Latency Hiding

Here, we show how the latency hiding mechanism described in Section 3.2 is implemented at the meta-level. The implementation consists of two parts: invoking methods to prefetch the arguments before their usage in an expression, and substitution of results of the prefetch where needed. Let us review the annotation syntax that requests prefetch method invocations, which is proposed in Section 3.2:

19#19

To perform method invocations in e_a, we define two evaluator classes prefetch-eval and prefetch-anno-eval, and a method for each--definitions are shown in Figure 10.

  

Figure 10: Meta-level program for latency hiding (1)

20#20

Expressions except for the prefetch method invocations in the annotation are evaluated as usual. This is implicitly achieved by the delegation mechanism. The latter part is for using prefetched values; i.e., do touch operations instead of method invocation. This is achieved by defining a method do-method-call of the class prefetch-eval (Figure 11). For each base-level method invocation, this user-defined do-method-call checks if the invocation is already performed in an annotation. If so, it touches a corresponding reply box, instead of actual invocation.

  

Figure 11: Meta-level program for latency hiding (2)

21#21