This paper is about two distinct but related things:
By publishing this paper, I hope to:
At the University of Washington, we have a large network with many hundreds of subnets and close to 120,000 IP devices on our campus network. Sometimes it is necessary to look at network traffic to diagnose problems. Recently, I began a project to allow us to capture subnet-level traffic remotely (without having to physically connect remotely) to make life easier for our Security and Network Operations groups and to help diagnose problems more efficiently.
Our Cisco 7600 routers have the ability to create a limited number of "Encapsulated Remote SPAN ports" (ERSPAN ports) which are similar to mirrored switch ports except the router "GRE" encapsulates the packets and sends them to an arbitrary IP address. (GRE is in quotes because the Cisco GRE header is larger than the standard GRE header (it is 50 bytes) so Linux and/or unmodified tcpdump can not correctly decapsulate it).
Because the router will send the "GRE" encapsulated packets without any established state or confirmation on the receiver (as if sending UDP), I don't need to establish a tunnel on Linux to receive the packets. I initially wrote a tiny (30-line) proof-of-concept decapsulator in C which could postprocess a tcpdump capture like this:
tcpdump -i eth1 -s0 -w - proto gre | conv > pcapfile or tcpdump -i eth1 -s0 -w - proto gre | conv | tcpdump -s0 -r - -w pcapfile ...
My initial measurements indicated that the percentage of dropped packets and CPU overhead of writing through the conversion program and then to disk were not significantly higher than writing directly to disk so I thought this was a reasonable plan. On my old desktop workstation, a 3.2GHz P4 Dell Optiplex 270 with slow 32-bit PCI bus and a built-in 10/100/1000 Intel 82540EM NIC) running Fedora Core 6 Linux (2.6.19 kernel, ethtool -G eth0 rx 4096), I could capture and save close to 180Mb/s of iperf traffic with about 1% packet loss so it seemed worth pursuing. Partly to facilite this and partly for unrelated reasons, I bought a newer/faster office PC.
To my surprise, my new office PC (a Dell Precision 690 with 2.66 GHz
quad-core Xeon x5355, PCI-Express-based Intel Pro-1000-PT NIC, faster
RAM and SATA disks) running the same (Fedora Core 6) OS, initially
dropped more packets than my old P4 system did, even though each of the 4
CPU cores does about 70% more than my old P4 system (according to my
benchmarks). I spent a long time trying to tune the OS by changing
various parameters in
/sys, trying to tune the e1000 NIC driver's tunable
parameters and fiddling with scheduling priority and processor affinity
(for processes, daemons and interrupts). Although the number of
combinations and permutations of things to change was high, I gradually
made enough progress that I continued down this path for far too long
before discovering the right path.
Two things puzzled me: "xosview" (a system load visualization tool) always showed plenty of idle resources when packets were dropped and writing packets to disk seemed to have a disproportionate impact on packet loss, especially when the system buffer cache was full.
It eventually occurred to me to try to decouple disk writing from packet reading. I tried piping the output of the capturing tcpdump program into an old (circa 1990) tape buffering program (written by Lee McLoughlin) which ran as two processes with a small shared-memory ring buffer. Remarkably, piping the output through McLoughlin's buffer program caused tcpdump to drop fewer packets. Piping through "dd" with any write size and/or buffer size or through "cat" did not provide any improvement. My best guess as to why McLoughlin's buffer helped is that even though the select(2) system call says writes to disk never block, they effectively do. When the writes block, tcpdump can't read packets from the kernel quickly enough to prevent the NIC's buffer from overflowing.
A quick look at the code in McLoughlin's buffer program convinced me I would do better starting from scratch so I wrote a simple multi-threaded ring-buffer program (which became Gulp). For both simplicity and efficiency under load, I designed it to be completely lock-free. The multi-threaded ring buffer worked remarkably well and considerably increased the rate at which I could capture without loss but, at higher packet rates, it still dropped packets--especially while writing to disk.
I emailed Luca Deri, the author of Linux's PF_RING NIC driver, and he (correctly) suggested that it would be easy to incorporate the packet capture into the ring buffer program itself (which I did). This ultimately was a good idea but initially didn't seem to help much. Eventually I figured out why: the Linux scheduler sometimes scheduled both my reader and writer threads on the same CPU/core which caused them to run alternately instead of simultaneously. When they ran alternately, the packet reader was again starved of CPU cycles and packet loss occurred. The solution was simply to explicitly assign the reader and writer threads to different CPU/cores and to increase the scheduling priority of the packet reading thread. These two changes improved performance so dramatically that dropping any packets on a gigabit capture, written entirely to disk, is now a rare occurrence and many of the system performance tuning hacks I resorted to earlier have been backed out. (I now suspect they mostly helped by indirectly influencing process scheduling and cpu affinity--something I now control directly--however on systems with more than two CPU cores, the inter-core-benchmark I developed may still be helpful to determine which cores work most efficiently together).
On some systems, increasing the default
size of receive socket buffers also helps:
echo 4194304 > /proc/sys/net/core/rmem_max;
echo 4194304 > /proc/sys/net/core/rmem_default
Our (pilot) production system for gigabit remote packet capture is a Dell PowerEdge model 860 with a single Intel Core2Duo CPU (x3070) at 2.66 GHz (hyperthreading disabled) running RedHat Enterprise Linux 5 (RHEL5 2.6.18 kernel). It has 2GB RAM, two WD2500JS 250GB SATA drives in a striped ext2 logical volume (essentially software RAID 0 using LVM) and an Intel Pro1000 PT network interface (NIC) for packet capture. (The builtin BCM5721 Broadcom NICs are unable to capture the slightly jumbo frames required for Cisco ERSPAN--they may work for non-jumbo packet capture but I haven't tested them. The Intel NIC does consume a PCI-e slot but costs only about $40.)
A 2-minute capture of as much iperf data as I can generate into a 1Gb ERSPAN port (before the ERSPAN link saturates and the router starts dropping packets) results in a nearly 14GB pcap file usually with no packets dropped by Linux. The packet rate for that traffic is about 96k pps avg. The router port sending the ERSPAN traffic was nearly saturated (900+Mb/s) and the sum of the average iperf throughputs was 818-897Mb/s (but unlike ethernet, I believe iperf reports only payload bits counted in 1024^2 millions so this translates to 857-940Mb/s in decimal/ethernet millions not counting packet headers). Telling iperf to use smaller packets, I was able to capture all packets at 170k pps avg but I could only 2/3 saturate the gigabit network using iperf and small packets with the hardware at my disposal.
A subsequent test using a "SmartBits" packet generator to roughly 84% saturate the net with 300-byte packets indicates I can capture and write to disk 330k pps without dropping any packets. Interestingly the failure mode at higher packet rates is that there is insufficient CPU capacity left to empty Gulp's ring buffer as fast as it fills. Gulp did not start dropping packets until its ring buffer eventually filled. This demonstrates that Linux can be very successful at capturing packets at high speed and delivering them to user processes as long as the reading process can read them from the kernel fast enough that the NIC-driver's relatively small ring buffer does not overflow. At very high packet rates, even though the e1000 NIC driver does interrupt aggregation, xosview indicated that much of the CPU was consumed with "hard" and "soft" interrupt processing.
In summary, I believe as long as the average packet size is 300 or more, our system should be able to capture and write to disk every packet it receives from a gigabit ethernet. The larger the average packet size, the more CPU headroom is available and the more certain is capturing every packet.
I should mention that I have been using Shawn Ostermann's "tcptrace" program to confirm that when tcpdump or Gulp reports that the kernel dropped no packets, this is indeed true. Likewise, when the tools report the kernel dropped some packets, tcptrace agrees. This means I have complete confidence in my claims above for capturing iperf data without loss. Although the SmartBits did not generate TCP traffic, it offered counts of how many packets it sent which agree with what was captured.
0) the Gulp manpage.pdf or Gulp manpage.html (converted with bold2html). 1) helping tcpdump drop fewer packets when writing to disk: (gulp -c can be used in any pipeline as it does no data interpretation) tcpdump -i eth1 -w - ... | gulp -c > pcapfile or if you have more than 2 CPUs, run tcpdump and gulp on different ones: taskset -c 2 tcpdump -i eth1 -w - ... | gulp -c > pcapfile (gulp uses CPUs #0,1 so taskset runs tcpdump on #2 to reduce interference) 2) a similar but more efficient capture using Gulp's native capture ability: gulp -i eth1 -f "..." > pcapfile 3) capture and GRE-decapsulate an ERSPAN feed and save the result to disk: gulp -i eth1 -d > pcapfile 4) capture, decapsulate and then filter with tcpdump before saving: gulp -i eth1 -d | tcpdump -r - -s0 -w pcapfile ... or if you have more than 2 CPUs, run tcpdump and gulp on different ones: gulp -i eth1 -d | taskset -c 2 tcpdump -r - -s0 -w pcapfile ... 5) capture everything to disk; then decapsulate offline: gulp -i eth1 > pcapfile1; gulp -d -i - < pcapfile1 > pcapfile2 6) capture, decapsulate and filter with ngrep: gulp -i eth1 -d | ngrep -I - -O pcapfile regex ... 7) capture, decapsulate and feed into ntop: gulp -i eth1 -d | ntop -f /dev/stdin -m a.b.c.d/x ... or mkfifo pipe; chmod 644 pipe; gulp -i eth1 -d > pipe & ntop -u ntop -f pipe -m a.b.c.d/x ... 8) capture, decapsulate and feed into wireshark: gulp -i eth1 -d | /usr/sbin/wireshark -i - -k 9) capture to 1000MB files, keeping just the most recent 10 (files): gulp -i eth1 -C 10 -W 10 -o pcapdir or with help from tcpdump: gulp -i eth1 | taskset -c 2 tcpdump -r- -C 1000 -W 10 -w pcapname
Normally if one is interested in capturing only a subset of the traffic on an interface, the pcap library can filter out the uninteresting packets in the kernel (as early as possible) to avoid the overhead of copying them into userspace and then discarding them.
Because neither the Linux GRE tunnel mechanism, i.e.:
# modprobe ip_gre # ip tunnel add gre1 local x.y.78.60 remote x.y.78.4 mode gre # ifconfig gre1 up # tcpdump -i gre1
nor the pcap code seems to be capable of decapsulating GRE packets with a non-standard header length (50 bytes in this case) and then applying normal pcap filters to what remains, I can do no in-kernel filtering on the contents of the ERSPAN packets--they must all be copied to userspace, decapsulated and then filtered again by tcpdump (wireshark or equivalent) as per examples #4-6 above.
Extensions to either the pcap code or the GRE tunnel mechanism should be able to add the ability to capture a subset of packets more efficiently by filtering them out in the kernel. I have not measured the overhead of "ip tunnel" but I presume doing this in the pcap code would be simplest and most efficient.
Perhaps select(2) should not always say a descriptor to an open file on disk will not block for write(2) or alternatively, perhaps the writes can be made faster so they agree with select(2) and don't block.
I think "
struct pcap_pkthdr" in
should be re-defined to be independent of
sizeof(long). In pcap
struct pcap_pkthdr precedes every packet.
Unfortunately, the size of
struct pcap_pkthdr (which contains
struct timeval) depends upon
This makes pcap files from 64-bit linux systems incompatible with those from
32-bit systems. Apparently as a workaround, some 64-bit linux distributions
are providing tcpdump and wireshark binaries which read/write 32-bit compatible
pcap files (which makes Gulp's pcap output appear to be corrupt).
(To build Gulp on 64-bit linux systems so that it reads/writes 32-bit compatible pcap files, try installng the 32-bit (i386) "libpcap-devel" package and making Gulp with "-m32" added to CFLAGS.)
To my surprise, I learned after completing this work that Luca Deri's PF_RING patch is NOT already incorporated in the standard Linux kernel (as I mistakenly thought) and the packet "ring buffer" that "ethtool" adjusts is something different. Though this misunderstanding is somewhat embarrassing to me, it seems likely that the benefits of Gulp and PF_RING will be cumulative and since my next obvious goal is 10Gb I look forward to confirming that.
Email -- corey @ u.washington.edu
Web -- http://staff.washington.edu/corey/
Date -- Tue Mar 18 14:14:09 PDT 2008