PFSense in cloud - low throughput



  • Have PFSense as a front end for my CouchDB server in the cloud.  Noticed on replication from development server to production at CloudSigma throughput was very slow.  Took 2hrs to replicate 215mb.

    Basic setup: public ip <=> PFSense <=> CloudSigma VLan <=> FreeBSD 10.2 <=> jail <=> CouchDB

    Worked with CloudSigma tech to isolate the issue.  Haven't found the cause or a fix.  Tech thinks it is something in PFSense.  Figures the VLan is good to go.

    Ran iperf3 tests between
    dev server and cloud PFSense: ~40Mbits/sec
    cloud PFSense <=> cloud FreeBSD server: ~190Mbits/sec
    dev server <=> PFSense <=>  FreeBSD server: ~400KBits/sec
    dev server <=> cloud FreeBSD server: ~40Mbits/sec (direct)

    No packages,
    turned off most services, 
    LAN is open out,
    WAN rules & NAT checked all very simple traffic on port to CouchDB server jail
    no difference in speed between jail and jail host
    latest PFSense 2.2.5 AMD64
    no load to speak of
    tried both vtnet and em0 nics

    Any ideas?  Am at a loss for what the issue could be.

    Peter



  • Solution:

    Use Intel Pro/1000 nic on both WAN and LAN.

    Had tried on LAN, but not both.  CloudSigma techs are testing why the difference, but solution gives solid throughput.



  • Could you give us a lowdown on your new solid iperf numbers. I would like to see what throughput your achieving.



  • 30 MBit/sec from FreeBSD server to remote office.  The limit is likely the throughput into our office.

    83Mbits/sec from FreeBSD server to PFSense gateway through VLAN

    Considerable better than 400Kbits/sec with vtnet0, but still considerable less than 1Gb nominal of the NIC. vtnet nics are rated at 10Gb by CloudSigma.

    Peter



  • What was your System Activity like during those super slow transfers? Where you getting a large number of interrupts and high kernel CPU? Even when I'm doing 70k-pps, I'm still around 150 interrupts per second. Intel i350-T2 is sweet.


Log in to reply