180 likes | 320 Views
PARSEC vs. SPLASH-2 :. A Quantitative Comparison of Two Multithreaded Benchmark Suites on Chip-Multiprocessors. 발표자 이보선. INTRODUCTION (1/2). Princeton Application Repository for Shared-Memory Computers What distinguishes PARSEC from other benchmark suites?
E N D
PARSEC vs. SPLASH-2 : A Quantitative Comparison of Two Multithreaded Benchmark Suites on Chip-Multiprocessors 발표자 이보선
INTRODUCTION (1/2) • Princeton Application Repository for Shared-Memory Computers • What distinguishes PARSEC from other benchmark suites? • SPLASH-2, SPEC OMP2001 - focus on High-Performance Computing • BioParallel - bioinformatics programs • ALPBench - suite of multimedia workloads • MineBench - data mining. • PARSEC vs. SPLASH-2
INTRODUCTION (2/2) • This paper makes three contributions • We compare SPLASH-2 with PARSEC to determine how much the program selections of the two suites overlap. Significant differences exist that justify an overhaul of the popular SPLASH-2 benchmark suite. • We identify workloads in both suites that resemble each other that can help researcher to interpret results. A few benchmarks of the two suites have similar characteristics. • We demonstrate how current technology trends are changing programs. The direct comparison of the PARSEC suite with SPLASH-2 shows that the proliferation of CMPs and the massive growth of data have a measurable impact on workload behavior.
OVERVIEW • The SPLASH-2 suite • one of the most widely used collections of multithreaded workloads • released at the beginning of the 90s • High-Performance Computing domain • PARSEC • released at the beginning of 2008 • PARSEC has the following five main features • Multithreaded Applications • Emerging Workloads • Diverse • Employ • Support Research
METHODOLOGY • Execution-driven simulation to obtain the relevant data • Standard statistical method to compute the similarity of the workloads
Experimental Setup & Removing Correlated Data • Simulate abstract cache hierarchy with CMP$im • Compute similarity with hierarchical clustering • Visualize results with dendrograms and scatter plots • Correlated characteristics can skew the redundancy analysis • It is therefore necessary to eliminate correlated information with Principal Component Analysis(PCA)
Measuring Similarity • The Euclidean distance between the program characteristics is a measure for the similarity of the programs • Hierarchical clustering works as follows: • 1. Assign each workload to its own cluster • 2. Compute the pair-wise distances of all clusters • 3. Merge the two clusters with the smallest distance • 4. Repeat steps 2 - 3 until only a single cluster is left
REDUNDANCY ANALYSIS RESULTS (1/2) • How much do the two program collections overlap? • In particular, which workloads of the PARSEC suite resemble which SPLASH-2 codes? • Which benchmark suite is more diverse?
REDUNDANCY ANALYSIS RESULTS (1/2) • Diversity by computing the total • SPLASH-2: 19.55, PARSEC: 18.98 • Direct comparison • the PARSEC suite Contains significantly more diversity than SPLASH-2
Multiple Differences • Instruction Mix Differences • Working Set Differences • Sharing Behavior Differences • No single source for the differences of the two suites.
Inclusion of Pipeline Model • Another difference between SPLASH-2 and PARSEC is the inclusion of workloads that employ the pipeline programming model in PARSEC.
Data Growth • World data is currently doubling every three years • These workloads employ models that allow them to have a basic understanding of the data they process. • For example, the bodytrackprogram employs a model of the human body to detect a person being shown in multiple video streams. • The compressed archive that contains the whole suite with all inputs is 16 MB in the case of SPLASH-2. • For PARSEC, it is 2.7 GB.
Large Working Sets More Common • For smaller caches PARSEC workloads have a significantly higher average miss rate. • The difference is 0.26% for a 1 MB cache, approximately one fourth more. • It decreases to 0.11% for an 8 MB cache. • SPLASH-2 workloads have an average miss rate 0.02% higher than PARSEC workloads if 16 MB caches are used. • This trend continues to the end of the spectrum of cache sizes.
CONCLUSIONS • PARSEC is the more diverse suite in direct comparison