I have got some kind of confuse in TCP performance depending on delay. My lab schema: [code]CPE --- (p1) Router (p2) --- Server[/code] The CPE connects to the Server by OpenVPN using TCP. I know about TCP-over-TCP but TCP proto is mandatory! Next I do series of network benchmarks through the OpenVPN tunnel. Series: [list=1]WITHOUT artificial delay. [*]WITH artificial delay of 30 ms (it's RTT) made on a Router's p1 and p2 interfaces. [*]WITH artificial delay of 60 ms. [/list] The result of series: [list=1]"to Server": 40 mbps, "from Server": 50 mbps [*]"to Server": 34 mbps, "from Server": 38 mbps [*]"to Server": 20 mbps, "from Server": 10 mbps [/list] Rate limit in test 1 due to CPU on CPE. It's normal. [b]But why third test with delay 60ms reduces performance "from Server" by 5 times while "to Server" only by 2 times?[/b] P.S. CPE is an OpenWRT on a Banana PI R2 board. Server is a Ubuntu server, Router is a Ubuntu server too. All devices connect via Ethernet. CPE intefrace is 100M. Others — 1G. TCP congestion control for both (CPE and Server) is BBR: [code]net.ipv4.tcp_congestion_control = bbr net.core.default_qdisc = fq [/code] Benchmark commands on a CPE: [code]"to Server": iperf3 -c ovpn_server_ip "from Server": iperf3 -c ovpn_server_ip -R [/code] Commands for artificial delay on a Router: test 2: [code]p1: sudo tc qdisc add dev ens224 root handle 1: netem delay 15ms p2: sudo tc qdisc add dev ens256 root handle 1: netem delay 15ms [/code] test 3: [code]p1: sudo tc qdisc add dev ens224 root handle 1: netem delay 30ms p2: sudo tc qdisc add dev ens256 root handle 1: netem delay 30ms [/code]

TCP performance vs delay

LovingFox

I have got some kind of confuse in TCP performance depending on delay.

My lab schema:

CPE --- (p1) Router (p2) --- Server

The CPE connects to the Server by OpenVPN using TCP. I know about TCP-over-TCP but TCP proto is mandatory! Next I do series of network benchmarks through the OpenVPN tunnel.

Series:

WITHOUT artificial delay.
WITH artificial delay of 30 ms (it's RTT) made on a Router's p1 and p2 interfaces.
WITH artificial delay of 60 ms.

The result of series:

"to Server": 40 mbps, "from Server": 50 mbps
"to Server": 34 mbps, "from Server": 38 mbps
"to Server": 20 mbps, "from Server": 10 mbps

Rate limit in test 1 due to CPU on CPE. It's normal.
But why third test with delay 60ms reduces performance "from Server" by 5 times while "to Server" only by 2 times?

P.S. CPE is an OpenWRT on a Banana PI R2 board. Server is a Ubuntu server, Router is a Ubuntu server too. All devices connect via Ethernet. CPE intefrace is 100M. Others — 1G.

TCP congestion control for both (CPE and Server) is BBR:

net.ipv4.tcp_congestion_control = bbr
net.core.default_qdisc = fq

Benchmark commands on a CPE:

"to Server": iperf3 -c ovpn_server_ip
"from Server": iperf3 -c ovpn_server_ip -R

Commands for artificial delay on a Router:

test 2:

p1: sudo tc qdisc add dev ens224 root handle 1: netem delay 15ms
p2: sudo tc qdisc add dev ens256 root handle 1: netem delay 15ms

test 3:

p1: sudo tc qdisc add dev ens224 root handle 1: netem delay 30ms
p2: sudo tc qdisc add dev ens256 root handle 1: netem delay 30ms