Myth - E1000 is faster than VMXNET3

guys, I have clocked two virtual machines running on the same physical host and found totally different results (one of them is alligned with your, though) - physical hosts : HP Proliant DL120 G6, one CPU, four cores but I don't remember exact model of CPU, 8GB physical RAM, local SATA storage 1x 160GB HDD [RAM and storage being not important in this test as long as we generate packets, right ?] - esx v5.0 version 5.0.0 U1 fully patched up to ESXi500-201207001 - re-tested again with v5.1 last week - simple vNetwork, Gigabit/auto or Gigabit/Full duplex manual settings do not change the results at all. Switch Dlink exact model unknown, 48 port enterprise class, 19" rackmountable, no other traffic there only my ESX hosts. - no other virtual machines installed/running on those hosts. This was dedicated platform for my testing only. - virtual machine v7 and v8 - Mikrotik linux distribution , versions v5.7, v5.15, v5.17 and current latest v5.20 tested. Each VM allocated one core, later two cores in repeated test which honestly didn't change the results at all. 512MB vRAM as it's plenty enough for this system and test (used around 40MB RAM only, hey this is linux right), vmxnet3, e1000, e1000e virtual adapters. Seems like the version of Mikrotik had absolutely no influence on result of test, read more below. There is built-in test in those Mikrotik distributions enabling user to test with random or compressible data. It is possible to have receive/send/both streams. Compressible data heavily load CPU as expected and is not representative measurement due to variations, but random generated uncompressible data leads to repeatitive statistically meaningful results as there is extremely low delta in results : usually it is within 3% which is excellent. As far as I understand, the built-in throughput test in Mikrotik is based on iperf. Mikrotik has their own CLI so it is not standard linux command line thus no verification of this is possible. But I think this is not important, as far as we believe the vendor and generate packets (for sure they are not read/written from disk !! easily verified) and those packets are not compressible (verifiable but sorry I won't capture the traffic and analyze it afterwards), the test is performed in absolutely repeated and correct way. Now the results : vmxnet3 gives little around 400Mbit/s throughput (380 to 410Mbit/s) while E1000 and E1000E both yield 990Mbit/s throughput with minimum of 975Mbit/s. Let me stress out stable and repeated EVERY TIME so this is extremely consistent, with every setting, every version of ESX, Mikrotik in virtual machines. This leads me to think : is the VMXNET3 driver in Mikrotik flawed and not working optimally ? The answer is probably no. Explanation - the result of 400Mbit/s I achieved with VMXNET3 is extremely aligned with the result presented here in your blog. The only huge difference is with the E1000/E1000E results where I have three times as much performance. I'm presenting detailed description so anyone of you can re-test. If you see any problems in my "methodology", let me know. So no eggs are created equal and it is heavily possible that E1000/E1000E are FASTER than VMXNET3. I dare to re-test this with Windows 2008, 2003 and Windows7, stay tuned :)