MikroTik

Hello,

I have an MT router CCR1016. Connected is a Cisco CMTS C4 for cable internet. Customers on this CMTS have between 200-400 MB of Internet download.
Unfortunately, the downloads come with a single download only about 60-80 Mbit / s. In a speed test with 20 simultaneous TCP connections, the booked speed comes on full. But not with a single connection.
The CCR is connected to the BGP router with MPLS / VPLS.
If I build another "Bandwidth Test" router to the CMTS directly, it runs.
Why can it be that I achieve the speed in the bandwidth test only with many simultaneous connections, but not with a single one?
The CPU load of the test routers is under no circumstances fully utilized. ROS of the CCR and BGP is at least 6.42.5.

many Greetings
Christian

Single router core deals with single connection. Your CCR has 16 cores, so if during single connection tests you see CPU load higher than 6%, one of cores is saturated and is a bottleneck. If you run 16 or more parallel connections, load is distributed over all router cores.

A powerful router with small number of cores (such as RB1100AHx4) would do better in this scenario.

Hello,

I have an MT router CCR1016. Connected is a Cisco CMTS C4 for cable internet. Customers on this CMTS have between 200-400 MB of Internet download.
Unfortunately, the downloads come with a single download only about 60-80 Mbit / s. In a speed test with 20 simultaneous TCP connections, the booked speed comes on full. But not with a single connection.
The CCR is connected to the BGP router with MPLS / VPLS.
If I build another "Bandwidth Test" router to the CMTS directly, it runs.
Why can it be that I achieve the speed in the bandwidth test only with many simultaneous connections, but not with a single one?
The CPU load of the test routers is under no circumstances fully utilized. ROS of the CCR and BGP is at least 6.42.5.

many Greetings
Christian

Bandwidth test with TCP connection count=1 uses only one core, that's why.
Ookla's Speedtest until now provides only single connection bandwidth test, but they're updating apps to make multiple connection bandwidth test.

OK. the router has 16 CPUs each with 1200mhz. But 200 MBit / s should also create at least a CPU with 1.2Ghz.
I'll take a look at the single CPU utilization during the bandwidth test.
Can it be something else? Queues are on all routers that are involved in "no queue" or "only-hardware-queue". Connection tracking is off on all routers. Firewalls settings are not synonymous internally. All routers treat traffic as FP traffic.

so. I have now done two more tests. with 20 simultaneous connections a CPU comes to 100%, it runs 940 Mbit / s (interface 1Gbps). With 1 possible connection is no !!! CPU to 100%, but only 300 Mbps. So not on the CPU?

o.k.

I see, the CPU of BWtest-Server is on 1 Session on 95%.
I will test with faster Router...

Even if no single CPU is maxed out, single connection is still sequential operation ... even if actual CPU core dealing with it changes.
Another "bonus" is that running btest on the device under throughput test seems to benchmark device's CPU almost more than connection itself.

So if the bandwidth test results seem sub-par, run iperf tests using PCs and let routers route. Note that sometimes even iperf tests can behave strangely depending on underlying IS. Once I ran UDP tests and iperf under Windows could not push more than 120Mbps in single stream. Iperf under linux did not have such limitation (at least not lower than 1Gbps).

How is it when I do a bandwidth test on multiple routers. I have a routed OSPF MPLS / VPLS backbone. Let's say I want to do a BB test from Router A to Router E via B, C, D. Where VPLS is directly between A and E. The data passes directly from A to E via VPLS, MPLS, and finally OSPF. A and E are two routers with high CPU performance. B, C and D are CCR1016. In how far is the CPU of B, C, D decisive? Is at least the CPU load correctly distributed within the VPLS transmission?

It really depends on amount of work every router is configured to perform for each packet to pass. However, if connection tracking is not active (and I can't imagine a good reason for core routers to care about connections), single connection isn't bound to single routing process so it's a good chance that intermediate routers are not CPU-bound bottlenecks.

Hi,

i have no connection-tracking activ. nowhere.

OK.
Unfortunately, after a few days I am not really smart.
I now have the following structure:

CCR1016 - -------- OSPF / MPLS / VPLS ---- ---- CCR1016----CMTS -----CM

All MT devices have latest software 6.43.8.

I do measurements from the CM (cable modem) to the 2 CCR1016 and under the two CCRs. Between the two CCR are about 20km of fiber and 2 pieces of CRS as a switch. Speed here is 1G full-duplex.

Between the two CCR I measure with only 2 simultaneous connections ~ 800 Mbit / s TCP (here the traffic is not through the VPLS tunnel, but routed directly).

I measure from the CM to the CCR before the CMTS, I also come to full-speed, as set 200MB with only 2 connections-limited by CMTS profile.

I measure from the CM to the CCR, which is 20km away (through the VPLS tunnel), definitely only 30-50MB with one connection. And also, more must be possible with an CCR1016.

Why? Are VPLS slowing down so much here?

CHristian

Hi,

It can only be a problem with MPLS / VPLS.
Without MPLS / VPLS go purely from IP to IP even with only one connection ~ 800 Mbit / s. If I give the VPLS ever an IP (from a / 30) and test here to the respective other IP go with new a connection ~ only 200 Mbit / s (when other Traffic past the VPL is going down to 50-70 Mbit/s) and the CPU utilization of the two routers is just once max. 20% per CPU. So here's a problem with VPLS. Only which one? I have no TE config or the like.

Christian