Generalized Data Transfers At Memory Bandwidth

Peter August Dinda

CMU SCS

Generalized Data Transfers At Memory Bandwidth

Generalized Data Transfers

Address Relations

Send/Recv Implementation

Storing Address Relations

Inspector/Executor [Salz, et al]

Context: Array Assignments

Distributed Arrays

Representative Assignments

Representing Address Relations

AAPAIR: Simple Representation

AABLK: Run-length Encoding

DMRLE: Handling Strides

DMRLEC: Repeated Strides

Address Relation Storage Costs

Copying & Superscalar Plateau

Paragon: No Superscalar Plat.

Pentium 90: Clear Plateau

DEC 3K/400a: Complex Plateau

Measurement Details

Measurement Testcases

Performance: DEC 3K/400a

Performance:IBM 250 (PPC601)

Performance: IBM SP2 (PWR2)

Performance: Paragon

Performance: Pentium 90

Performance: Pentium 133

Conclusions

Copying with Address Relations

A Simple Copy Engine