6.x, 4.x ipfw/dummynet pf/altq - network performance issues

Tue Feb 13 10:21:01 UTC 2007

  Clockrate is based off of my device_polling setup, which is configured 
to 4000. burst_max has a hard limit, can't go higher than it already is 
at 1000

  Could I get an explanation as to what the queue and isr sysctl values 
are actually doing? I'll be able to run some more basic tests tomorrow 
to see some results, but want to wrap my head around what's actually 
logically meant to be happening based on adjustments, etc. [I suspect 
this'll do nothing for the UDP issue, but at least I might be able to 
pipe some TCP traffic]

garcol at postino.it wrote:
> Hi,
>        I think you can try to check/tuning this sysctl variables and the isr 
> related variables:
>
> net.inet.ip.intr_queue_maxlen
> net.inet.ip.intr_queue_drops
>
> net.isr.enable   ... try to set
> net.isr.directed
> net.isr.queued 
> net.isr.drop
>
> and polling configuration:
>
> kern.clockrate
>
> kern.polling.burst_max
> ....  increase for high rate of small packets on GE 
> ....
>
> Alessandro
>
>
>
>
>   
>> Date: Wed, 07 Feb 2007 01:37:00 -0800
>> From: Justin Robertson <justin at sk1llz.net>
>> Subject: 6.x, 4.x ipfw/dummynet pf/altq - network performance issues
>> To: freebsd-performance at freebsd.org
>> Message-ID: <45C99DBC.1050402 at sk1llz.net>
>> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
>>
>>   It was suggested I post this to freebsd-performance, it's already in 
>> questions, isp, and net.
>>
>> I've been running some tests with using FreeBSD to filter and rate limit 
>> traffic. My first thoughts were to goto the latest stable release, which 
>> was 6.1 at the time. I've since done the same test under 6.2 and haven't 
>> seen any difference. I later migrated to running 4.11 to get away from 
>> these issues, but have discovered others.
>>
>> I've tested on an AMD 3200+ system with dual Intel 1000 series NICs, an 
>> AMD Opteron 165 with the same, and a Xeon 2.8 with the same. I've used 
>> both stock and intel drivers.
>>
>> 6.x;
>> Normal traffic isn't a problem. The second you get into the realm of 
>> abusive traffic, such a DoS/DDoS (over 100mbps) UDP floods the machine 
>> falls over. Little packets with ip lengths of 28-29 bytes seem to do the 
>> most damage. I've tried playing with various sysctl values and have seen 
>> no difference at all. By "falls over" I mean "stops sending all traffic 
>> in any direction". TCP syn packets have the same effect, tho not quite 
>> as rapidly (200~230mbps). I then tried moving filtering off to a 
>> transparent bridge. This improved the situation somewhat, but an extra 
>> 30-40mbps of UDP data and it would ultimately crumble. Overall the 
>> machine would be able to move between 300k-600k PPS before becoming a 
>> cripple, depending on packet length, protocol, and any flags. Without a 
>> specific pf or ipfw rule to deal with a packet the box would fall over, 
>> with specific block rules it would manage an extra 30-40mbps and then 
>> fall over.
>>
>> 4.11;
>> Again, normal traffic isn't a problem. When routing & filtering on the 
>> same system some of the problems found in 6.x are still apparent, but to 
>> a lesser degree. Splitting the task into a transparent filtering bridge 
>> with a separate routing box appears to clear it up entirely. UDP floods 
>> are much better handled - an ipfw block rule for the packet type and the 
>> machine responds as if there were no flood at all (until total bandwidth 
>> saturation or PPS limits of the hardware, which in this case was around 
>> 950Mbps). TCP syn attacks are also better handled, again a block rule 
>> makes it seem as if there were no attack at all. The system also appears 
>> to be able to move 800-900k PPS of any one protocol at a time. However, 
>> the second you try and queue abusive traffic the machine will fall over. 
>> Inbound floods appear to cause ALL inbound traffic to lag horrifically 
>> (while rate limiting/piping), which inherently causes a lot of outbound 
>> loss due to broken TCP. Now, I'm not sure if this is something to do 
>> with dummynet being horribly inefficient, or if there's some sysctl 
>> value to deal with inbound that I'm missing.
>>
>> I suppose my concerns are two-fold. Why is 6.x collapsing under traffic 
>> that 4.11 could easily block and run merrily along with, and is there a 
>> queueing mechanism in place that doesn't tie up the box so much on 
>> inbound flows that it ignores all other relevant traffic?
>>
>> (as a note, all tests were done with device polling enabled. Without it 
>> systems fall over pretty quickly. I also tried tests using 3com cards 
>> and had the same results)
>>
>>
>> In the event anybody is looking for basic errors, device polling is 
>> enabled and running at 4000 hz, which has proved to net the highest 
>> thruput in PPS. ADAPTIVE_GIANT is on (tests resulted in better pps 
>> thruput), all the other monitoring features are disabled, and here are 
>> my sysctl modifications related to networking (if there's something 
>> glaring let me know!);
>>
>> kern.polling.enable=1
>> kern.polling.burst_max=1000
>> kern.polling.each_burst=80
>> kern.polling.idle_poll=1
>> kern.polling.user_frac=20
>> kern.polling.reg_frac=50
>> net.inet.tcp.recvspace=262144
>> net.inet.tcp.sendspace=262144
>> kern.ipc.maxsockbuf=1048576
>> net.inet.tcp.always_keepalive=1
>> net.inet.ip.portrange.first=10000
>> kern.ipc.somaxconn=65535
>> net.inet.tcp.blackhole=2
>> net.inet.udp.blackhole=1
>> net.inet.icmp.icmplim=30
>> net.inet.ip.forwarding=1
>> net.inet.ip.portrange.randomized=0
>> net.inet.udp.checksum=0
>> net.inet.udp.recvspace=8192    (I've tried large and small, thinking 
>> perhaps I was fulling up buffers on udp floods and then causing it to 
>> drop tcp, there appears to be no difference)
>> net.inet.ip.intr_queue_maxlen=512
>> net.inet.tcp.delayed_ack=0
>> net.inet.tcp.rfc1323=1
>> net.inet.tcp.newreno=0 (I'd try this, but, the biggest problem is still 
>> with UDP, and I'd prefer something compatible with everything for now)
>> net.inet.tcp.delacktime=10
>> net.inet.tcp.msl=2500
>> net.inet.ip.rtmaxcache=1024
>> net.inet.raw.recvspace=262144
>> net.inet.ip.dummynet.hash_size=512
>> net.inet.ip.fw.dyn_ack_lifetime=30
>> net.inet.ip.fw.dyn_syn_lifetime=10
>> net.inet.ip.fw.dyn_fin_lifetime=10
>> net.inet.ip.fw.dyn_max=16192
>> net.link.ether.bridge.enable=0 (or 1 on when setup to bridge, obviously)
>> net.inet.ip.fastforwarding=1
>>
>> It has also been pointed out that using net.link.ether.ipfw=1 should 
>> negate the need for a transparent box, however the performance disparity 
>> between 6.x and 4.11 remains.
>>
>>     
> _______________________________________________
> freebsd-performance at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
>
>   