Hi Ilya,
On 6/4/24 6:27 AM, Ilya Maximets wrote:
On 6/4/24 11:54, Alin Serdean wrote:
Does it make sense to make this configurable via automake in order to avoid
future patches if we need to bump the value further?
I'm still not convinced this is an issue worth fixing.
The original discussion is here:
https://mail.openvswitch.org/pipermail/ovs-discuss/2024-April/053058.html
It's not clear if these reconnections actually cause any harm or are
even helpful in getting to connect to a less busy server in a general
case.
To me, not having to reconnect is a win, especially given the number of
client threads here. And 128 is still not that much for a server.
The original number 10 was clearly way too low for any reasonable workload.
But even with that we lived for a very long time with very large clusters
without any issues.
It could have also been that noone ever noticed, I know I didn't until
digging deeper into this.
The main problem in the original thread was that a lot of neutron clients
are having leader-only connections to the database for seemingly no reason.
That results in unnecessary mass re-connection on leadership change.
So, I'd prefer this fixed in OpenStack instead.
But isn't having many leader-only connections a perfectly valid config
when most if not all are writers? And if a small increase in the backlog
helps I don't see any reason not to do it. Fixing in Openstack would
involve one of two options from what I know - 1) Use a proxy; 2) Just
connect irregardless of leadership status. I'm not sure #2 won't cause
some other problems as the secondaries would have to forward things to
the leader, correct?
The best solution would be to have this configurable if you see that as
a viable option.
-Brian
Best regards, Ilya Maximets.
On Tue, Jun 4, 2024 at 11:05 AM Simon Horman <[email protected]> wrote:
+ Ihar
On Fri, May 31, 2024 at 03:40:08PM -0400, Brian Haley wrote:
An earlier patch [1] increased the size of the listen
backlog to 64. While that was a huge improvement over
10, further testing in large deployments showed 128
was even better.
nit: I would slightly prefer if a commit was referenced like this:
commit 2b7efee031c3 ("socket: Increase listen backlog to 64 everywhere.")
Looking at 'ss -lmt' output over more than one week for
port 6641, captured across three different controllers,
the average was:
listen(s, 10) : 1213 drops/day
listen(s, 64) : 791 drops/day
listen(s, 128): 657 drops/day
Looking at 'netstat -s | egrep -i 'listen|drop|SYN_RECV''
output over one week for port 6641, again captured across
three different controllers, the average was:
listen(s, 10) : 741 drops/day
listen(s, 64) : 200 drops/day
listen(s, 128): 22 drops/day
While having this value configurable would be the
best solution, changing to 128 is a quick fix that
should be good for all deployments. A link to the
original discussion is at [2].
[1]
https://github.com/openvswitch/ovs/commit/2b7efee031c3a2205ad2ee999275893edd083c1c
[2]
https://github.com/openvswitch/ovs/commit/2b7efee031c3a2205ad2ee999275893edd083c1c
nit: These two references are the same?
Signed-off-by: Brian Haley <[email protected]>
I'd value input on this from Ihar (CCed) who worked on the cited commit.
---
lib/socket-util.c | 2 +-
python/ovs/stream.py | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/lib/socket-util.c b/lib/socket-util.c
index c569b7d16..552310266 100644
--- a/lib/socket-util.c
+++ b/lib/socket-util.c
@@ -769,7 +769,7 @@ inet_open_passive(int style, const char *target, int
default_port,
}
/* Listen. */
- if (style == SOCK_STREAM && listen(fd, 64) < 0) {
+ if (style == SOCK_STREAM && listen(fd, 128) < 0) {
error = sock_errno();
VLOG_ERR("%s: listen: %s", target, sock_strerror(error));
goto error;
diff --git a/python/ovs/stream.py b/python/ovs/stream.py
index dbb6b2e1f..874fe0bd5 100644
--- a/python/ovs/stream.py
+++ b/python/ovs/stream.py
@@ -620,7 +620,7 @@ class PassiveStream(object):
raise Exception('Unknown connection string')
try:
- sock.listen(64)
+ sock.listen(128)
except socket.error as e:
vlog.err("%s: listen: %s" % (name, os.strerror(e.error)))
sock.close()
--
2.34.1
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
