WWW.FED4FIRE.EU

Joint experimentation of modern Internet application protocols with SDN (Go-Quick) Fed4FIRE+ Open Calls

March 15th, Paris, France

Dr. Ioannis GiannoulakisEight Bells Ltd

giannoul@8bellsresearch.comwww.8bellsresearch.com

WWW.FED4FIRE.EU

Ø Go-Quick experimentation aims to evaluate the performance of QUIC, as compared to HTTP and SPDY/HTTP2, through deploying virtual overlay networks on Fed4FIRE+ OpenFlow SDN-enabled infrastructure.

Ø Since sophisticated congestion avoidance and packet error correction mechanisms are used in all protocols, the actual network scenarios of the Go-Quick experiments determine which protocol performs best in each case.

Ø By deploying various network conditions and for different size of objects, performance metrics like throughput are assessed for the three protocols, QUIC, SPDY/HTTP2, and HTTP.

Ø Different virtual overlay networks are deployed on Fed4FIRE+ OpenFlowSDN-enabled infrastructure to adapt to changing network scenarios.

2

Project Summary

WWW.FED4FIRE.EU

Ø Deployment of a virtual overlay network on Fed4FIRE+ OpenFlow SDN-enabled infrastructure

Ø Installation and configuration of QUIC, APACHE servers and emulation of different network conditions

Ø Google’s QUIC Client and Server are meant mainly for integration testing: neither is performant at scale

Ø Go Quick has developed Bash scripts for full automation of the experiments

Ø Performance evaluation of QUIC as compared to other L7 protocols (e.g., HTTP/HTTP2)

3

Objectives

WWW.FED4FIRE.EU4

Go Quick Topology

HTTPServer

HTTP2Server

QUICServerFile A

OFELIA server host

OFELIA Switches

OFELIA client host

HTTPClient

HTTP2Client

QUICClient

OFELIA VTAM & OFAM aggregate managers

WWW.FED4FIRE.EU5

Internet Protocols in a nutshell

*P. Megyesi, Z. Kramer, S. Molnar, “How quick is QUIC?”, in Proc. of IEEE ICC 2016, 22-27 May 2016.

Ø One of the bottlenecks of HTTP performance is the opening of too many TCP connections to achieve concurrency. A large portion of HTTP data flows consist of small (less than 15KB), bursty data transfers over dozens of distinct TCP connections.

Ø Another limitation is that HTTP based web transfers are strictly initiated by the client. This presents a serious problem because it hurts performance significantly in the case of loading embedded objects.

Ø HTTP2 introduces request prioritization. The client is allowed to specify a priority level for each object and the server then schedules the transfer of the objects accordingly.

Ø QUIC is working over UDP. The protocol does not force in-order delivery of packets thus QUIC avoids HOL blocking.

WWW.FED4FIRE.EU

Ø QUIC aims to improve performance compared to HTTP, HTTP2 by multiplexing web objects in one stream over UDP

Ø “Go Quick” performed a comparative analysis

QUIC – SERVER & CLIENT

6

Results1 2 4 8 10 15 20 30 40 50 60 80 100

200

500

1000

0102030405060708090

Files

Mbp

s

QUIC - quic_client

QUIC quic_client10MBQUIC quic_client1MBQUIC quic_client100KBQUIC quic_client10KB

WWW.FED4FIRE.EU

Ø Performance is greatly affected by the file size and the number of concurrent file requests.

Ø OFELIA SDN physical switches were configured to route data and to create congestion and losses according to the network scenarios of the experiment.

QUIC SERVER – CHROME CLIENT

7

Results1 2 4 8 10 15 20 30 40 50 60 80 100

200

500

1000

01020304050607080

Files

Mbp

s

QUIC - Chrome

QUIC chrome10MBQUIC chrome1MBQUIC chrome100KBQUIC chrome10KB

WWW.FED4FIRE.EU

QUIC HTTPS HTTPS2 (100KB)

8

Results1 2 4 8 10 15 20 30 40 50 60 80 100

200

500

1000

0

50

100

150

200

250

Files

Mbp

s

QUIC - https - https/2 100KB

QUIC quic_client

QUIC chrome

HTTPS wget

HTTPS chrome (apache2.4.18 mpm prefork)HTTPS chrome (apache2.4.27 mpm event)HTTPS/2 chrome (apache2.4.27 mpm event)

Ø In all scenarios QUIC protocol is outperformed by HTTPS/HTTPS2

Ø For typical web object sizes and for a big number of objects (i.e., many parallel flows), HTTPS2 is up to 80% better.

Ø In future work, measurements with Google Sites Server will be collected to exploit multi-threadingcapabilities

WWW.FED4FIRE.EU

Ø Negligible delays and losses. Ø only delay by the interfaces and routing software

Ø Comparatively low values for delays and losses. Ø delay of 12.5 ms is added and losses of about 0.5%, both upstream and downstream

Ø Moderate values for delays and losses. Ø delay of 25 ms is added and losses of about 1%, both upstream and downstream

Ø Moderate values for delays and losses. Ø delay of 50 ms is added and losses of about 2%, both upstream and downstream

Ø Comparatively high values for delays and losses. Ø delay of 100 ms is added and losses of about 4%, on both upstream and downstream

4 SCENARIOS USING SDN AND LINUX TC TOOL

9

Network conditions under congestion

WWW.FED4FIRE.EU10

Throughput rate under the 4 scenarios

0.1

1

10

100

1000

0_0.0 12.5_0.5 25_1.0 50_2.0 100_4.0

https_1MB

http2_1MB

quic_1MB

https_100KB

http2_100KB

quic_100KB

https_10KB

http2_10KB

quic_10KB

Ø QUIC is affected less by losses and network delays

WWW.FED4FIRE.EU

Ø Go-Quick project leveraged the OpenFlow OFELIA testbed provided by Fed4FIRE+ (i2CAT) for the small-scale experiment on real equipment.

Ø Eight Bells plans to invest in studying network services and their interactions, therefore it had access to cutting edge technology environments for transport layer applications.

Ø Our analysts found into the Fed4FIRE+ experimentation testbed an SDN-capable platform with diverse set of capabilities and experimentation tools.

11

Benefits gained for Eight Bells

WWW.FED4FIRE.EU12

Business ImpactØ 8BELLS is a start-up company specializing in modelling and analysis

for businesses as well as in selected parts of Information ICT, based in Nicosia, Cyprus.

Ø The company has been established recently by ICT researchers and financial analysts pursuing the application of their research expertise and innovations in the ICT related arena.

Ø 8BELLS delivers customizable solutions that enhance modern communications relevant to the area of 5G Mobile Technology, NFV and also management solutions for Cloud infrastructures.

Ø 8BELLS translates business data into financial models, providing forecasts and supporting decision makers.

WWW.FED4FIRE.EU

Ø Eight Bells plans to provide evaluation reports and research studies based on the outcomes of Go-Quick.

Ø The company cooperates with network operators and other ICT stakeholders that will be interested on the experimental results.

Ø The innovative approach of Go-Quick experiment will consist a basis for a technical modelling framework in the existing company’s portfolio.

13

Business Impact

WWW.FED4FIRE.EU

Ø OFELIA testbed proved powerful enough to support the required VMs.

Ø Also flexible enough to adapt to the diverse topology demands, during the experiment.

Ø The support of the Patron (i2CAT) was valuable for the smooth execution of the experiment.

Ø Fruitful discussions and exchange of ideas took place during past meetings of Fed4FIRE.

14

Feedback to Fed4FIRE

This project has received funding from the European Union’s Horizon2020 research and innovation programme, which is co-funded by theEuropean Commission and the Swiss State Secretariat for Education,Research and Innovation, under grant agreement No 732638.

WWW.FED4FIRE.EU

www.8bellsresearch.com