Re: UDP packet handling weird behaviour of various operating systems

From: Adrian Chadd (adrianat_private)
Date: Fri Jul 27 2001 - 03:30:20 PDT

  • Next message: Radu-Adrian Feurdean: "Re: UDP packet handling weird behaviour of various operating systems"

    On Thu, Jul 26, 2001, Stefan Laudat wrote:
    > > Most UDP packets should be firewalled from the Internet.
    > 
    > Agree.
    > 
    > > This is only really useful if someone has access to the local network. Is
    > > Linux/UP actually *locking* or just temporarily unresponsive? Also, it is
    > > invalid to compare Windows ME running on $3000 hardware with Linux/*BSD
    > > running on an old Pentium. Are you running all of this on the same
    > > hardware? Obviously faster hardware is going to be affected less by a UDP
    > > flood. How about the network cards?
    > 
    > Identical network cards for Win2k, Linux SMP and OpemBSD processor (Intel
    > Pro 100). Linux was run on dual p3/1Ghz(SMP), Pentium2/400Mhz and P3/800Mhz
    > (UP). Windows 2000 was run on p3/1Ghz UP. I've made tests with same results
    > against Linux UP boxes running on Celeron/600 with 3com Vortex and realtek
    > 8139 NICs. I've outlined that the result is the same no matter if you hit
    > via 1Gbit or 100Mbit. 
    
    Guys, guys.
    
    The realtek cards suck. If you don't believe me, try reading the device
    driver code for them in FreeBSD. Bill Paul slightly rips into their
    lame design. I use a couple at home in my doze machines because they
    were lying about. Getting 100mbit is painful - I don't use the
    top-line hardware in the doze machines.
    
    > > I bet a Sun E10K with lots of NICs could flood the Sun UE3500 with lots of
    > > NICs, but that probably doesn't mean that the Solaris 8 network stack is
    > > better than the Solaris 8 network stack; it's because the E10K is faster.
    > 
    > well then someone will clear all this stuff for me.
    > 
    
    When you're seeing the PC lockup, run vmstat 1 on it.
    See how many interrupts/context switches are happening a second.
    I bet the INT levels are stupidly high.
    
    Case in point: When trying out squid on a pair of IDE disks hooked up
    to a linux-2.2 box, I noted that it was crawling. after running
    vmstat for a while, it was obvious that the box was handling an absolutely
    *stupid* amount of interrupts per second. Turning on DMA fixed that.
    
    Now, Gige. Id on't remember the details of the original post, but
    if you've got a gige card in a Win server, I'm betting that the
    basic TCP/UDP processing is occuring on the card, not on the box.
    Depending how much work went into the driver (read: I bet its more
    than the state of the gige drivers under free unices) they might
    even be generating the connection refused replies *on the card*.
    
    
    
    
    Adrian
    
    -- 
    Adrian Chadd			Yeah, for me its (XML) like the movie Titanic.
    <adrianat_private>	  Everybody loves it.
    				    I want to be different, so I hate it.
    					--Duane Wessels
    



    This archive was generated by hypermail 2b30 : Fri Jul 27 2001 - 09:12:42 PDT