GSOC2020Prague

Main Page - Roadmap - Summer Projects - Project Ideas - Developer FAQ - Tools - Related Projects

HOWTOs - Installation - Troubleshooting - User FAQ - Samples - Models - Education - Contributed Code - Papers

Back to GSoC 2020 projects

Project Overview

• Project name: TCP Prague model for ns-3
• Student: Deepak K
• Mentors: Ankit Deepak, Mohit Tahiliani, Vivek Jain, Viyom Mittal
• Project goals and outcomes: The original plan at the start of this project was to add TCP Prague to ns-3 mirroring ns-3 DCTCP, implement Classic ECN detection, RTT independence and Dynamic pacing. After several discussions with my mentors, these goals and priorities were altered. The following features have been implemented in ns-3 at the end of this project:
  • Dynamic pacing model in TCP, along with unit tests
  • TCP Prague that aligns with contemporary Linux implementation, along with unit tests
  • RTT independence in TCP Prague, along with unit tests
  • Experiments to generate plots and validate the ns-3 implementation of TCP Prague with Linux
• Links to code written in each phase:
  • GSoC 2020 Phase 1: Add dynamic pacing rate model to TCP
  • GSoC 2020 Phase 2: Add base TCP Prague model with RTT independence
  • GSoC 2020 Phase 3: Refactor TCP Prague and align with Linux
  • Alignment plots in the one-flow and two-flow scenarios.
• Future work:
  • The merge request for dynamic pacing is almost merged to ns-3, and is in its final stages of review This has been successfully merged to mainline ns-3.
  • Adding system-wide tests for TCP Prague is something that was in the scope of GSoC work, but could not be fully completed. (I am currently working on these tests). As such, the merge request for TCP Prague will soon be merged once these system tests are implemented.
  • A few features that we chose not to address during GSoC, but would definitely help the ns-3 Prague model include: merging Dual Queue model to ns-3, adding Classic ECN detection to TCP Prague, and merging Accurate ECN to ns-3.
• Project report: A project page was also added to give an elaborate explanation of the different features implemented as part of this work.
• Repository: https://gitlab.com/deepakkavoor/ns-3-dev
• About me: I am a third-year undergraduate student pursuing Computer Science and Engineering at the National Institute of Technology Karnataka (NITK), India. My interests include computer networks and cryptography. Previously I have worked on implementing the Set-Associative Hashing feature for Fq-CoDel AQM in ns-3 (link to my commit).

Milestones and Deliverables

Phase 1

This phase implemented the dynamic pacing feature in ns-3 TCP.

Link to Phase 1 code: click here

• The outcome was achieving alignment of ns-3 pacing with Linux implementation, which sets the current pacing rate based on the congestion window and prior RTT measurements.
• On receipt of each ACK, a sender would update the pacing rate as follows: pacingRate = factor * cWnd * MSS / rtt
• The value of factor depends on whether TCP is currently in Slow Start or Congestion Avoidance. The default values used by Linux are 2 and 1.2 respectively, and we use the same. A higher factor in Slow Start allows TCP to probe for higher speeds early on.
• A test suite was also added to check that packets are correctly paced out as per the above update equation.
• The work done in this phase was documented in ns-3.
• I also tried to designed experiments in ns-3 to show improvements when using dynamic pacing.

Phase 2

This phase implemented the RTT independence feature in ns-3 TCP Prague.

Link to Phase 2 code: click here

• The outcome was addition of ns-3 TCP Prague as a mirror of ns-3 DCTCP and implementation of the RTT independence feature following Linux Prague.
• The cWnd increment during Congestion Avoidance is modified, so that Prague behaves as if it is operating at a (usually higher) target RTT. This reduces unfairness when Prague and other classic congestion controls (such as Reno) coexist within a common bottleneck.
• There are currently three RTT scaling heuristics in Linux: Rate, Scalable and Additive, and all were implemented in ns-3.
• As an example, the update equation for congestion window increment on receipt of each ACK using the "Rate" heuristic would be: increment = (curRTT / targetRTT) * (curRTT / targetRTT) * 1 / cWnd, where curRTT denotes the current RTT measurement and targetRTT denotes the target RTT at which Prague is expected to operate.
• Unit tests were added to check that RTT independence in ns-3 Prague works as expected.
• The work done in this phase was documented in ns-3.

Phase 3

This phase refactored the ns-3 Prague code to align with Linux.

Link to Phase 3 code: click here

• The structure of ns-3 Prague was moved away from ns-3 DCTCP. A "PRR-like" cWnd reduction approach (that Linux follows) was added on receipt of an ACK with ECE mark.
• Along with this, experiments were conducted to obtain aligning plots between ns-3 and Linux implementations of Prague.
• We used the one-flow scenario (and two-flow scenario in case of RTT independence) from the l4s-evaluation module to validate ns-3 Prague with Linux. The topology was evaluated in Linux using network namespaces (specifically using the NeST framework).

A more elaborate description of the features added in these phases can be found in the project site. Along with the above features, I also worked on the following aspects that were indirectly related to TCP Prague:

• ECN++ and Accurate ECN, which were implemented in ns-3 as part of a 2018 GSoC project was rebased onto ns-3-dev, along with a few revisions to test suites. The rebased branch can be found here.

Weekly Reports

Community Bonding Period (May 4 - May 31)

• Established the Wiki page for this project.
• Rebased the GSoC 2018 implementation of ECN++ and AccECN (without TCP options) to the latest ns-3-dev.
• Went through the documentation for TCP model in ns-3 and read about paced chirping.

After a discussion with mentors and the L4S Team, it was decided that the first phase of this project should focus on dynamic pacing in TCP Prague.

Week 1 (June 1 - June 7) (Start of Phase 1)

• Agreed upon the pacing structure for TCP Prague in ns-3. The document can be found here.
• Integrated basic pacing model for New Reno, DCTCP and TCP Prague in ns-3 by adding methods PacingEnabled () and UpdatePacingRate (). This allows any congestion control to dynamically update its pacing rate depending on whether it is in Slow Start or Congestion Avoidance.

Week 2 (June 8 - June 14)

• Designed a test suite TcpPacingTest to test TCP packet pacing rate during Slow Start and Congestion Avoidance.
• Rebased Joakim Misund's work on ns-3 Prague to latest ns-3-dev.
• Generated an MR to add support for dynamic pacing, along with the above test suite.

Week 3 (June 15 - June 21)

• Worked on fixing a few issues with the MR generated in Week 2.
• Went through the code for RTT Independence in Linux Prague, and gathered a few notes in this document.
• Added code to enable different scaling heuristics (similar to Linux).

  typedef enum
  {
    RTT_CONTROL_NONE,      //!< No RTT Independence
    RTT_CONTROL_RATE,      //!< Flows with e2e RTT < target try to achieve same throughput
    RTT_CONTROL_SCALABLE,  //!< At low RTT, trade throughput balance for same marks/RTT
    RTT_CONTROL_ADDITIVE   //!< Behave as a flow operating with extra target RTT
  } RttScalingMode_t;

Week 4 (June 22 - June 28)

• Divided the MR (from Week 2) to now separately add support for dynamic pacing (!339) and TCP Prague (!373).
• Ran experiments using Linux namespaces to check the behaviour of TCP flows when dynamic pacing is enabled.
• Compared the plots from Linux and ns-3 for congestion window and pacing rate, and observed several differences; I am trying to identify why Linux has more frequent cWnd oscillations compared to ns-3.

Week 5 (June 29 - July 5)

• Generated merge requests for ECN++ and AccECN (rebased from Wenying Dai's GSoC 2018 work).
• Added dynamic pacing to other congestion controls in ns-3.
• Analyzed packet traces (using Wireshark) for topologies designed in Linux namespaces, and inferred the following: if a Linux TCP sender has more than two eligible packets to be sent, the first two are sent back-to-back, and the rest are paced out.

Week 6 (July 6 - July 12)

• Extended examples/tcp/tcp-pacing.cc to highlight dynamic pacing and produce time-series plots of Congestion Window, Slow Start threshold and current Pacing Rate.
• Added documentation regarding dynamic pacing in ns-3, and its similarities and differences to that in Linux.

Week 7 (July 13 - July 19)

• Explored different topologies and configurations that would show the benefits of dynamic pacing.
• Modified the pacing example to follow the topology as given in Figure 2 here.
• Continued to work on RTT Independence and added/modified the following methods:
  • TcpPrague::AiAckIncrease (): Update the per-ACK cWnd AI increase factor depending on the current RTT scaling heuristic.
  • TcpPrague::IncreaseWindow (): Consider the RTT scaling factor when updating cWnd during AI.
  • TcpPrague::PktsAcked (): Apply the EWMA update equation and increase round count only if the value returned by TcpPrague::ShouldUpdateEwma () is true.

Week 8 (July 20 - July 26)

• Improved and refactored the code for Prague RTT independence to align with Linux.
• Added unit tests to check correct increment of Congestion window for different RTT scaling heuristics.
• Fixed bugs in ECN++ and AccECN code that were causing DCTCP tests to fail.
• Participated in the IETF 108 Hackathon and helped with integration of ns-3 Prague model alongside other L4S components (such as Dual Queue).

Week 9 (July 27 - August 02)

• Setup a Linux kernel supporting TCP Prague, Dual Queue model and Accurate Ecn by referring to the contemporary Linux development branch maintained by the L4S team.
• Worked on setting up a topology in this kernel corresponding to the one-flow scenario using Linux namespaces.

Week 10 (August 03 - August 09)

• Worked on experiments using the ns-3 l4s-evaluation module to show that Prague experiences higher throughput in presence of other Classic flows when RTT independence is disabled in ns-3 Prague.

Week 11 (August 10 - August 16)

• Major refactoring of ns-3 Prague code to match the Linux implementation. The following methods were added:
  • TcpPrague::UpdateCwnd (): Update the congestion window on receipt of each ACK, and use a "PRR-like" reduction on ECE marks.
  • TcpPrague::UpdateAlpha (): Update the value of alpha (which corresponds to an EWMA of the fraction of bytes that were CE marked) based on TcpPrague::ShouldUpdateEwma ().
  • TcpPrague::SlowStart (): When cWnd < ssThresh / 2 use a slow start algorithm similar to ns3::TcpLinuxReno.
  • TcpPrague::CwndChanged (): Call TcpPrague::AiAckIncrease whenever congestion window is updated.
  • TcpPrague::EnterLoss (): On encountering a packet loss, gradually reduce congestion window to half its original value.
• Validated this refactored implementation against Linux namespaces and obtained alignment in the one-flow scenario.

Week 12 (August 17 - August 23)

• Implemented a topology in Linux namespaces corresponding to the two-flow scenario, with Prague and Reno running over a common bottleneck supporting Dual Queue.
• Obtained aligning results in the two-flow scenario that validates RTT independence in ns-3 Prague with that of Linux.
• Worked on adding final documentation of TCP Prague to ns-3.
• Worked on the project site.

Week 13 (August 24 - August 30)

• Added doxygen comments to the ns-3 Prague model.
• Updated the MR for ns-3 Prague model.
• Fixed issues related to delayed acknowledgements in the TCP pacing test suite.
• Worked on understanding how TCP system tests are written in ns-3.