What is ICC and GCC?

 GCC vs. ICC comparison using PARSEC Benchmarks Abedalmuhdi Almomany, Afnan Alquraan, Lakshmy Balachandran Abstract--Our goal is to compare the impact of various compiler optimizations on program performance using two widely used state-of-the-art compiler suites: GNU C Compiler and Intel’s C/C++ Compiler using PARSEC benchmarks. Compiler optimization is the process of tuning the output of a compiler to minimize or maximize some of the attributes of an executable computer program. Optimization of a compiler can be done by turning on optimization flags. In this paper, we investigate the chances of enhancing the program performance by better utilization of the existing architectural features such as compiler optimization. Proper utilization of such architectural features would not only enhance the program performance, but also reduce the need for costly upgrades as well as the system cost under development. Index Terms -- compiler, icc, gcc, PARSEC. I. INTRODUCTION “Good, fast and cheap” is one of the major slogans for this century. This is applicable to each and every field we are dealing with. In the field of Microprocessors also, we could see that architectural researchers are immensely working to fulfill the same principle. This can be only achieved by increasing the performance of microprocessors. Hence architectural researches are mostly relying on improving the performance of microprocessors. Analysis of the performance improvement can be only done by comparing the novel with existing versions using widely accepted benchmarks. Basis of our study is a similar study [1] that conducted to compare performance of various optimizations using matrix multiplication program as the benchmark. As per our review, a detailed evaluation of the characteristics of gcc and icc compliers using PARSEC benchmark suites would certainly help the ongoing architectural researches. The GNU Compiler Collection (GCC) is a compiler system produced by the GNU Project supporting various programming languages. GCC was originally written as the compiler for the GNU operating system. GCC stands for “GNU Compiler Collection”. The GNU system was developed to be 100% free software, free in the sense that it respects the user's freedom. Then Intel C++ Compiler, also known as icc , is a group of C and C++ compilers from Intel available for OS X, Linux, Windows and Intel-based Android devices. Manuscript Received on December 2014. Abedalmuhdi Almomany, A PHD candidate in the University of Alabama in Huntsville ,he has a Master Degree in Embedded Systems from Yarmouk University - Jordan. Afnan Alquraan, has a bachelor Degree in Computer Engineering from the Yarmouk University- Jordan. Lakshmy Balachandran, MSE in Computer Engineering, University of Alabama in Huntsville, USA. We consider six binaries generated by gcc and icc compilers for the given analysis. Namely, we consider three binaries generated by gcc: (1) gcc –O0 assumes no optimization, (2) gcc –O2 assumes level 2 optimization, and (3) gcc –O3 assumes level 3, the highest optimization level. Similarly, we consider four binaries generated by icc: (4) icc – O0 assumes no optimization, (5) icc –O2 assumes level 2 optimization, (6) icc –O3 assumes level 3 optimization. II. PARSEC BENCHMARK The Princeton Application Repository for Shared-Memory Computers (PARSEC) is a benchmark suite composed of multithreaded programs. The suite focuses on emerging workloads and was designed to be representative of nextgeneration shared-memory programs for chipmultiprocessors. The benchmark suite with all its applications and input sets is available as open source free of charge. Most of the workloads inside the benchmark suite were by researchers from Intel and Princeton University. Several other Benchmark suites including SPEC OMP2001, SPLASH-2, ALP Bench, BioParallel etc were widely used before the introduction of PARSEC benchmark suites but each one has its own limitations. The lack of good benchmark suites can adversely affect the parallel architecture research as well as reduce its impact. To address this problem, PARSEC benchmark suite was created with a goal to create a suite of emerging workloads that can drive CMP research. Due to the high importance of PARSEC benchmark suites in Computer Architectural field, a lot of researches are conducting in the areas of PARSEC benchmark suites with an intention to help the ongoing parallel architecture research. III. BENCHMARKS The current version of PARSEC suite contains the following 13 programs from many different areas such as computer vision, video encoding, financial analytics, animation physics and image processing. This will help to use the PARSEC as a benchmark suite for wide variety of applications. GCC vs. ICC comparison using PARSEC Benchmarks 77 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. Table: 1 PARSEC workloads IV. EXPERIMENTAL SETUP ENVIRONMENTS
CPU info: o processor : 0-7 o vendor_id : GenuineIntel o CPU family : 6 o model : 42 o model name : Intel(R) Xeon(R) CPU E31270 @ 3.40GHz o stepping : 7 o microcode : 0x23 o CPU MHz : 1600.000 o cache size : 8192 KB
OS: Ubuntu 12.04
Linux kernal: 3.2.0-60-generic
gcc version: 4.6.3
icc version: 12.1.5
PARSEC version : 3.0
Gcc Optimizations : gcc-O2,O3 (-funroll-loops - fprefetch-loop-arrays -fpermissive -fnoexceptions)
Icc Optimizations : icc-O2,O3 (-funroll-loops - opt-prefetch -fpermissive -fno-exceptions) V. IMPACT OF COMPILER OPTIMIZATIONS: A CASE STUDY USING PARSEC BENCHMARKS Evaluating the impact compiler optimizations may have on benchmark program performance is a rather challenging task. A number of factors such as execution time, cache hierarchy, CPU utilization and their interactions may influence program performance. Here, we are performing a detailed study of different parameters that affect the performance of different compilers using PARSEC benchmarks. VI. EXECUTION TIME FOR THE WORKLOADS USING GCC AND ICC As mentioned the introduction, our plan is to determine program performance of the gcc and icc compilers when using common compiler optimizations. We consider six binaries generated by gcc and icc compilers for the analysis. Namely, we consider three binaries generated by gcc: (1) gcc –O0 assumes no optimization, (2) gcc –O2 assumes level 2 optimization, and (3) gcc –O3 assumes level 3, the highest optimization level. Similarly, we consider three binaries generated by icc: (4) icc – O0 assumes no optimization, (5) icc –O2 assumes level 2 optimization, (6) icc –O3 assumes level 3 optimization. Note 1: We ran the benchmark program using different inputs such as simsmall, simlarge and native. We observed a proportional pattern in the generated output for all the above mentioned inputs. Hence we are using the output pattern of native to explain our results. Note 2: We conducted our study for different number of threads namely 1, 2, 4 and 8. A detailed evaluation of the results for different number of threads shows that there is a proportional behavior in the output for these different numbers of threads. Hence we are explaining the pattern of ‘8’ number of threads in detail. Note 3: Both Note 1 and Note 2 are applicable to all the results that we are explain here. Table: 2 Illustration - Execution Time of different threads for Blackscholes (native) Table: 2 explains the execution time of blackscholes with native inputs for 1, 2, 4 and 8 threads. As per the table values, we can observe that for 2,4 and 8 threads execution time is less for icc-O3 optimization with values 54.707s, 36.224s, 31.214s respectively. For 1 thread, the execution time is less for icc-O2 but there is only a very small minimal different between the execution time of icc-O2 and icc-O3 here. Hence we can conclude that for Blacksholes, the execution time is less for icc optimization than gcc and there is only a small difference between icc-O2 and icc-O3. Programs Application Domain Parallelization Blackscholes Financial Analysis Data-Parallel Bodytrack Computer Vision Data-Parallel Canneal Engineering Unstructured Dedup Enterprise Storage Pipeline Facesim Animation Data-Parallel Ferrret Similarity Search Pipeline Fluidanimate Animation Data-Parallel Freqmine Data Mining Data-Parallel Raytrace Rendering Data-Parallel Streamcluster Data Mining Data-Parallel Swaptions Financial Analysis Data-Parallel Vips Media Processing Data-Parallel X264 Media Processing Data-Parallel Number of Threads gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 1 197.53s 133.8230s 137.765s 175.435s 91.646s 91.743s 2 108.977s 76.481s 78.355s 97.827s 54.865s 54.707s 4 65.09s 47.887s 48.926s 59.084s 36.260s 36.224s 8 51.158s 37.591s 41.431s 47.040s 31.277s 31.214s International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-4 Issue-7, December 2014 78 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. Table: 3 Execution Time for the Programs (native-8 threads) Table: 3 explains the execution time for the all workloads in PARSEC benchmarks (except Dedup) for different optimizations. For example, we can see for workload Bodytrack, the minimum execution time is for icc-O2 and maximum is for gcc-O0. Also, for X264 the minimum execution time is for icc-O3 and maximum is for gcc-O0. The last row of Table: 3 is the total execution time of all the workloads (we have not considered Dedup workload here) in PARSEC benchmarks. The result shows that iccO2 has least overall execution time compared to all other optimizations we have considered. Speed up could be calculated from the execution time. The Table.4 below shows the speed up factor of each benchmark for different number of optimizations. The speedup factor is calculated with reference to gcc-O0 of one thread. Speed Up = (Execution time of gcc-O0) / (Execution time of gcc/icc optimized). Table: 4 Speed-Ups for different optimizations normalized to the gcc-O0 of 1 thread (native) Program 1 Thread –Speed UP 2 Threads –Speed UP gccO0 gcc-O2 gcc-O3 iccO0 icc-O2 icc-O3 gcc-O0 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 Blackscholes 1.00 1.48 1.48 1.13 2.16 3.61 1.81 2.58 2.52 2.02 3.60 3.61 Bodytrack 1.00 4.61 4.61 0.86 5.80 10.26 1.86 8.08 8.15 1.86 10.25 10.26 Canneal 1.00 1.01 1.01 0.89 0.89 1.42 1.61 1.62 1.61 1.33 1.44 1.42 Facesim 1.00 1.00 1.00 1.00 1.01 1.86 1.85 1.86 1.86 1.85 1.89 1.86 Ferrret 1.00 2.37 2.37 1.00 2.36 4.53 1.90 4.52 4.11 1.91 4.50 4.53 Fluidanimate 1.00 6.65 6.65 0.99 6.06 11.14 1.91 12.05 11.96 1.90 10.87 11.14 Freqmine 1.00 2.70 2.70 0.91 2.57 4.92 1.92 5.23 5.30 1.77 4.93 4.92 Raytrace 1.00 0.52 0.52 0.84 0.06 8.22 1.31 0.94 4.26 1.20 8.02 8.22 Streamcluster 1.00 2.71 2.71 0.99 5.08 8.93 1.95 5.03 4.90 2.01 8.96 8.93 Swaptions 1.00 2.16 2.16 1.05 3.10 6.09 1.95 4.19 4.26 2.04 6.05 6.09 Vips 1.00 0.97 0.97 1.06 1.06 0.71 1.94 1.87 1.93 2.05 2.07 0.71 X264 1.00 1.13 1.13 0.61 1.12 2.80 1.95 2.93 2.90 1.62 2.62 2.80 4 Threads -Speed UP 8 Threads –Speed UP Blackscholes 3.03 4.12 4.04 3.34 5.47 5.45 3.86 5.25 4.77 4.20 6.32 6.33 Bodytrack 3.21 12.30 12.29 2.83 16.08 15.70 3.93 16.46 17.05 3.39 20.75 20.48 Canneal 2.34 2.35 2.36 2.09 2.08 2.05 2.72 2.74 2.72 2.44 2.44 2.39 Facesim 3.08 3.10 3.09 3.09 3.08 3.10 3.49 3.52 3.48 3.52 3.52 3.51 Ferrret 3.13 7.85 7.07 3.10 7.79 7.72 3.51 9.48 8.42 3.49 9.40 9.31 Fluidanimate 3.54 20.53 20.32 3.53 18.91 19.01 4.67 23.32 23.16 4.35 22.93 22.44 Freqmine 3.67 9.73 9.93 3.35 9.20 9.23 4.65 12.15 12.39 4.36 11.91 11.72 Raytrace 1.55 1.55 5.88 1.40 8.01 8.23 1.64 2.20 6.42 1.46 8.01 8.28 Streamcluster 3.65 8.77 8.62 3.75 12.80 12.87 4.10 11.70 11.57 4.33 12.30 12.26 Swaptions 3.69 7.89 8.02 3.87 11.45 11.57 4.49 10.39 10.42 4.66 14.44 14.43 Vips 3.65 3.29 3.66 3.44 4.02 1.22 4.25 4.30 4.41 3.89 4.40 1.47 X264 3.08 5.04 4.90 2.68 4.09 4.50 3.90 5.41 5.29 3.44 5.02 5.44 Program gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 Blackscholes 51.158s 37.591s 41.431s 47.04s 31.277s 31.21s Bodytrack 156.47s 37.375s 36.089s 181.244s 29.657s 30.04s Canneal 77.417s 76.831s 77.288s 86.234s 86.119s 87.96s Facesim 88.362s 87.798s 88.7s 87.812s 87.715s 87.84s Ferrret 169.205s 62.684s 70.577s 170.239s 63.251s 63.87s Fluidanimate 374.82s 75.046s 75.564s 402.196s 76.319s 77.97s Freqmine 224.665s 85.917s 84.244s 239.321s 87.671s 89.11s Raytrace 341.335s 253.986s 87.156s 383.028s 69.929s 67.6s Streamcluster 235.31s 82.572s 83.486s 222.877s 78.485s 78.75s Swaptions 99.828s 43.122s 42.999s 96.187s 31.034s 31.04s Vips 22.034s 21.771s 21.257s 24.113s 21.315s 63.84s X264 32.357s 23.351s 23.88s 36.713s 25.128s 23.21s Total Execution Time 1872.961s 888.044s 732.671s 1977.004s 687.9S 732.4s GCC vs. ICC comparison using PARSEC Benchmarks 79 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. From the Table. 4, we could see that for Blacksholes the speed up is maximum for icc-O3 and minimum for gcc-O0. Similarly for streamcluster the speed up of icc-O2 is 12.30 times gcc- without any optimization. VII. COMPARISON OF CACHE MISSES OF WORKLOADS FOR DIFFERENT COMPILER OPTIMIZATIONS We used Perf Tool to analyze the factors like cache misses, page fault, CPU utilized etc. The cache misses counted by the perf tool is the number of memory access that cannot be served by any level of caches. Table: 5 shows the cache misses of different threads for Blackscholes (native). Table: 5 Illustration – Cache Misses of different threads for Blackscholes (native) Number of Threads gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 1 36398320 34228256 34624479 37894909 34228256 159124 2 35745956 33577988 34813256 37562361 33577988 30657717 4 40393484 36345194 40919286 39323089 36345194 30499104 8 37972212 32955782 92835740 36733608 32209138 31194108 Table: 6 Cache Misses for different optimizations normalized to the gcc-O0 (native – 8 threads ) Programs gcc-O0 gccc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 Blackscholes 1.000 0.868 2.445 0.967 0.848 0.821 Bodytrack 1.000 0.992 0.935 0.991 1.002 0.971 Canneal 1.000 0.958 0.958 0.962 0.963 1.001 Facesim 1.000 0.992 0.989 0.989 0.987 0.988 Ferrret 1.000 0.777 0.771 0.966 0.989 1.021 Fluidanimate 1.000 1.064 1.080 0.975 1.170 1.175 Freqmine 1.000 1.043 1.038 0.996 1.026 1.034 Raytrace 1.000 6.570 7.522 0.956 0.869 0.800 Streamcluster 1.000 0.966 0.970 1.002 1.016 1.018 Swaptions 1.000 0.942 1.205 1.566 0.738 0.695 Vips 1.000 0.959 0.968 0.926 0.905 0.911 X264 1.000 0.972 0.978 0.939 0.962 0.958 From Table: 6 we can understand that which optimization has least number of cache misses compared to others. For instance, workload Bodytrack has least number of cache misses in gcc-O3 while for Raytrace the cache misses are minimal for icc-O3. VIII. COMPARISON OF NUMBER OF CYCLES IN THE WORKLOADS FOR DIFFERENT COMPILER OPTIMIZATIONS The performance of a compiler also depends upon the number of cycles used for running the program. Table:8 shows the Number of Cycles for different optimizations normalized to the gcc -O0 (native – 8 threads). Table: 7 Illustration – Number of Cycles of different threads for Blackscholes (native – normalized to the gcc-O0) Number of Threads gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 1 1.000 0.677 0.695 0.889 0.464 0.464 2 1.000 0.679 0.696 0.890 0.465 0.465 4 1.000 0.686 0.698 0.889 0.465 0.465 8 1.000 0.648 0.645 0.903 0.481 0.481 International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-4 Issue-7, December 2014 80 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. Table: 8 Number of Cycles for different optimizations normalized to the gcc -O0 (native – 8 threads) Program gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 Blackscholes 1.000 0.648 0.646 0.904 0.481 0.481 Bodytrack 1.000 0.203 0.200 1.210 0.167 0.167 Canneal 1.000 0.996 0.995 1.092 1.090 1.103 Facesim 1.000 1.007 0.995 1.004 1.003 1.004 Ferrret 1.000 0.370 0.417 1.006 0.375 0.377 Fluidanimate 1.000 0.202 0.203 1.081 0.205 0.209 Freqmine 1.000 0.380 0.371 1.068 0.388 0.395 Raytrace 1.000 1.748 0.404 1.060 0.097 0.095 Streamcluster 1.000 0.351 0.351 0.951 0.336 0.337 Swaptions 1.000 1.004 1.001 2.236 0.722 0.723 Vips 1.000 0.998 0.998 0.923 0.922 0.923 X264 1.000 0.695 0.697 1.126 0.685 0.686 The overall numbers of cycles are considered as the total number of cycles for execution of all the programs (except Dedup) in the PARSEC benchmarks. The overall number of cycles is less for icc-O2 and high for icc-O0 . IX. COMPARISON OF AMOUNT OF PAGE FAULTS OBSERVED IN THE WORKLOADS FOR DIFFERENT COMPILER OPTIMIZATIONS. Using perf tool we identified the amount of page faults in different workloads for different optimizations. Table: 10 illustrate the page faults for different optimizations normalized to the gcc-O0 (native – 8 threads). As per the analysis the Overall page faults is less for icc-O3 and high for gcc-O3. Table: 9 Illustration – Page Faults of different threads for Blackscholes (native - normalized to the gcc-O0) Table: 10 Page Faults for different optimizations normalized to the gcc-O0 (native – 8 threads) Benchmark gccO0 gcccO2 gccO3 iccO0 iccO2 iccO3 Blackscholes 1.000 1.000 1.000 1.001 1.001 1.001 Bodytrack 1.000 0.998 1.013 0.989 0.993 0.997 Canneal 1.000 1.000 1.000 1.000 1.000 1.000 Facesim 1.000 1.000 1.000 1.000 1.000 1.000 Ferrret 1.000 1.004 1.022 1.207 1.287 0.948 Fluidanimate 1.000 1.005 0.995 1.003 1.000 1.000 Freqmine 1.000 0.999 0.995 0.963 0.944 0.979 Raytrace 1.000 0.999 1.001 1.000 0.995 0.995 Streamcluster 1.000 1.000 1.000 1.000 1.000 1.000 Swaptions 1.000 0.974 1.042 0.962 0.965 0.972 Vips 1.000 1.002 1.140 1.032 1.032 1.029 X264 1.000 1.000 1.000 1.000 1.000 1.000 Number of Threads gcc-00 gcc-O2 gcc-O3 icc-O0 icc-O2 icc-O3 1 1 0.99998 0.99998 1.00069 1.00070 1.00069 2 1 0.999987 0.99994 1.000658 1.00066 1.000670 4 1 1.00001 1.00002 1.00067 1.00068 1.00068 8 1 0.9999 1.00012 1.00058 1.00062 1.00063 GCC vs. ICC comparison using PARSEC Benchmark X. COMPARISON OF AMOUNT OF CPU Table: 11 CPU Utilization for different optimizations normalized to the gcc Program gcc-00 Blackscholes 1.000 Bodytrack 1.000 Canneal 1.000 Facesim 1.000 Ferrret 1.000 Fluidanimate 1.000 Freqmine 1.000 Raytrace 1.000 Streamcluster 1.000 Swaptions 1.000 Vips 1.000 X264 1.000 As per our analysis, CPU utilization is maximum for gccO0 and minimum for icc-O3. This strengthens best optimization will have the less amoun utilization. XI. COMPARISON OF COMPILE TIMES OF GCC AND ICC. We also observed the time taken to compile each in gcc and icc and calculated the overall compile time all the workloads using both. We have observed that the overall compile time is less for gcc and high for icc. Figure: 0 500 1000 GCC vs. ICC comparison using PARSEC Benchmarks 81 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. CPU UTILIZATIONS OF WORKLOADS FOR DIFFERENT COMPILER OPTIMIZATIO e: 11 CPU Utilization for different optimizations normalized to the gcc-O0(native gcc-O2 gcc-O3 icc-O0 0.854 0.782 0.977 0.907 0.909 1.000 1.000 1.015 0.995 0.999 0.999 0.999 0.998 0.998 0.882 1.000 1.008 1.000 0.991 0.988 1.000 1.000 1.592 0.941 0.996 0.986 1.000 1.000 1.002 1.001 0.980 1.000 0.947 0.997 0.995 0.991 utilization is maximum for gccO0 This strengthens the fact that, the the less amount of CPU TIMES OF GCC AND We also observed the time taken to compile each workload in gcc and icc and calculated the overall compile time for We have observed that the ime is less for gcc and high for icc. Table: 12 Compilation Time of gcc v/s icc Program Blackscholes Bodytrack Canneal Dedup Faceim Ferrret Fluidanimate Freqmine Raytrace Treamcluter Swaptions Vips X264 Total Compilation Time in seconds Figure: 1 Compilation Time of gcc v/s icc Ferrret Freq … Strea … Vips Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. COMPILER OPTIMIZATIONS. O0(native – 8 threads) icc-O2 icc-O3 0.771 0.465 0.936 0.936 0.995 0.988 1.000 1.000 0.999 0.999 1.000 1.003 0.994 0.991 0.452 0.452 1.004 1.004 1.002 1.000 0.898 0.957 0.960 0.979 Table: 12 Compilation Time of gcc v/s icc gcc Icc 0.759 1.160 35.198 49.970 1.721 1.126 0.574 0.973 125.177 97.144 208.902 313.661 1.960 2.864 3.545 4.594 464.526 560.357 1.7103s 2.202 2.317 3.419 232.099 400.576 41.739 73.485 Total Compilation Time 1120.215 1511.531 gcc icc International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-4 Issue-7, December 2014 82 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Pvt. Ltd. XII. CONCLUSION In this project we analyzed the impact of various compiler optimizations on program performance using two widely used state-of-the-art compiler suites: GNU C Compiler and Intel’s C/C++ Compiler using PARSEC benchmark suites. The results indicate that Intel’s compiler with optimization (O2, O3) in general outperforms GNU C compiler in almost all the parameters we observed. Speed up is the factor we primarily taken into account. We have observed that the overall speed up is best for icc-02 for 1, 2 4 and 8 threads. The overall speed up with icc-O2 is almost 3 times compared to the speed up of gcc without any optimization. Table: 13 Total Speed-Ups all the programs normalized to the gcc-O0 (native). Number of Threads (native) Best for /Speed Up Worst for /Speed UP 1 icc-O2 / 3.01 icc-O0/ . 95 2 icc-O2 / 3.27 icc-O0/ .974 4 icc-O2 / 2.8 icc-O0/ .95 8 icc-O2 / 2.72 icc-O0/ .94 Also, the experimental results help to identify the best and worse optimizations for each parameter we have observed. Table: 14 shows the best and worse optimizations for different parameters for the overall PARSEC benchmark suite. Table:14 Best and worse optimization for different parameters Parameters Best Worse Over-all Speed up icc -O2 icc -O0 Over-all Cache Misses icc-O3 gcc-O3 Over-all Number of Cycles icc-O2 icc-O0 Over-all page faults icc-O3 gcc-O3 Total Compilation Time gcc icc XIII. ACKNOWLEDGEMENT We would like to thank Dr. Aleksandar Milenkovic in University of Alabama in Huntsville, USA for giving complete support and insightful comments about the study. REFERENCES [1] Compiler Optimizations in Multi-core Era: A Performance Study Using Intel C++/Fortran and GNU C++/Gfortran , Aleksandar Milenkovic LaCASA Laboratory, Electrical and Computer Engineering The University of Alabama in Huntsville. [2] Christian Bienia, Sanjeev Kumar, Jaswinder Pal Singh and Kai Li. The PARSEC Benchmark Suite: Characterization and Architectural Implications. In Proceedings of the 17th International Conference on Parallel Architectures and Compilation Techniques, October 2008. [3] Christian Bienia, Sanjeev Kumar and Kai Li. PARSEC vs. SPLASH-2: A Quantitative Comparison of Two Multithreaded Benchmark Suites on Chip-Multiprocessors. In Proceedings of the 2008 Annual IEEE International Symposium on Workload Characterization, September 2008. [4] http://parsec.cs.princeton.edu/download/tutorial/3.0/parsectutorial.pdf [5] J. L. Hennessy and D. A. Patterson, Computer Architecture: A Quantitative Approach. San Francisco: Morgan Kaufmann, 2007. Abedalmuhdi Almomany, A PHD candidate in the University of Alabama in Huntsville ,he has a Master Degree in Embedded Systems from Yarmouk University - Jordan. Has a published paper “Real time Radio Frequency Identification Vehicles Data logger Traffic Management System”. Current working focus on parallel programming using CUDA and OpenCL and parallel processing. Afnan Alquraan, has a bachelor degree in computer engineering from the Yarmouk University- Jordan, her working focus on programming languages and networking. Lakshmy Balachandran, MSE in Computer Engineering, University of Alabama in Huntsville, USA. Major research work includes Data Aggregation in Wireless Sensor Network (WSN) using STDCA method, Application of Digital Image Processing in Drug Industry, MIL-STD 1553B ENCODER AND DECODER..