Mikrotik router on AMD Phenom 1090T (6 core), 3.6 Ghz with 2 port intel 82576 ethernet.
I run 7 udp bandwidth tests from this router to 7 other router with udp size packet=150 , after some time router show incredible numbers )) (conntrack enabled, no firewall rules exept 1 nat rule)
640000 pps !!! Mikrotik - good work !

nice hardware

What mother board are you using?

Tell us ya system specs and bits!

Cheers

tx drops here... in the first screenshot. Why?

Router 1:
Xeon 4 cores 2400Mhz
Intel Gigabit Desktop Adapter (82574L) PCI-E

<->

Router 2:
Core Q8400 2660Mhz
Intel Gigabit Desktop Adapter (82574L) PCI-E

Result:
rx/tx - 950000 pps

2 bda - port flapping on swich, thats why tx drops.
system specs -
AMD Phenom II X6 Black 1100T
Intel 82576 2 port ethernet

Router 1:
Xeon 4 cores 2400Mhz
Intel Gigabit Desktop Adapter (82574L) PCI-E

<->

Router 2:
Core Q8400 2660Mhz
Intel Gigabit Desktop Adapter (82574L) PCI-E

Result:
rx/tx - 950000 pps

Nice result, but CPU utilization when doing bit more real world work would be more interesting. By real world i mean PPPoE encaps/decaps, firewall, NAT, some mangling for load balancing/policy routing ....

These units are rockets..

The above results show connection test only between 2 mikrotiks . Can anybody explain how real world routing where udp packets are sent to many many hundreds of hosts affect the above results ?

The reason I ask is wouldnt IRQ usage with so many packets passing through router affect ?
IF not I am having trouble with couple of routers with 8 cores and 4 cores handling 100kpps and 100mbps (very small packet size) if the above works then i wouldnt mind switching over to the AMD 6 core with 82576 quad port adapter.

Can someone please explain how the numbers will be affected.

i've just tested two intel SR1630GP (i3 CPU, 2G ram) servers against each other (with bandwidth test)

it's happily can fill up two 1Gig NICs even with 300byte frames - that means 250k pps per direction / card!

with the high UDP load, the server with a dualport multiqueue adapter (82576 i think) have about 10% less CPU load than the other with the onboard NICs.

what about testing with 64byte or smaller packets?

Hello,

we are just working on our next High-Performance Router, the hardware is ready and supports:

Ether1 & Ether2 = Either SFP or 10/100/1000BaseT (Intel Chipset)
Ether3 .. Ether8 = 10/100/1000BaseT (Intel Chipset)
An optional 4 * SFP modul can be plugged in (Intel Chipset)
All modern Intel CPUs are supported, following tests are done with a small i3-540 (73Watt CPU) with 4 cores.

We only have to finish the 1U chassis design, than router is ready to sell. Some performance test, all traffic is generated by the 2 Boards under test!

first pictures: RX packet drops?

Hi,

need some more test & engineering, also real RFC 2544 test will follow.


Regards
Lutz

what intel chipset / nic is this? are you using bonding?

Hello,

detailed description of used hardware and features will come with the datasheet for the system when everything is ready!

Where at what do you want to bond? All tests above are w/o any kind of bonding.


Regards
Lutz

Something isnt right, with that new of hardware you should be able to triple / quaruple that. I am getting ready to experiment.

http://www.h1x.com/mt/gigabit/gigabit2.html

The above URL is a screen video I took probably 4-5 years ago, version 2.9.24. Shows the older style P4s with Intel server nics doing 600,000pps - single core days... why is the test on nicer hardware not even as good?

Sorry for my offtopic but see that hardware:

http://www.knvnet.cz/eudemon1000e.pdf

1000E-X3 is priced around 12 000 USD, but it's completelly hardware and 6 gbit/s troughput, 200M connections and 100K new connections/sec is absolutelly better then every x86 based router. If you want do that on x86, you need cca 4 PC's on 4x3,4Ghz. Power consuption of this box is max 100W. 4 those PC's is consumpting around 600W. Count how many you spend for elektricity when you have air conditioner in server house and backing up power, it grows to 850W. When you pay 0,5 USD per kWh it's 3723 USD per year versus 438 USD per year.

Conlusion? Good bye mikrotik.

PS: We have 2x Eudemon 200E-X3 ( 340K connections, 2 gbit/s troughput, 30K new conn/sec ), power consuption is 55W for 2840 USD. It works's as master/master with fully redundancy with connection synchronizing.

What's the point of that post?

Would you like me to find you a $50,000 Juniper box that performs even better than that Huawei thing you just posted? More money buys you ASICs, which raises performance. That's neither new nor surprising.

You're either trolling or spamming.

tomiczech, can I have all those features for 200 Mbps at price of $400 (in proportion) from Huawei?

p.s. is hardware antivirus upgradeable?

tomiczech
OMG! You're saying that dedicated network hardware is faster than x86?!

It's not news to anyone. People use x86 based networking gear because it's cost effective and has a high performance to cost ratio. There comes a point when you have to use hardware specifically designed for processing packets and that costs money.

I don't touch Huawei gear in my line of work. It's expressly forbidden by regulation. I'm sure there's been thorough examination down to the chips used, kinda makes me wonder if that rumor of a Chinese government backdoor has a bit of weight to it...

Back on topic

As to the testing I'm more interested in seeing forwarding performance with the test unit between and isolated from the units generating and receiving packets.

They need to add 10Gbps interface support but I want to see how RouterOS can do on the same hardware used in these tests. Intel was pushing 20Gbps with their setup.
http://download.intel.com/embedded/proc ... 322973.pdf

AMD 2650 (AM1) only one cpu usage,
But good performance
http://forum.mikrotik.com/viewtopic.php?t=104585