overall load situation.
The best outcome would be if we could test this in large clusters and
find that unicast always is best. Then no switch will be needed.
The second best, given the assumption above, would be some algorithm
that auto-adapts (probably with hysteresis) and does the switch based on
counters. E.g, number of identical unicast re-transmissions as a
percentage of cluster size in the unicast->broadcast direction, and
number of broadcast duplicates received by the peers in the
broadcast->unicast direction. But my assumption may of course be totally
wrong and a different algorithm is needed, or none at all.
Anyway, series
Acked by Jon Maloy <jma...@redhat.com>
On 4/7/20 2:52 AM, Xue, Ying wrote:
Good work Tuong!
Acked-by: Ying Xue <ying....@windriver.com>
-----Original Message-----
From: Tuong Lien [mailto:tuong.t.l...@dektech.com.au]
Sent: Saturday, March 28, 2020 12:03 PM
To: jma...@redhat.com; ma...@donjonn.com; Xue, Ying;
tipc-discussion@lists.sourceforge.net
Cc: tipc-...@dektech.com.au
Subject: [PATCH RFC 0/4] tipc: add some improvements for broadcast
Hi Jon, all,
Please find the full series here,
+ For the 1st patch: it's really the last one I sent before, so you have
ack-ed already.
+ For the other ones, please help take a look. Also, I will send another
patch for iproute2/tipc which is user-space part of the last one in this
series i.e. broadcast rcv stats dumping.
Thanks alot!
Tuong Lien (4):
tipc: introduce Gap ACK blocks for broadcast link
tipc: add back link trace events
tipc: enable broadcast retrans via unicast
tipc: add support for broadcast rcv stats dumping
net/tipc/bcast.c | 22 ++-
net/tipc/bcast.h | 9 +-
net/tipc/link.c | 500 +++++++++++++++++++++++++++++++----------------------
net/tipc/link.h | 11 +-
net/tipc/msg.c | 9 +-
net/tipc/msg.h | 16 +-
net/tipc/netlink.c | 2 +-
net/tipc/node.c | 75 ++++++--
net/tipc/sysctl.c | 9 +-
net/tipc/trace.h | 17 +-
10 files changed, 424 insertions(+), 246 deletions(-)
_______________________________________________
tipc-discussion mailing list
tipc-discussion@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/tipc-discussion
