Rebecca Portelli: CheesePi: Delay Characterization through TCP-based Analysis from End-to-End Monitoring

Student: Rebecca Portelli
Thesis: CheesePi: Delay Characterization through TCP-based Analysis from End-to-End Monitoring
University: KTH, School of Electrical Engineering (EES)
Year: 2016
Co-advisor: Liam McNamara
Now at: Spotify, Stockholm

Links

About the thesis

Programme: Information and Communication Technology, KTH (30 credits), 2016
Supervisors at SICS: Ian Marsh and Liam McNamara
Project: CheesePi

People expect web pages to respond quickly, and web browsing is among the most delay-sensitive common services. Yet most users cannot measure delay from their home network themselves, and ISPs cannot see the network from the user's side. The usual tool, ping, sends ICMP packets that routers often treat differently from real traffic, or block.

This thesis measured delay inside the protocol: from the packets of ordinary web requests over TCP. Every TCP connection has three phases, the handshake, the data transfer and the teardown, and each can act as a timer.

TCP flow when retrieving a web page
A TCP flow retrieving a web page: three-way handshake, HTTP request and data transfer, then teardown.

Method

  • Capture: packets were captured on Raspberry Pis in the CheesePi network, first at the client only, then at both client and server.
  • Automation: a web client fetched the same resource repeatedly: 3,000 TCP flows per experiment, about an hour of traffic, enough to capture the network's behaviour without loading it.
  • Analysis: offline, each flow was dissected into its phases to compute round-trip delay, one-way delay, end-host processing delay, web transfer delay and packet loss.
  • Experiments: client-only measurements to servers in Sweden (cheesepi.sics.se), the UK (cambridge) and Australia (sydney.edu.au); then two-ended experiments from SICS in Kista to Pis at Karlstad University (KU) and Luleå University of Technology (LTU), fetching a 57.4 kB resource, repeated morning, afternoon and evening on a weekday (19 May 2016).
End-to-end monitoring at both client and server
Monitoring at both ends of the client–server communication.

Results

TCP beats ping. The client's partial handshake RTT, once the server's processing time is removed, gives an accurate picture of the pure network delay. It measures what real traffic experiences, which is not always what ICMP experiences.

Histograms of ping RTT and TCP partial handshake RTT
3,000 pings (green) against 3,000 TCP partial handshake RTTs (blue) on the same path. The handshake includes a small, measurable processing delay.

Both ends see more than one. With captures at client and server:

  • Network delay and kernel processing delay can be separated during the handshake and teardown.
  • The two RTTs within a handshake (client-side and server-side) are not symmetrical.
  • The teardown RTT is longer than the handshake RTT in the same flow, since the server takes measures against SYN-flooding attacks.
  • Routing asymmetry shows up by comparing one-way delay on the forward and reverse paths.
  • Apparent losses can be told apart from packets that were simply reordered or delayed.
CDFs of data transfer delay over a day
Web transfer delay to Karlstad in the morning, afternoon and evening: medians of about 25–28 ms.

Conclusions and future work

Monitoring TCP flows from both ends of a client–server communication gives a more accurate and richer view of network delay than monitoring from the client alone, and TCP-based analysis characterises round-trip delay better than ICMP-based analysis. TCP is the common case in IP networks, so measuring it gives a more realistic view.

Future work suggested: varying page sizes and content to see how data size affects loss and transfer delay, and measuring under different network conditions and loads.

The technique is published as Delay characterisation through in-protocol measurements (MoMM 2016), with Rebecca as co-author, and described in CheesePi: Research, Results, and the Regulator.

Thesis

Full text (DiVA)

Original entry

Rebecca Portelli, CheesePi: Delay Characterization through TCP-based Analysis from End-to-End Monitoring, 2016, now at Spotify, Stockholm, Sweden, thesis page (co-advised with Liam McNamara).

Project

Related publications

BibTeX

@mastersthesis{Portelli1039071,
  author   = {Portelli, Rebecca},
  title    = {CheesePi: Delay Characterization through TCP-based Analysis from End-to-End Monitoring},
  school   = {KTH, School of Electrical Engineering (EES)},
  series   = {EES Examensarbete / Master Thesis},
  number   = {TRITA EE 2016:105},
  pages    = {89},
  year     = {2016},
  url      = {http://www.diva-portal.org/smash/record.jsf?pid=diva2:1039071},
  keywords = {Active IP Network traffic measurement, TCP flow analysis, network delay characterization, end-to-end monitoring, Aktiv IP-nätverk trafik mätning, TCP flödesanalys, karakterisering av nätverksfördröjning, end-to-end-monitoring}
}

Source: DiVA record diva2:1039071

More in Masters theses: ← Sagar Sharma (2016)  |  Biniam Gebregergs Guulay (2015) →

← All supervised theses