The POOMA operating system

Conformal Computing: Algebraically connecting the hardware/software boundary using a uniform approach to high-performance computation for software and hardware applications

“…Origin 2000 Size fftpsirad2 IMSL NAG SCSL FFTW 2 3 0.190 0.064 0.010 0.065 0.013 2 4 0.018 0.061 0.010 0.047 0.013 2 5 0.018 0.062 0.010 0.065 0.014 2 6 0.017 0.116 0.011 0.073 0.013 2 7 0.019 0.063 0.010 0.068 0.015 2 8 0.018 0.062 0.011 0.105 0.014 2 9 0.017 0.122 0.011 0.069 0.014 2 10 0.021 0.065 0.013 0.056 0.015 2 11 0.021 0.064 0.016 0.058 0.017 2 12 0.021 0.067 0.023 0.067 0.023 2 13 0.022 0.075 0.036 0.065 0.030 2 14 0.024 0.144 0.065 0.066 0.051 2 15 0.030 0.120 0.135 0.110 0.082 2 16 0.040 0.209 0.296 0.080 0.189 2 17 0.065 0.335 0.696 0.072 0.395 2 18 0.126 0.829 3.205 0.075 0.774 2 19 0.238 3.007 9.538 0.096 2.186 2 20 0.442 9.673 18.40 0.143 4.611 2 21 0.884 23.36 38.93 0.260 9.191 2 22 1.910 46.70 92.75 0.396 19.19 2 23 4.014 109.4 187.7 0.671 48.69 2 24 7.550 221.1 442.7 1.396 99.10…”

“…Therefore we achieved a performance increase for every size except one, and we were able to run successfully for a vector 4,194,304 elements larger. IBM's SP2 Size fftpsirad2 ESSL FFTW 2 3 0.010 0.013 0.020 2 4 0.010 0.053 0.020 2 5 0.010 0.013 0.020 2 6 0.010 0.013 0.020 2 7 0.010 0.013 0.027 2 8 0.010 0.016 0.023 2 9 0.010 0.010 0.020 2 10 0.010 0.013 0.023 2 11 0.013 0.010 0.023 2 12 0.020 0.020 0.023 2 13 0.033 0.030 0.030 2 14 0.040 0.043 0.053 2 15 0.070 0.060 0.083 2 16 0.140 0.120 0.143 2 17 0.280 0.160 0.283 2 18 0.580 0.310 0.683 2 19 1.216 0.610 1.456 2 20 2.580 1.276 3.273 2 21 5.553 3.663 6.770 2 22 12.12 Failed 15.553 2 23 25.25 Failed Failed Table 1.2. Real Execution Times (seconds) of fftpsirad2 and ESSL on Maui HPCC IBM SP2.…”

“…The OS is IRIX 6.5. 2. an IBM SP2 at the MAUI High Performance Computing Center 4 , with 32, P2SC 160Mhz processors, and 1 GB of memory. The L1 cache size is 160KB (32KB Icache and 128KB Dcache), there is no L2 cache.…”

“…As was discussed for the situation illustrated by Eqs. 3.19 and 3.20, we form the weight vector ξ from the elements of M (4) and reshape to give: Next we carry out the reshape-transpose to give: of Eq. 3.29.…”

“…The approach is similar in spirit to other efforts using libraries of algorithm building blocks based on C++ template classes. In POOMA for example, expression templates using the Portable Expression Template Engine (PETE) (http://www.acl.lanl.gov.pete) were used to achieve efficient distribution of array indexing over scalar operations [1,2,3,4,5,6,7,8,9,10].…”