1st DFG/GENI
Doctoral Consortium,
San Juan, PR
March 13th-15th, 2011
GENI Experiments in Optimizing Network
Environments using XSP
Ezra Kissel and Martin Swany
University of Delaware
Abstract
Research Objectives
Our proposal is to build, deploy and study an Optimizing
Network Environment (ONE) based on our eXtensible
Session Protocol (XSP) within the NSF's Global
Environment for Network Innovation (GENI.) Our approach
for this work is based on a new layer of protocols and
services that exists architecturally atop the current Internet
architecture, while incorporating support for emerging
technologies like the reconfigurable networks enabled by
GMPLS and Passive Optical Networks (PONs.) To this end,
we have designed ONE, which is an architecture that
enables network adaptation using network service
gateways and more visibility and control of the underlying
infrastructure.
We intend to investigate a network service-oriented
framework on GENI that allows future experiments
based on the expressive power of our sessionbased network inlay.
A number of proposed
experiments will exercise the XSP and ONE
concept and will utilize a number of existing GENI
components as well as unique resources such as
the Supercharged Planetlab Platform (SPP).
We intend to address a number of network
research problems in this exploratory work:
• Routing, pathfinding, and intelligent resource
allocation algorithms
Initial Experiments
Phoebus-XSP “Inlay”– Providing locationindependent access to private experiment
network resources remains a challenge. We
have investigated a solution where XSP has
the ability to “stitch” various GENI
components via transparent pathfinding and
forwarding over Phoebus Gateways (PGs).
XSP establishes a communication session
with the PG currently mapping the specified
endpoint ID (e.g. RFC1918 address) allowing
external hosts the ability to access a GENI
experiment as if they were on the same
network. Our experiment has tested this
functionality across existing ProtoGENI and
Planetlab resources.
pg
• Provide location and identification separation
via XSP and session-aware network gateways
• Acting on network measurement and topology
information to optimize network performance
and access
pg
pg
UDEL
GridFTP
Utah exp.
192.168.4.0/24
headnode
PL West
PL South
GENI stitching via
XSP and Phoebus
e.g.: ‘ssh headnode’
Phoebus Gateway
(At Internet2 PoPs)
1G ION circuit
pg
Adapting to Network Conditions – Utilizing the
inlay described above, we evaluated the
ability of PGs to accelerate transfers over ION
and
ProtoGENI
backbone
resources.
Phoebus implements protocol translation and
adaptation that can significantly improve
performance over LFNs and challenged
network paths. Our early testing shows a 5060% improvement when utilizing the
Phoebus-XSP approach for common network
transfer tools and benchmarks such as
GridFTP and iperf (Figure 1).
Future
experiments will allow PGs to select optimal
paths based on prevailing network conditions.
• Investigation of scalability of the architecture,
including
forwarding
performance,
state
management and authentication
Fig 1: GridFTP and iperf transfers between UDEL and LBL.
Direct connections compared to using Phoebus-XSP inlay over
ION and Protogeni backbone tunnels.
PL East
LBL
GridFTP
• Empirical evaluation of performance benefits
over a wide variety of network conditions and
applications
ProtoGENI backbone node
(Running Phoebus Gateway)
1G experiment tunnel
Fig 2: SLaBS versus direct transfers with increasing latency. 4
parallel GridFTP streams competing for 5G bottleneck link [1].
Session Layer Burst Switching – Our previous SLaBS work [1]
demonstrates the ability of an intelligent buffer-and-burst approach
to improve transfer performance over dedicated and bottleneck
network paths. To avoid contention and increase utilization, XSP
enables framing of incoming flows into “right-sized” bursts that can
be intelligently scheduled for transmission across a network of
SLaBS devices (PGs). Figure 2 shows SLaBS performance on a
10G network testbed. Our proposed experiments aim to evaluate
SLaBS fast-path optimizations using programmable network
resources such as NetFPGAs and SPP. We intend to investigate
distributed SLaBS burst scheduling approaches across the
available dynamic backbone resources and existing PGs.
Future Work
Use of GENI Infrastructure
Current and Proposed Publications
• Extend Phoebus-XSP inlay to configure Openflow
switches and other configurable network devices
Our running experiments utilize some of the
currently available GENI resources including
Emulab PCs and backbone nodes and Planetlab
virtual PCs. We were able to achieve the ability to
create ION circuits between ProtoGENI backbone
nodes and existing PGs, which we plan to further
evaluate in future testing. We have also investigated
some of the I&M systems (LAMP and INSTTOOLS)
under development in order to instrument our
experiments and provide measurement feedback.
Our research efforts have resulted in the following
publications:
• Develop SLaBS “fast path” support using
programmable hardware platforms and RDMA
• Integrate with I&M systems and actively use
measurement feedback and topology information
• Explore optimizing wireless node and sensor mote
communications over wide area networks
As a result of this early testing, we have identified
the ability to allocate and setup a stable set of
resources as a major challenge in performing longterm experiments. Existing resources are often
being contended for among developers and other
experimenters alike, which does makes the process
exciting. We expect to take advantage of additional
unique GENI resources as they become more fully
integrated and available.
[1] Session Layer Burst Switching for High
Performance Data Movement. Kissel, E.,
Swany, M. In Proceedings of PFLDNet 2010.
The proposed research effort hopes to result in the
following theses/dissertations:
[2] Improving WAN Performance with XSP – A
Protocol for New Internet Architectures. Ezra
Kissel, PhD Dissertation, Expected 2012
Acknowledgements
Internet2 – Matt Zekauskas, Jason Zurawski
Utah Emulab – Rob Ricci, Leigh Stoller
GPO/BBN – Mark Berman, Tom Mitchell