Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 1 Based upon slides from Jay Lepreau, Utah Emulab Introduction Shiv Kalyanaraman Google: “Shiv RPI”

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 2 Based upon slides from Jay Lepreau, Utah Recall: Simulation vs Measurement/Implementation q 1. Real system is more credible, but more complex – lot of auxiliary concerns & murphy’s law strikes often! q But simulation must be thought of as a first step to real implementation q (I.e. to get a stable design that must be validated by implementation and/or analysis) q 2. Measurement of Internet traffic may not be the same as measurement tomorrow (real, but still random samples!) q Representative measurement traces can be used to drive simulation (I.e. trace-driven simulation) q 3. New emulation platforms: Utah’s emulab (next slide) q Takes out the configuration complexity from small/medium sized real experiments! q 4. Bottom line: mix and match both tools depending upon the problem at hand

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 3 Based upon slides from Jay Lepreau, Utah Utah Emulab and Click: Emulation and Modular Implementation Platforms Utah’s Emulab Testbed: control & interconnect the kernels of 100s of machines just by using ns-2 scripting!!! MIT’s Click Modular Router On Linux: Forwarding Plane Implns

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 4 Based upon slides from Jay Lepreau, Utah This is done! You just write ns-2 scripts!

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 5 Based upon slides from Jay Lepreau, Utah What is Emulab? q A configurable Internet emulator in a room q Today: 200 nodes, 500 wires, 2x BFS (switch) q virtualizable topology, links, software q Bare hardware with lots of tools q An instrument for experimental CS research q Universally available to any remote experimenter q Simple to use: just write ns-2 scripts

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 6 Based upon slides from Jay Lepreau, Utah Why? q “We evaluated our system on five nodes.” -job talk from university with 300-node cluster q “We evaluated our Web proxy design with 10 clients on 100Mbit ethernet.” q “Simulation results indicate...” q “Memory and CPU demands on the individual nodes were not measured, but we believe will be modest.” q “The authors ignore interrupt handling overhead in their evaluation, which likely dominates all other costs.” q “Resource control remains an open problem.”

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 7 Based upon slides from Jay Lepreau, Utah Why 2 q “You have to know the right people to get access to the cluster.” q “The cluster is hard to use.” q “ runs FreeBSD 2.2.x.” q “October’s schedule for is…” q “ is tunneled through the Internet”

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 8 Based upon slides from Jay Lepreau, Utah Key Design Aspects q Allow experimenter complete control q … but provide fast tools for common cases q OS’s, disk loading, state mgmt tools, IP, traffic generation, batch,... q Virtualization q of all experimenter-visible resources q node names, network interface names, network addresses q Allows swapin/swapout

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 9 Based upon slides from Jay Lepreau, Utah Design Aspects (cont’d) q Flexible, extensible, powerful allocation algorithm q Persistent state maintenance: q none on nodes q all in database q leverage node boot time: only known state! q Separate control network q Familiar, powerful, extensible configuration language: ns

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 10 Based upon slides from Jay Lepreau, Utah Some Unique Characteristics q User-configurable control of “physical” characteristics: shaping of link latency/bandwidth/drops/errors (via invisibly interposed “shaping nodes”), router processing power, buffer space, … q Node breakdown (old): q 40 core, 160 edge

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 11 Based upon slides from Jay Lepreau, Utah More Unique Characteristics q Capture of low-level node behavior such as interrupt load and memory bandwidth q User-replaceable node OS software q User-configurable physical link topology q Completely configurable and usable by external researchers, including node power cycling

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 12 Based upon slides from Jay Lepreau, Utah An “Experiment” q emulab’s central operational entity q Directly generated by an ns script, q … then represented entirely by database state q Steps: Web, compile ns script, map, allocate, provide access, assign IP addrs, host names, configure VLANs, load disks, reboot, configure OS’s, run, report

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 13 Based upon slides from Jay Lepreau, Utah Mapping Example

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 14 Based upon slides from Jay Lepreau, Utah Automatic mapping of desired topologies and characteristics to physical resources q Algorithm goals: q minimize likelihood of experimental artifacts (bottlenecks) q “optimal” packing of multiple simultaneous experiments q Extensible for heterogenous hardware, software, new features q Randomized heuristic algorithm: simulated annealing q May move to genetic algorithm

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 15 Based upon slides from Jay Lepreau, Utah Virtual Topology

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 16 Based upon slides from Jay Lepreau, Utah Mapping into Physical Topology

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 17 Based upon slides from Jay Lepreau, Utah For more info…. q Emulab website: