Stenographer is a full-packet-capture utility for buffering packets to disk for intrusion detection and incident response purposes. It provides a high-performance implementation of NIC-to-disk packet writing, handles deleting those files as disk fills up, and provides methods for reading back specific sets of packets quickly and easily.

It is designed to:

• Write packets to disk, very quickly (~10Gbps on multi-core, multi-disk machines)
• Store as much history as it can (managing disk usage, storing longer durations when traffic slows, then deleting the oldest packets when it hits disk limits)
• Read a very small percentage (<1%) of packets from disk based on analyst needs

It is NOT designed for:

• Complex packet processing (TCP stream reassembly, etc)

• It’s fast because it doesn’t do this.  Even with the very minimal, single-pass processing of packets we do, processing ~1Gbps for indexing alone can take >75% of a single core.

• Processing the data by reading it back from disk also doesn’t work:  see next bullet point.

• Reading back large amounts of packets (> 1% of packets written)

• The key concept here is that disk reads compete with disk writes… you can write at 90% of disk speed, but that only gives you 10% of your disk’s time for reading.  Also, we’re writing highly sequential data, which disks are very good at doing quickly, and generally reading back sparse data with lots of seeks, which disks do slowly.

For further reading, check out DESIGN.md for a discussion of stenographer's design, or read INSTALL.md for how to install stenographer on a machine.

A user requests packets from stenographer by specifying them with a very simple query language. This language is a simple subset of BPF, and includes the primitives:

host 8.8.8.8          # Single IP address (hostnames not allowed)
net 1.0.0.0/8         # Network with CIDR
net 1.0.0.0 mask 255.255.255.0  # Network with mask
port 80               # Port number (UDP or TCP)
ip proto 6            # IP protocol number 6
icmp                  # equivalent to 'ip proto 1'
tcp                   # equivalent to 'ip proto 6'
udp                   # equivalent to 'ip proto 17'

# Stenographer-specific time additions:
before 2012-11-03T11:05:00Z      # Packets before a specific time (UTC)
after 2012-11-03T11:05:00-07:00  # Packets after a specific time (with TZ)
before 45m ago        # Packets before a relative time
after 3h ago         # Packets after a relative time

NOTE: Relative times must be measured in integer values of hours or minutes as demonstrated above.

Primitives can be combined with and/&& and with or/||, which have equal precendence and evaluate left-to-right. Parens can also be used to group.

(udp and port 514) or (tcp and port 8080)

The stenoread command line script automates pulling packets from Stenographer and presenting them in a usable format to analysts. It requests raw packets from stenographer, then runs them through tcpdump to provide a more full-featured formatting/filtering experience. The first argument to stenoread is a stenographer query (see 'Query Language' above). All other arguments are passed to tcpdump. For example:

# Request all packets from IP 1.2.3.4 port 6543, then do extra filtering by
# TCP flag, which typical stenographer does not support.
$ stenoread 'host 1.2.3.4 and port 6543' 'tcp[tcpflags] & tcp-push != 0'

# Request packets on port 8765, disabling IP resolution (-n) and showing
# link-level headers (-e) when printing them out.
$ stenoread 'port 8765' -n -e

# Request packets for any IPs in the range 1.1.1.0-1.1.1.255, writing them
# out to a local PCAP file so they can be opened in Wireshark.
$ stenoread 'net 1.1.1.0/24' -w /tmp/output_for_wireshark.pcap

To download the source code, install Go locally, then run:

$ go get github.com/google/stenographer

Go will handle downloading and installing all Go libraries that stenographer depends on. To build stenotype, go into the stenotype directory and run make. You may need to install the following Ubuntu packages (or their equivalents on other Linux distros):

• libaio-dev
• libleveldb-dev
• libsnappy-dev
• g++
• libcap2-bin
• libseccomp-dev

This is not an official Google product (experimental or otherwise), it is just code that happens to be owned by Google.

This code is not intended (or used) to watch Google's users. Its purpose is to increase security on our networks by augmenting our internal monitoring capabilities.