PLINK 1.0’s basic association analysis commands were quite flexible, but the powerful max(T) permutation test suffered from poor performance. PRESTO [42] and PERMORY introduced major algorithmic improvements (including bit population count) which largely solved the problem. Table 6 demonstrates that PLINK 1.9 successfully extends the PERMORY algorithm to the full range of PLINK 1.0’s association analyses, while making Fisher’s exact test practical to use in permutation tests. (There is no 64-bit Windows PERMORY build, so the comparisons on the Win64-2 machine are between 64-bit PLINK and 32-bit PERMORY.)Table 6 Association analysis max(T) permutation test times (sec) Other parameter(s)DatasetMachinePLINK 1.07PERMORY 1.1PLINK 1.90Ratio–trend (C/C)synth1Mac-2∼20 k18.71.1 kMac-12∼16 k2.85.7 kLinux32-2∼21 k65.0320Linux64-512∼17 k285.02.8Win32-2∼35 k1444.261.5Win64-2∼25 k889.714.4synth2Mac-2∼170 k42.44.0 kMac-12∼180 k6.428 kLinux32-2∼410 k391.01.0 kLinux64-512∼200 k580.913.7Win32-2∼1100 k2362.5198.0Win64-2∼370 k1423.634.0–fisher (C/C)synth1Mac-2∼150 k21.96.9 kMac-12∼150 k3.741 kLinux32-2∼170 k57.82.9 kLinux64-512∼120 k3.435 kWin32-2∼440 k64.96.8 kWin64-2∼200 k22.09.1 ksynth2Mac-2∼890 k49.818 kMac-12∼690 k7.691 kLinux32-2∼1300 k393.73.3 kLinux64-512∼720 k13.055 kWin32-2∼3600 k208.317 kWin64-2∼1700 k35.648 k–assoc (QT)synth1Mac-2∼30 k148.0200Mac-12∼22 k22.6970Linux32-2∼68 k847.280Linux64-512∼29 k29.2990Win32-2∼58 k896.165Win64-2∼36 k264.2140–assoc lin (QT)synth1Mac-2606.8Mac-1234.7Linux32-23212.6Linux64-5121259.846.427.2Win32-22115.73062.70.69Win64-2972.6336.62.89All runs are with 10000 permutations and –seed 1. The PLINK 1.9 implementation extends Pahl et al.’s PERMORY algorithm [29] with multithreading (note the 12- and 512-core machine results) and additional low-level optimizations.
