High Capacity Network Link Emulation Using Network Processors
Abhijeet A. Joglekar
School of Computing
University of Utah, Salt Lake City, UT 84112
abhijeet@cs.utah.edu
Master's Thesis
May, 2004
Abstract
Network link emulation constitutes an important part of network
emulation, wherein links in the topology are emulated to subject the
network tra c to di erent bandwidths, latencies, packet loss
distributions, and queuing models. Increasingly, experimenters are
creating topologies with substantial emulation bandwidths; contributed
both by a large number of low-speed links and a small number of
high-speed links. It is a signi cant challenge for a link emulator to
meet this requirement in real time. Existing solutions for link
emulation use general-purpose PC-class machines; the well-understood
hardware and software PC platform make it attractive for quick
implementation and easy deployment. A PC architecture is largely
optimized for compute bound applications with large amounts of
exploitable instruction-level parallelism (ILP) and good memory
reference locality. Networking applications, on the other hand, have
little ILP and instead exhibit a coarser packet-level parallelism.
In this thesis, we propose using network processors for building high
capacity link emulators. Network processors are programmable
processors that employ a multithreaded, multiprocessor architecture to
exploit packet-level parallelism, and have instruction sets and
hardware support geared towards e cient implementation of common
networking tasks. To evaluate our proposal, we have designed and
implemented a link emulator, LinkEM, on the IXP1200 network
processor. We present the design and a mapping of LinkEM's tasks
across the multiple micro- engines and hardware threads of the
IXP1200. We also give a detailed evaluation of LinkEM, which includes
validating its emulation accuracy, and measuring its emulation
throughput and link multiplexing capacity. Our evaluation shows that
LinkEM has a factor of between 1.6 and 4.6 higher throughput for small
packets, and link multiplexing capacity between 1.4 and 2.6 higher for
low bandwidth links than Dummynet, a popular PC based link emulator,
on a comparable PC platform.
The full thesis is available in
gzip'ed Postscript
and PDF formats.