Skip to content

Latest commit

 

History

History
193 lines (129 loc) · 6.87 KB

README.md

File metadata and controls

193 lines (129 loc) · 6.87 KB
Raw

P4Benchmark

Tool to benchmark P4 Compilers and Targets

Installation

Run the following commands::

pip install -r requirements.txt
sudo python setup.py install

Generate P4 Program and PCAP file for testing

'--version' option can be used to set the P4 Program language to 14/16

  • Benchmark parse field

The generated P4 program parses Ethernet, PTP and a customized header containing 4 fields and each field is 16-bit wide::

p4benchmark --feature parse-field --fields 4 --version 16
  • Benchmark parse header

The generated P4 program parses Ethernet, PTP and a customized number of headers each containing a customized number of fields. Each field is 16-bit wide::

p4benchmark --feature parse-header --fields 4 --headers 4 --version 16
  • Benchmark parse complex

The generated P4 program parses Ethernet, PTP and a parse graph that has the depth of 2 and each node has 2 branches::

p4benchmark --feature parse-complex --depth 2 --fanout 2 --version 16
  • Benchmark action complexity

The generated P4 program has N=2 number of set-field operations::

p4benchmark --feature set-field --operations 2 --version 16
  • Benchmark header addition

The generated P4 program adds N=2 number of headers to packets::

p4benchmark --feature add-header --headers 2 --version 16
  • Benchmark header removal

The generated P4 program removes N=2 number of headers to packets::

p4benchmark --feature rm-header --headers 2 --version 16
  • Benchmark processing pipeline

The generated P4 program applies N=2 number of tables::

p4benchmark --feature pipeline --tables 2 --table-size 32 --version 16
  • Benchmark Read State

The generated P4 program declares 1 register and performs 10 number of read operations::

p4benchmark --feature read-state --registers 1 --operations 10 --version 16
  • Benchmark Write State

The generated P4 program declares 1 register and performs 10 number of write operations::

p4benchmark --feature write-state --registers 1 --operations 10 --version 16

Generated Files

The output directory contains::

$ ls output
commands.txt  main.p4  run_switch.sh  run_test.py  test.pcap

1. main.p4        The desired program to benchmark a particular feature of the P4 target
2. test.pcap      Sample packet crafted to match the parser or tables
3. run_switch.sh  Script to run and configure bmv2
4. commands.txt   Match-action rules for tables
5. run_test.py    Python packet generator and receiver

Run Behavioral Target

Require Behavioral Model 2 and P4C

$ cd output
$ ./run_switch

Run Python packet generator

In another terminal, run::

$ cd output
$ sudo ./run_test.py --nb-headers 1 --nb-fields 4

PKTGEN (Send PCAP file)

Or, you could use a high performance packet generator that sends the prepared PCAP file and sniffs for returning packets

Build

Requrire cmake and libpcap::

cd pktgen
mkdir build
cd build
cmake ..
make

Run pktgen

This packet generator reads the prepared PCAP file and send c copies of that packet at t Byte per second out of the interface veth4. The result is stored in the result directory::

$ p4benchmark/pktgen/build
$ sudo ./p4benchmark -p ../../output/test.pcap -i veth4 -c 10000 -t 10000 -o result

TO MEASURE LATENCIES

On compiling the p4 program, with the suitable compiler backend, a configuration file is produced. This configuration can be dumped onto the respective hardware component, say FPGA/PISCES. The output produced by Whippersnapper 2.0 is capable of running the p4 program on p4 software switch aka bmv2 or behaviour model version 2. The test.pcap file has the structure of the packet to be sent. To measure the latency produced by simple_switch (the time for packet to be processed and pass out of the switch) can be measured in two ways:

Using pktgen: Build pktgen, by the above procedure

Setup

	sudo ./veth_setup.sh -- once per session
	g++ DataAlgo.cpp -o DataAlgo
	g++ Percent.cpp -o Percent
	Change lines 23, 63, 65 in measure_latency.sh according to the feature being tested.

Run Test

    sudo ./measure_latency.sh -- give desired no of packets, transmission rate and version, ex: 10000, 10000, 16

The latency values will be stored to ---.txt and the normalised percentage alues will be stored to ----Percent.txt

DataAlgo filters the outliers in the dataset to some extent. It first calculates the mean of entire data, then clusters data surrounding the mean within an offset of standard deviation and then recalculates the mean of this clustered data.

Using tshark (3rd party)

Setup

	sudo apt install tshark
	cd /usr/share/wireshark/
	nano init.lua  -- In line 29 set disable_lua = true
	change line 30 of latency_new.sh according to the feature being tested.

Run Test

	sudo ./latency_new.sh -- give desired no of packets and version, ex: 10000, 16

The results will be saved in latency.csv

Working

The above script is to automate the testing of a feature completely. The actual process going on is

  1. The p4benchmark will produce the output directory to test certain feature.
  2. The main.p4 program will be compiled with p4c-bm2-ss compiler.
  3. tshark will monitor the interfaces being used by the switch, and to print timestamp to csv files.
  4. A simple_switch will be setup with the main.json file and with some veth interfaces as its ports.
  5. RuntimeCLI will populate the match-action tables of simple_switch from commands.txt.
  6. The run_test.py file will send n copies of test.pcap file to simple_switch port.
  7. The simple_switch will process the packet recieved on the ingress port and send the output packet to the egress port.
  8. The packet arrival epoch timestamps and their number will be printed to a file on both ingress and egress interfaces.
  9. Average of difference of these timestamps is taken for all the packets, which represents the latency.
  10. An algorithm is used to eliminate buggy values due to glitches in packet transfer, droppings. This is based on the fact that latency values are expected to be similar for each packet. The latency values are divided into various category. The category having the highest frequency will be the one to be selected and average of all values of only that category will be calculated. Note that if two categories are having major frequency count which is a rare case, we don't get much error by considering only one of them.

Contributors

Madhav Tummala
Siddhant Bajaj
Abhinav Jindal
Pavan Kumar Vaddineni