Packet reordering and blocking problem at gigabit with 2.4 kernel

From: Hyunje Park (hjpark_at_zooin.net)
Date: 09/23/03

  • Next message: Thiago Rondon: "Re: cryptoloop hard lockup 2.6.0-test5-mm4" 
    To: <linux-kernel@vger.kernel.org>
Date:	Tue, 23 Sep 2003 22:27:14 +0900

    I am developing application level multicast router with Xeon processor.
    The system lost lots of packets and reordered the packets, and finally
    was blocked with heavy load.
    Some of packets are lost and the lost packet appeared after a few
    seconds from its normal sequence.

    The specification of the systems are:
    Intel SE7501VW (7501 chipset with 533MHz interface) or Tian S2723(7500
    chipset with 400MHz,)
    2.4 GHz or 3.0 GHz Xeon (two CPUs, one CPU without hyperthreading was
    tested)
    Intel 82544SX Gigabit card

    Software vesions experimented:
    Linux kernel 2.4.18, 2.4.20, 2.4.22
    E1000 drivers: 4.x, 5.1.x, 5.2.16 (with/without NAPI option)

    Every channel sends 6Mbps with 2048 bytes packet. The senders sends upto
    100~120 channel to the multicast
    relay and the relay forwards the channels to the receiver using unicast
    or multicast.

    The L3 switch, Alteon 780 cannot support IGMP snooping currently. It
    means the multicast traffics may be bounced
    to the multicast relay as broadcast. Every channel are generating about
    800 pps: 100 channel will generate 80Kpps.

    I tried to experiment a lot of diffenent combinations with software,
    hardware, or configuration.
    I heard that SMP machines has inherant reordering issues, so I turned
    off SMP, but no improvement.
    I also tried to turn on NAPI, but no improvment.
    I tried to Interrupt mitigation by setting (RX, TX)Intdelay between 100
    and 500. The performance becomes better, but the
    reordering problems are not discarded.
    With the faster CPU, there are no differences either.

    I don't know what else I have to try to fix the bugs.
    Is it inherant problems with current version of Linux kernel?
    Or is there a bug of E1000 driver?
    Am I giving too much stress to the system?
    I will test with Sysconnect gigabit ethernet card tomorrow, but I am not
    sure if it solve the problem.

    Pease CC me with any suggestions or comments since I am not subscribed
    to the list.
    Thanks in advance.

    Hyunje Park

    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

  • Next message: Thiago Rondon: "Re: cryptoloop hard lockup 2.6.0-test5-mm4"

    Relevant Pages

    • Kismet and wi0
      ... between Kismet and the Orinoco Gold card. ... MPSAFE network stack disabled, ... # to set the channel the interface listens on. ... # filter_dump - Packets filtered at the dump level are tracked, displayed, ...
      (freebsd-questions)
    • Re: UDP multicast packets not seen on listening interface in BETA5
      ... > I'm having a bit of trouble with a program I wrote to listen for ... > the multicast packets, ... > Also I verified the program is indeed listening with sockstat: ...
      (freebsd-current)
    • Re: Spoofing multicast addresses
      ... }>> connections and it wasteful to do unnecessary checking on these packets. ... reducing the overhead of multicast packets wouldn't make ... doing the PCB lookup before the source address check, then the attacker ... able to keep things in the non-optimal state by varying the packet stream. ...
      (FreeBSD-Security)
    • Re: Multiple multicast apps
      ... Looks like an error in the multicast filter setup for your network ... Rather than having to take all multicast packets off the ... It's a bug and I would assume it has been fixed later in the 4.x ...
      (comp.unix.bsd.freebsd.misc)
    • multicast routing issue?
      ... I have an application which should join a multicast channel at beginning ... But my problem is that if I join the channel from the VPN router, I can receive the traffic ok. ... But the request to join the channel from my program didn't get to the router and forward alone. ...
      (comp.unix.admin)