The CrOS Project

  The CrOS project from CalTech [24] uses a crystalline programming framework. The authors present an MIMD model with nodes calling library functions to communicate with each other. The CrOS target hardware did not have a separate router, and accordingly the processors themselves performed the communication over dedicated point-to-point links. Their loosely synchronous model requires that a node cannot write to another node unless the destination node is prepared to receive the message; similar in spirit, at the processor level, to scheduled routing. Since CrOS was purely run-time, they did not support any notion of phases, or any distinction between compile-time and run-time values.

  

Figure 3.4: CrOS code for a parallel prefix

However, CrOS does include a notion of operators similar to that used in COP. In addition to simple stream-style connections, they include primitives for doing transpose and various other array-style communications. A fragment of a simple sum-of-squares program is presented in Figure 3.4; the vm_combine() function is a CrOS operator that does a parallel prefix, or scan, across the nodes in the machine.

CrOS supports a version of online routing called the crystal_router, which handles online routing in a manner similar to that used by COP. Without the ability to extract information at compile-time, however, they do an `inspector'-style step where routing information is collected in the first phase, then subsequent phases use the collected information to avoid having to examine all possible streams for input.