J-machine
Problem:
* how can a massively parallel machine be built out of simple nodes
o need a good interconnect
o need efficeint and general communication syncronization/communication
o explore a different execution model
* alternative to shared or distributed memory on a SPMD
* would like mechanisms that are implementable on a simple CPU
* would like a processor that is both cheap and universal
Users:
* High Performance Computing (HPC)
o Most HPC applications require floating point but J-machine did not support floating point
o The J-machine is not necessarily usable for HPC but it is enough to demonstrate mechanisms that are useful for HPC
* Readers are architects
Uniqueness:
* Very high performance interconnect
* Packaging enabled high performance interconnect
* Message driven processing
o similar to event driven processing at each node
* Active messages
o messages contain a handler and data
o both high and regular priority messages
o high priority messages may suspend regular priority messages
* Tagged memory
o full empty bits (for producer/consumer data passing)
o futures
* Support for global address space
Evaluation:
* used both micro and macro benchmarks
o micro-benchmarks used for interconnect, messages, and syncronization evaluation
o macro-benchmarks used for relevent algorithms
* they actually built the machine
* lacked application comparisons to other machines (this was done in a later paper).
