Poor performance on igb driver

bdaniel7

igb0 is WAN, igb1 is LAN.

I'm starting top -aSH as you suggested, then during the peak transfer, I exit from top with q.

I had powerD enabled, with all (AC power, Battery power, Unknown power) set to Maximum.
I disabled powerD but there is no difference.

And I get this sysctl: unknown oid 'dev.cpu.0.freq'

tman222

@tman222 said in Poor performance on igb driver:

Hi @bdaniel7

I also agree that the CPU should be able to handle 1Gbit speeds fairly easily, especially if you are not trying run any IDS/IPS on top regular kernel packet processing.

FreeBSD's network defaults aren't tuned too well for very high speed connections by default (although this is getting better in newer versions). Here is a link to a thread with some more parameters you can tune on your Intel NIC's:

https://forum.netgate.com/topic/117072/dsl-reports-speed-test-causing-crash-on-upload

Of those parameters, I"d probably adjust the RX/TX descriptors and processing limits first and see if that yields any improvements.

Hope this helps.

Hi @bdaniel7 - have you also tried tuning some of the additional parameters that I suggested? If yes, what were the results?

stephenw10

Sorry I meant where are you testing between? Speedtest client on igb1 connecting to a server via igb0?

Steve

bdaniel7

@stephenw10
Yes, the mediaconverter is connected to igb0, my Windows 10 client is connected to the igb1 port.

stephenw10

I don't see it having been asked so, are you connecting using PPPoE?

Steve

bdaniel7

@stephenw10
Yes, I'm using PPPoE.

stephenw10

Ah, then that is the cause of the problem. You can see that all the loading is on one queue and hence one CPU core while the others are mostly idle. It's unfortunately a known issue with PPPoE in FreeBSD/pfSense right now.
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=203856

However there is something you can do to mitigate it to some extent, set:
sysctl net.isr.dispatch=deferred

You can add that as a system tunable in System > Advanced if it makes a significant difference.

Be aware that doing so may negatively impact some other things, ALTQ traffic shaping in particular.

Steve

bdaniel7

Thank you for the clarification.
I should've stated from the beginning that I'm on PPPoE.
I added the net.isr.dispatch setting, but I don't have any improvements in speed.

I am now evaluating which option is cheaper and faster, buying a different board, with other (Intel) cards and keeping pfSense, or moving to Linux.

bdaniel7

These are my settings, by the way:

hw.igb.fc_setting=0
hw.igb.rxd="4096"
hw.igb.txd="4096"
net.link.ifqmaxlen="8192"
hw.igb.max_interrupt_rate="64000"
hw.igb.rx_process_limit="-1"
hw.igb.tx_process_limit="-1"
hw.igb.0.fc=0
hw.igb.1.fc=0
net.isr.defaultqlimit=4096
net.isr.dispatch=deferred
net.pf.states_hashsize="2097152"
net.pf.source_nodes_hashsize="65536"
hw.igb.enable_msix: 1
hw.igb.enable_aim: 1

stephenw10

Hmm, you should see some improvement in speed with that setting. You may need to restart the ppp session or at least clear the firewall state. Or reboot if it's being applied by system tunables.

Steve

bblacey

@bdaniel7 said in Poor performance on igb driver:

These are my settings, by the way:

hw.igb.fc_setting=0
hw.igb.rxd="4096"
hw.igb.txd="4096"
net.link.ifqmaxlen="8192"
hw.igb.max_interrupt_rate="64000"
hw.igb.rx_process_limit="-1"
hw.igb.tx_process_limit="-1"
hw.igb.0.fc=0
hw.igb.1.fc=0
net.isr.defaultqlimit=4096
net.isr.dispatch=deferred
net.pf.states_hashsize="2097152"
net.pf.source_nodes_hashsize="65536"
hw.igb.enable_msix: 1
hw.igb.enable_aim: 1

I recently went through the process if identifying the performance culprit on the Intel NICs using a Lanner FW-7525A. It turns out, that for the igb driver, you want hw.igb.enable_msix=0 or hw.pci.enable_msix=0 to nudge the driver towards using msi interrupts over the less-performant MSIX interrupts (suggested here). This made a 4x difference on my system. It is also recommended to disable tso and lso on the igb drivers so include net.inet.tcp.tso=0 as well. Hope this helps.

stephenw10

Hmm, interesting. I wouldn't have expected msi to any better than msix.
What sort of figures did you see?

Steve

bblacey

@stephenw10 said in Poor performance on igb driver:

Hmm, interesting. I wouldn't have expected msi to any better than msix.
What sort of figures did you see?

Steve

Hmmm, I'm back to msix interrupts so that was a red herring. I'm able to fully saturate my 400/20 link (achieve 470/24) with both inbound and outbound firewall rules enabled. Here is my current config that seems to achieve this:

[2.4.4-RELEASE][root@firewall.home]/root: cat /boot/loader.conf 
kern.cam.boot_delay=10000
# Tune the igb driver
hw.igb.rx_process_limit=800  #100
hw.igb.rxd=4096  #default 1024
hw.igb.txd=4096  #default 1024
# Disable msix interrupts on igb driver either via hw.pci or the narrower hw.igb
#hw.pci.enable_msix=0   #default 1 (enabled, disable to nudge to msi interrupts)
#hw.igb.enable_msix=0
#net.inet.tcp.tso=0  #confirmed redundant with disable in GUI
#hw.igb.fc_setting=0
legal.intel_ipw.license_ack=1
legal.intel_iwi.license_ack=1
boot_multicons="YES"
boot_serial="YES"
console="comconsole,vidconsole"
comconsole_speed="115200"
autoboot_delay="3"
hw.usb.no_pf="1"

Basically, I'm using the defaults other than increasing the igb driver rx_process_limit, rxd and txd. I have disabled tso, lro and checksum offloading via the gui under System->Advanced->Networking (checked means disabled) and set kern.ipc.nmbclusters to 262144 under System->Advanced->Tunables.

Hardware:

CPU: Intel(R) Atom(TM) CPU  C2358  @ 1.74GHz (1750.04-MHz K8-class CPU)
  Origin="GenuineIntel"  Id=0x406d8  Family=0x6  Model=0x4d  Stepping=8
  Features=0xbfebfbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CLFLUSH,DTS,ACPI,MMX,FXSR,SSE,SSE2,SS,HTT,TM,PBE>
  Features2=0x43d8e3bf<SSE3,PCLMULQDQ,DTES64,MON,DS_CPL,VMX,EST,TM2,SSSE3,CX16,xTPR,PDCM,SSE4.1,SSE4.2,MOVBE,POPCNT,TSCDLT,AESNI,RDRAND>
  AMD Features=0x28100800<SYSCALL,NX,RDTSCP,LM>
  AMD Features2=0x101<LAHF,Prefetch>
  Structured Extended Features=0x2282<TSCADJ,SMEP,ERMS,NFPUSG>
  VT-x: PAT,HLT,MTF,PAUSE,EPT,UG,VPID
  TSC: P-state invariant, performance statistics

You might want to go back to the pfSense defaults and enable all networking offloading options are disabled (checked in the GUI), then tweak the igb driver elements as I did above test and then adjust key tunables such as kern.ipc.nmbclusters but more isn't necessarily better.

MarcoP

I've just noticed you are on PPPoE, would increasing MSS clamping on the interface or setting MTU at 1492 help.

stephenw10

You should put custom settings in /boot/loader.conf.local to avoid them being overwritten at upgrade. Create that file if it's not there.

Steve

Nonconformist

Hi @bdaniel7, any luck on achieving gigabit speeds after your tweaks? I’ve been running into the same issues as you with the same Qotom box.

Posted about it [here] (https://forum.netgate.com/topic/137196/slow-gigabit-download-on-a-quadcore-intel-celeron-j1900-2-41ghz), and then used the tweaks in this thread.

Still getting only about 730mbps on wired.

bdaniel7

@nonconformist
Hi, nope, I couldn't get any speed higher than 550 Mbps when I tried the tweaks.
Then I abandoned the subject, due to lack of time.

I will try the tweaks from the article you posted.

MarcoP

any dropped packets?

netstat -ihw 1

MarcoP

Since you have cores waiting, you could try to avoid locks when switching between them with:
net.isr.bindthreads="1"

Nonconformist

@marcop Couldn't check through the week so doing this over the weekend. Long story short, no dropped packets.

net.isr.bindthreads="1"

actually brought the download/upload speeds down to 680/800 from 740/934.

Reading more about this, it's beginning to look like achieving 1G download isn't possible with the igb0 driver with a PPPoE WAN connection.