Dr. Dobb's Journal - June 2008 - (Page 42)

d06keas_p4ds 4/10/08 1:48 PM Page 42 State of the Art PERFORMANCE PORTABLE C++ continued from page 40 Comparing different bar colorings to determine the fastest compiler won’t work. When interpreting results, you should pick two bar colorings, then compare how those colors interact across all four memory layouts. For example, when looking at Figure 3, you can compare the Core2/pathscale3.0 results to the Power5/xlC7.0 results to see that the fastest memory layout for one configuration is the slowest for the other. The performance difference here is as much as 13 percent. By using performance portable array classes, you can get the best performance in both environments. Figure 4 is also interesting. Looking at the graph, the most efficient arrangement of data is to implement the Point class as individMnemonic Processor Compiler Intel GNU PGI Pathscale v10.0 v4.1.1 v6.2-3 v3.0 ual arrays, and the Brick connectivity class as an array of structs. Compare this to the layout where Point and Brick are both implemented as arrays of structs. The performance degradation for using a Struct-like Point class with the icpc v10.0 compiler on the Itanium2 is about 36 percent. This is surprising because many scientific applications use Struct-like Point classes almost exclusively. Also, the graphs show that compiling a given benchmark with the same compiler but on two different architectures gives different results. Looking at Figure 2, a comparison of the Core2/PGI6.2-3 with the Opteron/PGI6.2-3 shows a performance Compiler Options -fast -O3 -mfpmath=sse -static -fast -Mipa=fast,inline -Bstatic -Ofast -static -O3 -mfpmath=sse -static -fast -Mipa=fast,inline -Bstatic -Ofast -static -fast -O3 -static -O5 -qnoeh -O3 -static Core2/icpc10.0 Core2/g++4.1.1 Intel Xeon Core2/PGI6.2-3 E5345 Core2/path3.0 Opteron/g++4.1.1 AMD Opteron/PGI6.2-3 8216 Opteron/path3.0 Itanium2/icpc10.0 Intel Itanium2/g++3.4.4 Power5/xlC7.0 IBM Power5/g++3.3.3 GNU v4.1.1 PGI v6.2-3 Pathscale v3.0 Intel GNU xlC GNU v10.0 v3.4.4 v7.0 v3.3.3 Table 1: Processor/compiler details of each benchmark. Each test was compiled on the same processor it was run on, except the Core2 runs that had to be compiled on a 2.4-GHz Intel Xeon using an E75xx chipset. Figure 2: Point class/Triangle class. 42 Dr. Dobb’s Journal l www.ddj.com l June 2008 http://www.mcobject.com http://www.imagix.com http://www.mcobject.com/enterprise http://www.ddj.com

Table of Contents Feed for the Digital Edition of Dr. Dobb's Journal - June 2008

Dr. Dobb's Journal - June 2008 Contents Friday Night Fish Fry Alia Vox Developer Diaries There Must Be Contest Conversations Building a Test Harness for RTOS QT and Windows CE Software to Hardware Parallelization Performance Portable C++ Effective Concurrency The Agile Edge Swaine's Flames

Dr. Dobb's Journal - June 2008

For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.