Parallel Processing of Proxylets - Possibilities
| Jon Crowcroft | |
| Jon@cs.ucl.ac.uk | |
| inc |
| Application Level Programmable InterNetwork Environment | |
| BT funded work at IC, UCL, UTS, Lancaster and Sussex | |
| Collaborate with EU Android Project, and DARPA Radioactive Project. | |
| Platform called funnelWeb includes | ||
| Execution Environment for Proxylets | ||
| Toolbox of Proxylets | ||
| Infrastructure access – e.g. directory for proximity, neighbour discovery, policy and metrics | ||
| JVM=Sandbox;EEP=Node;Plet=program | ||
Not Programmable h/w or Packet
| We’re not doing programmable packets | ||
| We’re not looking at IEEE Programmable Hardware | ||
| Although can interface to both: | ||
| Both x-bone ISI work, and Lanode, Lancaster work have both been tried. | ||
| JVM is somewhat heavyweight awy of doing sandbox – Sussex Safetynet work will sidestep this | |
| Meanwhile, to trial things, want faster platform | |
| Java threads and Multiple JVMs can exploit parallelism – how best? |
| Services are built from multiple proxylets (routelet, transcoder, encrypter, multicaster, pgm/rlc, ntp, etc etc). | |
| Coarse grain CPU and resource management can trade off load | |
| Could look at “micro-proxylets” and threads | |
| Prior work in HIPPARCH eval benefits and pitfalls of micro and macro and intra and inter-protocol module parallelism |
| Need to look at classic tradeoffs: | |
| Memory hierarchies in cluster computers – locality and communications costs | |
| Bus/Backplane overhead | |
| Scheduler Costs (context switch etc) |
| Have cluster at UCL; one at DARPA (beowolf style – I.e. linux/pc + Gbps) |
| Have idea and code | |
| Have cluster in rack | |
| Await new staff! | |
| Need IPv6 capable JVM | |
| Its research… |
| More flexible service deployment | |
| Data plane processing for heterogeneous networks | |
| Control Plane customisation (Virtual Networks) | |
| Let researchers get back into the game… |