MPTCP Userland implementation

• Download: version 0.1 (tar.gz)
• Requirements:
  • 32 bit, Linux 2.6 kernel (important - does not compile on 64 bit)
  • flex
  • bison
  • linux headers 2.6.17 (download here)
• Install: unzip, edit Rules.mk (see README in archive), compile.
• Test: run send_tcp and recv_tcp programs in src/test
• Userland code: The code implements multipath on top of a userland port of the Linux 2.3.29 stack by Kandula et. al (description here). The port simply takes the unmodified ip and tcp code from the kernel, and wrapps it around with pcapture to receive data, and raw sockets to send data.Despite the kernel being dated, the userland version allows testing and development much more quickly than in-kernel implementations.
• MPTCP Protocol: the code implements a subset of the current multipath TCP draft. It uses options to advertise additional addresses, and includes a data sequence number to allow the receiver to reorder data at connection level. It does not implement the full mapping as in the draft: it only includes the starting connection level offset in each segment. This means the connection will likely fail unpredicatbly if it goes through middleboxes that resegment.
• Congestion Control:: the code implements the linked increases algorithm for multipath tcp, as described in the congestion control draft. It allows the user to choose the congestion control algorithm by changing defines in src/user-tcp/ipv4/fake-linux.h.
• Limitations: the code is single threaded, thus client must use asynchronous socket calls. It will not work with unmodified TCP apps; for examples of use please look at the test programs provided.

htsim simulator (for TCP and MPTCP)

• Download: version 0.2 (tar.gz)
• Requirements:
  • Linux, recent g++
• Install: unzip, run make in sim directory (see README in archive)
• Test: run htsim in sim directory. It will provide little useful output on the console. It stores the output in a binary format in the data directory.
• Process results:install R first (http://cran.r-project.org/). You will need to install various R packages, including a custom package by Damon Wischik (djwutils_2.10.tar.gz). In R, set the working directory to analysis and load analysis.R This will parse the data/logout.dat file. See README file, and the docs/ directory in the archive for some basic examples that plot various data. A brief introduction to R is given here
• Fast simulator: htsim allows testing both high speed and large numbers of tcp flows concurrently. It models the network as a collection of pipes (that add delays) and queues (with fixed processing capacity, and finite buffers), allowing the creation of arbitrary network topologies. Each packet is allocated only once (references are passed between network hops), and there is a memory pool of allocated packets to minimize malloc/free overhead. Finally, all flows' timeouts are triggered at the same time: the code performs a periodic sweep of all the connections and triggers RTO handlers as needed. This unnecesarily synchronizes the RTOs of different flows, and will bias results if RTOs are frequent. The console output prints an info for every RTO; eyeballing their count is one way of ensuring the results are not distorted by RTOs.
• MPTCP Implementation: the code only implements congestion control coupling. It implements the linked increases algorithm for multipath tcp, as described in the congestion control draft, as well as Fully Coupled, and independent tcps. It allows the user to choose the congestion control algorithm by passing a parameter to htsim. It maintains the window in bytes (as opposed to the Linux based implementations, which maintain the window in packets).
• Limitations: the code does not implement reordering at connection level, of any other parts of the multipath protocol. The subflows are independent TCP connections, that are only linked through congestion control.

Confidential Content-Based Publish/Subscribe