[dpdk-users] Packet losses using DPDK

[David Fernandes <dfernandes@toulouse.viveris.com>]

Hi !

I am working with MoonGen which is a fully scriptable packet generator 
build on DPDK.
(→ https://github.com/emmericp/MoonGen)

The system on which I perform tests has the following characteristics :

CPU : Intel Core i3-⁠6100 (3.70 GHz, 2 cores, 2 threads/⁠core)
NIC : X540-⁠AT2 with 2x10Gbe ports
OS : Linux Ubuntu Server 16.04 (kernel 4.4)

I coded a MoonGen script which requests DPDK to transmit packets from 
one physical port and to receive them at the second physical port. The 2 
physical ports are directly connected with an RJ-45 cat6 cable.

The issue is that I perform the same test with exactly the same script 
and the same parameters several times and the results show a random 
behavior. For most of the tests there is no losses but for some of them 
I observe packet losses. The percentage of lost packets is very 
variable. It happens even when the packet rate is very low.

Some examples of random failed tests :

# 1,000,000 packets sent (packets size = 124 bytes, rate = 76 Mbps) → 
10170 lost packets

# 3,000,000 packets sent (packets size = 450 bytes, rate = 460 Mbps) → 
ALL packets lost


I tested the following system modifications without success :

# BIOS parameters :

     Hyperthreading : enable (because the machine has only 2 cores)
     Multi-⁠⁠processor : enable
     Virtualization Technology (VTx) : disable
     Virtualization Technology for Directed I/⁠⁠O (VTd) : disable
     Allow PCIe/⁠⁠PCI SERR# Interrupt (=PCIe System Errors) : disable
     NUMA unavailable

# use of isolcpus in order to isolate the cores which are in charge of 
transmission and reception

# hugepages size = 1048576 kB

# size of buffer descriptors : tried with Tx = 512 descriptors and Rx = 
128 descriptors and also with  Tx = 4096 descriptors and Rx = 4096 
  descriptors

# Tested with 2 different X540-⁠T2 NICs units

# I tested all with a Dell FC430 which has a CPU Intel Xeon E5-2660 v3 @ 
2.6GHz with 10 Cores and 2threads/Core (tested with and without 
hyper-threading)
     → same results and even worse


Remark concerning the NIC stats :
      I used the rte_eth_stats struct in order to get more information 
about the losses and I observed that in some cases, when there is packet 
losses,  ierrors value is > 0 and also ierrors + imissed + ipackets < 
opackets. In other cases I get ierrors = 0 and  imissed + ipackets = 
opackets which has more sense.

What could be the origin of that erroneous packets counting?

Do you have any explanation about that behaviour ?

Thanks in advance.

David