Engine Comparison

Being a comprehensive packet processor costs speed: pcapkit’s own engine is slower than every third-party engine below except PyShark, as Test Results measures. engine= selects the engine:

engine=	Engine
'default' or 'pcapkit'	pcapkit’s own parser, the only one with no third-party requirement
'scapy'	Scapy
'dpkt'	DPKT
'pyshark'	PyShark
'pypcap'	PyPCAP
'pcap_ct'	pcap-ct
'pypcapfile'	PyPCAPFile

PyPCAP and pcap-ct are independent distributions of the same libpcap(3) interface, and both install a top-level pcap module: upstream PyPCAP stops at Python 3.11, pcap-ct covers 3.10 and newer. Install exactly one of them – with both present, pcap-ct wins the import and the other becomes unselectable.

Every third-party engine supports less than the default one. PyPCAP and PCAP_CT perform no protocol dissection at all, so they offer neither reassembly nor flow tracing, and PyPCAPFile has no IPv6 decoder, so IPv6 reassembly is unavailable. Each gap is announced with a warning or an exception, except for these, which pass without a warning:

• PyPCAPFile leaves TCP behind a VLAN tag, in a tunnel or behind an extension header (an IPv6 one, or AH in IPv4) out of TCP reassembly and flow tracing;

• DPKT leaves TCP in an IP-in-IP tunnel or behind an AH or IPv6 extension header in IPv4 out of TCP reassembly and flow tracing, and IPv4 in IPv6 or IPv6 in IPv4 out of IP reassembly;

• Scapy reads IPv6 only as far as Scapy dissects it, so IPv6 behind some BSD loopback headers, and TCP and IPv6 fragments behind some IPv6 extension headers, go undissected.

Engine Support tabulates them.

Every engine also answers a preflight check, unsupported_reason(), which the Extractor constructor consults before importing the engine. An engine that cannot run here therefore gives one warning naming the cause (a Python version, a missing tshark, a missing libpcap(3), the wrong pcap distribution) and falls back to the default engine, rather than raising from inside the third-party package. The engine selection graph draws each fallback in order.

Note

The figures under Test Environment, Test Results and Test Results (Relative) are copied here by hand from a benchmark report; nothing regenerates them on this page. The Engine Benchmark workflow, .github/workflows/benchmark.yml, measures every engine on every interpreter it is tested on, weekly and on dispatch, and posts its environment and ratio tables to the benchmark: weekly engine benchmark report thread in the CI reports Discussions category (see CI Reports (#2030)); the per-version table is in the run’s job summary and its benchmark-out artifact. examples/benchmark/run.sh produces the same tables on any machine with Docker, and examples/benchmark/README.md explains what each one measures.

Engine Support by Python Version

Which engines can run, by interpreter, verified by installing each engine and extracting a capture on 3.10 through 3.15.

Engine	3.10	3.11	3.12	3.13	3.14	3.15
pcapkit	yes	yes	yes	yes	yes	yes
dpkt	yes	yes	yes	yes	yes	yes
scapy	yes	yes	yes	yes	yes	yes
pcap_ct	yes	yes	yes	yes	yes	yes
pypcap	yes	yes	no	no	no	no
pypcapfile	yes	yes	no	no	no	no
pyshark	yes	yes	yes	yes	no	no

pypcap and pypcapfile stop at 3.11, and pyshark at 3.13, for the reasons under Engine Prerequisites. Python 3.11 is the last version on which every engine can run – though pypcap and pcap_ct are mutually exclusive there too, so no single environment has every engine at once.

Engine Prerequisites

Some engines need more than a pip install. The unsupported_reason() preflight above enforces each constraint below in code.

PyShark

None of its requirements is visible to an import: the package imports cleanly and then fails when used.

• Drives Wireshark’s tshark binary, which need not be on PATH: pyshark looks at tshark_path in its config.ini first, then PATH on POSIX, both Program Files directories on Windows, and /Applications/Wireshark.app on macOS. Install Wireshark (or just tshark) from your platform’s package manager.

• Requires Python 3.13 or older. pyshark 0.6 builds its event loop with asyncio.get_event_loop_policy().get_event_loop(), and from Python 3.14 asyncio.get_event_loop() raises RuntimeError when there is no current event loop instead of creating one. Measured: a loop is returned silently on 3.10 and 3.11, with a DeprecationWarning on 3.12, and refused on 3.14; pyshark extracted a capture on 3.13.

PyPCAP

PyPCAP ships no wheels – only an sdist – so pip compiles it, and the build needs both libpcap’s headers (pcap.h) and its shared or static library:

# Debian/Ubuntu
sudo apt-get install libpcap-dev
# RHEL/Fedora/Amazon Linux
sudo dnf install libpcap-devel
# macOS
brew install libpcap

Caveats, each an upstream packaging problem:

• pypcap 1.3.0 ships a pre-generated pcap.c produced by Cython 0.29.x, which does not compile against the Python 3.12+ C API, so on 3.12 or newer the build fails whatever else is present. Use Python 3.11 or older for this engine, or regenerate pcap.c with Cython 3 yourself.

• Its setup.py does not consult CFLAGS/LDFLAGS or pkg-config. It searches a fixed list of prefixes – /usr, sys.prefix, /opt/libpcap*, ../libpcap*, ../wpdpack* and the macOS SDKs – so a libpcap installed anywhere else, notably Homebrew’s keg-only prefix on Apple Silicon (/opt/homebrew/opt/libpcap), is not found. Install it into sys.prefix or /opt/libpcap instead.

PCAP_CT

The way to drive the same libpcap interface on Python 3.12 and newer, where pypcap cannot be built. Nothing to compile and no pcap.h needed: pcap-ct is a ctypes reimplementation, and both it and its libpcap dependency ship py3-none-any wheels. Verified reading a capture on Python 3.10 and 3.14.

Caveats:

• A system libpcap(3) is still required at run time. The libpcap distribution ships a vendored libpcap.so and, as published, does not use it: its libpcap.cfg says LIBPCAP = None, which sends its loader to ctypes.util.find_library(). So the library actually loaded is the host’s libpcap.so.1, and with none present import pcap raises OSError rather than ImportError. Set LIBPCAP = tcpdump in libpcap.cfg to use the vendored copy instead.

• Both distributions are pre-releases, and pcap-ct documents itself as tracking the pypcap 1.2.3 interface. Every attribute the engine uses was measured behaving identically to pypcap 1.3.0, on the versions tested.

PyPCAPFile

PyPCAPFile 0.12.0 imports the imp module, which was removed in Python 3.12, so pcapfile.savefile – the module needed to read a capture – cannot be imported on 3.12 or newer. Upstream master has fixed this, but no release carries the fix, so this engine requires Python 3.11 or older.

Note

pcapkit itself and its default, dpkt and scapy engines need nothing beyond their pip install – dpkt 1.9.8 and scapy 2.8.0 were both measured reading a capture on Python 3.14. The scapy engine needs scapy 2.6.0 or newer, which the Scapy extra requires: 2.5.0 can drop a PCAP-NG interface’s timestamp resolution, reading nanoseconds as microseconds (#1572). pcapkit does not check the version at run time, so upgrade a scapy installed some other way.

Test Environment

These figures come from one run of the Engine Benchmark workflow on a GitHub-hosted runner, measured natively rather than in the examples/benchmark/run.sh container images, so the interpreters’ patch releases, tshark and libpcap are the runner’s own, as recorded below.

Run	Engine Benchmark run 40062653380, on GitHub’s hosted ubuntu-26.04-arm runner
Image	GitHub-hosted ubuntu-26.04-arm, image ubuntu26-arm64 20261023.150.3
pcapkit revision	c264485
Python	CPython 3.10.22, 3.11.17, 3.12.15, 3.13.16, 3.14.8, 3.15.1
Architecture	aarch64
Capture	in.pcap – 605 bytes, SHA-256 7aabadca49b329c6...
Packets per extraction	6
Iterations	1000 timed extractions per engine per pass, the first discarded as a warm-up
Passes	3 over the whole engine set, in 8 environments: 3.10-pcap_ct, 3.10-pypcap, 3.11-pcap_ct, 3.11-pypcap, 3.12-pcap_ct, 3.13-pcap_ct, 3.14-pcap_ct, 3.15-pcap_ct
tshark	TShark (Wireshark) 4.6.4.
libpcap	libpcap version 1.10.6 (64-bit time_t, with TPACKET_V3)

Resolved package versions:

Package	3.10-pcap_ct	3.10-pypcap	3.11-pcap_ct	3.11-pypcap	3.12-pcap_ct	3.13-pcap_ct	3.14-pcap_ct	3.15-pcap_ct
dpkt	1.9.8	1.9.8	1.9.8	1.9.8	1.9.8	1.9.8	1.9.8	1.9.8
libpcap	1.11.0b29	absent	1.11.0b29	absent	1.11.0b29	1.11.0b29	1.11.0b29	1.11.0b29
lxml	6.1.3	6.1.3	6.1.3	6.1.3	6.1.3	6.1.3	6.1.3	6.1.3
pcap-ct	1.3.0b3	absent	1.3.0b3	absent	1.3.0b3	1.3.0b3	1.3.0b3	1.3.0b3
pypcap	absent	1.3.0	absent	1.3.0	absent	absent	absent	absent
pypcapfile	0.12.0	0.12.0	0.12.0	0.12.0	0.12.0	0.12.0	0.12.0	0.12.0
pypcapkit	1.5.0b14	1.5.0b14	1.5.0b14	1.5.0b14	1.5.0b14	1.5.0b14	1.5.0b14	1.5.0b14
pyshark	0.6	0.6	0.6	0.6	0.6	0.6	0.6	0.6
scapy	2.8.0	2.8.0	2.8.0	2.8.0	2.8.0	2.8.0	2.8.0	2.8.0

Test Results

Measured with examples/benchmark/native.sh, which the Engine Benchmark workflow runs: 1000 timed extractions of examples/captures/in.pcap per engine and Python version. The first extraction is discarded as a warm-up. Values are milliseconds per packet, the median of 3 passes.

All columns come from one run on one machine, differing in the interpreter and nothing else, so they may be compared with each other. They may not be compared with figures from another machine: absolute times are a property of the host as much as of the engine.

Engine	3.10	3.11	3.12	3.13	3.14	3.15
pypcapfile [*]	0.0375	0.0262	–	–	–	–
pypcap [†]	0.0479	0.0342	–	–	–	–
pcap_ct	0.0667	0.0489	0.0564	0.0565	0.0613	0.0534
dpkt	0.0796	0.0599	0.0607	0.0616	0.0631	0.0858
scapy	0.2399	0.2028	0.2085	0.2021	0.2083	0.2208
pcapkit	0.6962	0.5709	0.5837	0.5660	0.6025	0.6128
pyshark [‡]	17.0812	17.6570	17.2537	17.1367	–	–

The unavailable cells were attempted. They are not zeroes and must not be compared with a measured row.

[*]

Not measured, for the reason recorded at the time:

• on 3.12 – pypcapfile does not support Python 3.12; its linklayer module imports `imp`, removed in Python 3.12

• on 3.13 – pypcapfile does not support Python 3.13; its linklayer module imports `imp`, removed in Python 3.12

• on 3.14 – pypcapfile does not support Python 3.14; its linklayer module imports `imp`, removed in Python 3.12

• on 3.15 – pypcapfile does not support Python 3.15; its linklayer module imports `imp`, removed in Python 3.12

[†]

Not measured, for the reason recorded at the time:

• on 3.12, 3.13, 3.14, 3.15 – pypcap is not installable on this interpreter, so it was not built here: pypcap 1.3.0 ships a pcap.c pre-generated by Cython 0.29.32, which does not compile against the Python 3.12+ C API. The engine is unsupported here rather than broken here.

[‡]

Not measured, for the reason recorded at the time:

• on 3.14 – pyshark does not support Python 3.14; it builds its event loop with `asyncio.get_event_loop_policy().get_event_loop()`, which raises RuntimeError from Python 3.14 when no current event loop exists

• on 3.15 – pyshark does not support Python 3.15; it builds its event loop with `asyncio.get_event_loop_policy().get_event_loop()`, which raises RuntimeError from Python 3.14 when no current event loop exists

Test Results (Relative)

Times relative to pcapkit, which is 1 by definition; lower is faster. Each ratio is taken against the baseline measured in the same environment on the same pass, so the machine cancels out. “Observed range” is the smallest and largest ratio actually seen across the passes.

Ratios are pooled across every environment in the run, which means across every Python version measured as well as both pcap providers. The observed range therefore includes any disagreement between interpreters about an engine; the per-version table is where that is broken out.

Engine	Relative time	Observed range	Samples
pypcapfile	0.0496	0.0456-0.0542	12
pypcap	0.0650	0.0600-0.0702	6
pcap_ct	0.0967	0.0852-0.103	18
dpkt	0.107	0.103-0.141	24
scapy	0.353	0.340-0.369	24
pcapkit	1	baseline	24
pyshark [§]	29.7	23.4-31.5	18

[§]

Measured, but on less evidence than the rows beside it. What was collected is real; what was lost is named here rather than left to be inferred from the pass count: 3 individual extraction(s) failed and were discarded.

The single absolute figure, for converting the ratios back: pcapkit itself ran at 0.5917 ms per packet on this run (0.566-0.715 across 24 passes). Absolute times are a property of the machine as much as of the engine, which is why every other figure here is a ratio.

Run-to-run spread: the median row varied by 17% of its ratio across the passes and the widest by 35%. A gap between two rows that is smaller than that is noise, not a result.