NAME
fprobe - a NetFlow probe
SYNOPSIS
fprobe [options] remote:port[/[local][/type]] ...
DESCRIPTION
fprobe - libpcap-based tool that collect network traffic data and emit
it as NetFlow flows towards the specified collector.
OPTIONS
-h Display short help
-p Dont put the interface into promiscuous mode. Note that even
if this option is used, the interface might be in promiscuous
mode for some other reason.
-i <interface>
Listen on interface. If unspecified, fprobe will use result of
pcap_lookupdev() function. On Linux systems with 2.2 or later
kernels, an interface argument of ‘any’ can be used to capture
packets from all interfaces. Note that captures on the ‘any’
device will not be done in promiscuous mode.
You may use ‘-’ as interface name to process files produced by
tcpdump with -w flag. fprobe will read data from stdin.
-f <expression>
Filter expression selects which packets will be captured. If no
expression is given, all packets on the net will be captured.
Otherwise, only packets for which expression is ‘true’ will be
captured.
fprobe use silly IP-packet detection method, so it is bad idea
to leave the filter empty. For general use ‘ip’ (-fip) is good
filter expression.
Read tcpdump manual for detailed expression syntax.
-s <seconds>
How often scan for expired flows. [default=5]
-g <seconds>
Fragmented flow lifetime. [default=30]
-d <seconds>
Idle flow lifetime (inactive timer). [default=60]
-e <seconds>
Active flow lifetime (active timer). [default=300]
-n <version>
NetFlow version for use (1, 5, 7). [default=5]
-a <address>
Use address as source for NetFlow flow.
-x <inputID>[:<outputID>]
Workaround for SNMP interfaces indexes. [default=0]
The second parameter may be omitted - in this case its value
will be equal to the first.
See BUGS section.
-b <flows>
Memory bulk size. [default=200 or 10000]
Note that maximum and default values depends on compiling
options (--with-membulk parameter).
-m <kilobytes>
Memory limit for flows cache (0=no limit). [default=0]
-q <flows>
Pending queue length. [default=100]
Each captured packet at first puts into special buffer called
‘pending queue’. Purpose of this buffer is to separate most
time-critical packet capture thread from other.
-B <kilobytes>
Kernel capture buffer size (0=don’t change). [default=0]
Increase kernel capture buffer size is most adequate way to
prevent packets loss. Unfortunately, at present there is no
straight way to set the buffer size throught libpcap, so this
option is a hack. Moreover, now this hack take effect only on
socket()-based capture mechanisms: it mean that it work on Linux
and don’t work on BSD systems with their bpf().
Note that maximum allowed size of the buffer in Linux limited
and generally relatively small, so it should need to change the
maximum: sysctl -w net/core/rmem_max=4194304
-r <priority>
Real-time priority (0=disabled). [default=0]
If parameter greater then zero fprobe will use real-time
scheduling policy to prevent packets loss. Note that possible
values for this option depends on operating system.
-t <B:N>
Emitting rate limit (0:0=no limit). [default=0:0]
Produce N nanosecond delay after each B bytes sent. This option
may be useful with slow interfaces and slow collectors. Note
that the suspension time may be longer than requested because
the argument value is rounded up to an integer multiple of the
sleep resolution (it depends on operating system and hardware)
or because of the scheduling of other activity by the system.
See BUGS section.
-S <bytes>
Snaplen (0=whole packet). [default=256]
Number of bytes to capture from packet on wire.
-K <bytes>
Link layer header size. By default fprobe take this information
from libpcap, but sometimes obtained size unsuitable for our
purpose. It occurs, for example, on trunk interfaces in VLAN
enviroment, where link layer header contain additional VLAN
header.
See EXAMPLES section.
-k Don’t exclude link layer header from packet size. By default
fprobe counts only IP-part of packet.
-c <directory>
Directory to chroot to.
-u <user>
User to run as.
-v <level>
Maximum displayed log level. (0=EMERG, 1=ALERT, 2=CRIT, 3=ERR,
4=WARNING, 5=NOTICE, 6=INFO, 7=DEBUG) [default=6]
-l <[dst][:id]>
Log destination (0=none, 1=syslog, 2=stdout, 3=both) and
log/pidfile identifier. [default=1]
This option allows to select opportune log destination and
process identifier. The identifier helps to distinguish pidfile
and logs of one fprobe process from other.
Note that if log destination contains ‘stdout’ (equal 2 or 3)
fprobe will run in foreground.
remote:port/local/type
Parameters remote and port are respectively define address and
port of the NetFlow collector.
The local parameter allows binding certain local IP address with
specified collector. If the parameter is omitted the value (if
any) of -a option will be used.
The type parameter determines emitting behavior. It may be ‘m’
for mirroring (by default) and ‘r’ for collectors round-robin
rotating.
You may specify multiple collectors.
EXAMPLES
Web traffic trivial capturing:
fprobe -ippp0 -f"tcp&&port 80" localhost:2055
Capturing from trunk interface:
fprobe -ieth0 -f"vlan&&ip" -K18 localhost:2055
Reasonable configuration to run under heavy load:
fprobe -fip -B4096 -r2 -q10000 -t10000:10000000 localhost:2055
Send packets to collector at 10.1.1.1:2055 and distribute them between
collectors at 10.1.1.2:2055 and at 10.1.1.3:2055 on a round-robin
basis:
fprobe 10.1.1.1:2055 10.1.1.2:2055//r 10.1.1.3:2055//r
BUGS
SNMP interfaces indexes and packet direction.
Unfortunately libpcap don’t provide any routing-related information
about captured packet, therefore there are no straight ways to
determine and distinguish input and output interfaces. However -x
option at least can tell that flow was passed through the certain
interface. Also you may launch several instances of the program with
tricky set of filters to mark out each possible packet direction:
fprobe -x1:2 -ieth1 -f"ip&&dst net 10.2" localhost:2055
fprobe -x2:1 -ieth2 -f"ip&&dst net 10.1" localhost:2055
Slow interfaces and slow collectors.
There are may be problems with slow interfaces and slow collectors. It
effects as emitted packets loss. On the one hand silent non-blocking
sendto() implementation can’t guarantee that packet was really sent to
collector - it may be dropped by kernel due to outgoing buffer shortage
(slow interface’s problem) and on the other hand packet may be dropped
on collector’s machine due the similar reason - incoming buffer
shortage (slow collector’s problem).
Use -t option as workaround for this issue.
SEE ALSO
tcpdump pcap(3)
http://www.cisco.com/go/netflow