2 unstable releases

0.5.0 Jul 15, 2021
0.4.0 Oct 30, 2018

#217 in Debugging

ISC license

405KB
10K SLoC

Rust 6.5K SLoC // 0.0% comments C 3K SLoC // 0.1% comments Python 689 SLoC // 0.1% comments

pyportus

This module provides a python interface to the Portus CCP implementation.

Setup

To build and use the python bindings, you also need setuptools_rust

sudo pip install setuptools_rust

You need to tell it to use the nightly version of rust since some features are still experimental:

  • Check where the package was installed: pip show setuptools_rust
  • Edit packages/setuptools_rust/build.py
  • Search for the line containing "rustc" (should be ~102), and change the args to be ["cargo", "+nightly", "rustc", ...

Now, rather than running make, you can build with

python setup.py develop

Depending on your python environment setup, you may need to run this with sudo (and ensure that your PATH variable is preserved):

sudo env PATH=$PATH python setup.py develop

Now you should be able to import the package like so:

import portus

Writing Algorithms

Overview

An algorithm in portus is represented by a Python class and an instance of this class represents a single TCP flow. A new instance is created for each flow.

This class must be a subclass of portus.AlgBase and must implement the two following method signatures:

  • on_create(self)
  • on_report(self, r)
    • r is a Report object containing all the fields defined in your datapath program, as well as the current Cwnd and Rate. Suppose your program defines just a single variable: (def (acked 0)), where acked adds up the total bytes acked since the last report. This value can be accessed as r.acked. Similarly, you can access the cwnd or rate as r.Cwnd and r.Rate (captialization important!).

Each instantiation of the class will automatically have two fields inside self:

  • self.datapath is a pointer to the datapath object that can be used to install new datapath programs. It has two available methods:
    1. datapath.install( str ), which takes a datapath program as a string. It compiles the program and installs it in the datapath. It does not return anything, though it may raise an exception if your program fails to compile.
    2. datapath.update_field(field, val), which takes a variable in the Report scope of your datapath program and sets the value to val. For example, to update just the cwnd, you could use datapath.update_field("Cwnd", 10000) (note: cwnd is denoted in bytes, not packets).
  • self.datapath_info is a struct containing fields about this particular flow from the datapath (this could be used, for example, in on_create to set an initial cwnd based on the datapath's mss)
    • sock_id: unique id of this flow in the datapath
    • init_cwnd: the initial congestion window this flow will have until you set it
    • src_ip, src_port, dst_ip, dst_port: the ip address and port of the source and destination for the flow

Datapath Programs

Datapath programs are used to (1) define which statistics to send back to your usespace program and how often and (2) set the congestion window and/or pacing rate. A datapath program is written in a very simple lisp-like dialect and consists of a single variable definition line followed by any number of when clauses:

(def ( ... ) ( ... ))
(when (event) (
  do_stuff ...
)
(when (other_event) (
    do_other_stuff ...
)

NOTE: the following info is out of date as the datapath program API has been

updated

1. Report Variable Definitions

Example: (def (Report.acked 0) (Report.rtt 0) (Report.timeout false))

This line defines the names and initial values of variables in the report scope. Calling (report) in your datapath program results in a call to your algorithm's on_report function with the current value of these variables. After the call these variables are reset back to their initial value.

NOTE: Variables in datapath programs are written as {scope}.{name}. For example, the acked variable in the Report scope is written as Report.acked. Therefore, all variables defined in this line must start with Report. However, when you access them in on_report, you just provide the variable name. In our example above, Report.rtt defines the variable rtt in the Report scope. If we want to access this value in on_report(r), we'd use r.rtt (i.e. not r.Report.rtt).

2. When Clauses

When clauses consist of a boolean expression and a set of instructions. On each ack, the datapath checks the boolean expression, and if it evaluates to true, it runs the set of instructions. For example, the following when clause sends a report (i.e. calls the on_report function) once every rtt:

(when (> Micros Flow.rtt_sample_us)
    (report)
)

Putting it all together

A sample algorithm definition showing the full API:

import portus

# Class must sublcass portus.AlgBase
class SampleCCAlg(portus.AlgBase):
  # Init must take exactly these parameters
  def __init__(self, datapath, datapath_info):
    # Store a copy of the datapath and info for later
    self.datapath = datapath
    self.datapath_info = datapath_info
    
    # Internally store an initial cwnd value
    self.cwnd = 10 * self.datapath_info.mss
    
    # Install an initial datapath program to keep track of the RTT and report it once per RTT
    # The first when clause is true on every single ack,
    #    which means the 'Report.rtt' field will always keep the latest rtt sample
    # The second when clause is true once one rtt's worth of time has passed, 
    #    at which point it will trigger on_report, and Micros (and Report.rtt) will be reset to 0
    self.datapath.install("""\
    (def
        (Report.rtt 0)
    )
    (when true
        (:= Report.rtt Flow.rtt_sample_us)
        (fallthrough)
    )
    (when (> Micros Flow.rtt_sample_us)
        (report)
    )
    """)

  # This function will be called once per RTT, and the report struct `r` will contain:
  # "rtt", "Cwnd", and "Rate"
  def on_report(self, r):
      # Compute new cwnd internally 
      # If the rtt has decreased, increase the cwnd by 1 packet, else decrease by 1 packet
      if self.last_rtt < r.rtt:
          self.cwnd += self.datapath_info.mss
      else:
          self.cwnd -= self.datapath_info.mss
      self.last_rtt = r.rtt
      
      # Send this new value of cwnd to the datapath
      self.datapath.update_field("Cwnd", self.cwnd)
    

Important Notes

  1. You should install an initial datapath program in your __init__ implementation, otherwise you will not receive any reports and nothing else will happen. You can always install a different datapath program later when handling on_report.
  2. If you want to print anything, you should use sys.stderr.write() (note that you need to import sys and that it doesn't automatically add new lines for you like print does).
  3. You must store a reference to datapath in self called "datapath" (i.e. self.datapath = datapath), because the library internally uses this to access the datapath struct as well.

Starting CCP

The CCP entry point is portus.connect(ipc_type, class, debug, blocking):

  • ipc_type (string): (netlink | unix | char) on linux or (unix) on mac
  • class: your algorithm class, e.g. SampleCCAlg
  • debug (bool): if true, the CCP will log all messages passed between the ccp and datapath
  • blocking (bool): if true, use blocking ipc reads, otherwise use non-blocking

For example: portus.connect("netlink", SampleCCAlg, debug=True, blocking=True).

Regardless of whether you use blocking or non-blocking sockets, connect will block forever (to stop the CCP just send ctrl+c or kill the process).

Example

For a full working example of both defining an algorithm and running the CCP, see the simple AIMD scheme in ./aimd.py and try running it: sudo python aimd.py.

Dependencies

~8.5MB
~156K SLoC