Reproducing step:
Take 3 machines, connect them on the same network through a single switch (to
limits network bandwith/latency effects)
run uperf on 3 of them, 2 slave and one master.
I am using the following uperf script:
<?xml version="1.0"?>
<profile name="xml">
<group nthreads="1">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h1 protocol=tcp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="read" options="size=8k"/>
<flowop type="write" options="size=8k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
<group nthreads="1">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h2 protocol=tcp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="write" options="size=8k"/>
<flowop type="read" options="size=8k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
<group nthreads="1">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h1 protocol=udp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="read" options="size=1.4k"/>
<flowop type="write" options="size=1.4k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
<group nthreads="$n">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h2 protocol=udp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="write" options="size=1.4k"/>
<flowop type="read" options="size=1.4k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
<group nthreads="$n">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h1 protocol=udp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="read" options="size=0.2k"/>
<flowop type="write" options="size=0.2k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
<group nthreads="$n">
<transaction iterations="1">
<flowop type="connect" options="remotehost=$h2 protocol=udp
wndsz=50k tcp_nodelay"/>
</transaction>
<transaction duration="30s">
<flowop type="write" options="size=0.2k"/>
<flowop type="read" options="size=0.2k"/>
</transaction>
<transaction iterations="1">
<flowop type="disconnect" />
</transaction>
</group>
</profile>
if you put n big enough (50 in my case seems a good value) on the same hardware
you will have significant throughput differences between Bionic and Xenial.
On hundred runs, I see an avergae throuhput 719,8 Mbits/s in and 719,9 Mbits/s
out on xenial and 608,2 Mbits/s in and 608,3 Mbits/s out on Bionic. That
represent 20% network performance on bionic.
Also if I run sudo tc -s qdisc or netstat, we see packets been dropped
on Bionic but not Xenial.
I ran that test because we see network performance issue (TCP
timeouts/packet drop) when validating our application on bionic in our
datacenter, which is using a completly different hardware.
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/1862044
Title:
packet loss, extra latency and lower bandwidth on bionic
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/systemd/+bug/1862044/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
