The trade-off between implicit and explicit data distribution in shared-memory programming paradigms

This paper explores previously established and novel methods for scaling the performance of OpenMP on NUMA architectures. The spectrum of methods under investigation includes OS-level automatic page placement algorithms, dynamic page migrationd manual data distribution. The trade-off that these methods face lies between performance and programming effort. Automatic page placement algorithms are transparent to the programmer, but may compromise memory access locality. Dynamic page migration is also transparent, but requires careful engineering of online algorithms to be effective. Manual data distribution on the other requires substantial programming effort and architecture-specific extensions to OpenMP, but may localize memory accesses in a nearly optimal manner. The main contributions of the paper are: a classification of application characteristics, which identifies clearly the conditions under which transparent methods are both capable and sufficient for optimizing memory locality in an OpenMP program; and the use of two novel runtime techniques, runtime data distribution based on memory access traces and affinity scheduling with iteration schedule reuse, as competitive substitutes of manual data distribution in several important classes of applications.

[1]  D. Lenoski,et al.  The SGI Origin: A ccnuma Highly Scalable Server , 1997, Conference Proceedings. The 24th Annual International Symposium on Computer Architecture.

[2]  Joel H. Saltz,et al.  Runtime compilation techniques for data partitioning and communication schedule reuse , 1993, Supercomputing '93. Proceedings.

[3]  Michael Frumkin,et al.  Implementation of NAS Parallel Benchmarks in High Performance Fortran , 2000 .

[4]  Eduard Ayguadé,et al.  A case for user-level dynamic page migration , 2000, ICS '00.

[5]  Eduard Ayguadé,et al.  Is Data Distribution Necessary in OpenMP? , 2000, ACM/IEEE SC 2000 Conference (SC'00).

[6]  H. Jin,et al.  - 3-The OpenMP Implementation of NAS Parallel Benchmarks and Its Performance , 1999 .

[7]  Evangelos P. Markatos,et al.  Using processor affinity in loop scheduling on shared-memory multiprocessors , 1992, Supercomputing '92.

[8]  Hugh Garraway.  Parallel Computer Architecture: A Hardware/Software Approach , 1999, IEEE Concurrency.

[9]  Nenad Nedeljkovic,et al.  Data distribution support on distributed shared memory multiprocessors , 1997, PLDI '97.

[10]  Joel H. Saltz,et al.  Run-time parallelization and scheduling of loops. Contract report , 1988 .

[11]  Multiprocessors.  Using Processor A � nity in Loop Scheduling on Shared Memory , 1994 .

[12]  Joel H. Saltz,et al.  Run-Time Parallelization and Scheduling of Loops , 1991, IEEE Trans. Computers.

[13]  Eduard Ayguadé,et al.  UPMLIB: A Runtime System for Tuning the Memory Performance of OpenMP Programs on Scalable Shared-Memory Multiprocessors , 2000, LCR.

[14]  Leonid Oliker,et al.  A Comparison of Three Programming Models for Adaptive Applications on the Origin2000 , 2000, ACM/IEEE SC 2000 Conference (SC'00).

[15]  Ricardo Bianchini,et al.  Using simple page placement policies to reduce the cost of cache fills in coherent shared-memory systems , 1995, Proceedings of 9th International Parallel Processing Symposium.

[16]  John S. Keen,et al.  Measuring Memory Hierarchy Performance of Cache-Coherent Multiprocessors Using Micro Benchmarks , 1997, ACM/IEEE SC 1997 Conference (SC'97).

[17]  Siegfried Benkner,et al.  Exploiting Data Locality on Scalable Shared Memory Machines with Data Parallel Programs , 2000, Euro-Par.

[18]  Jonathan Harris,et al.  Extending OpenMP For NUMA Machines , 2000, ACM/IEEE SC 2000 Conference (SC'00).