前回の実験の続きで最適なブロックサイズについて。さらに大きな問題で測定できないので何とも言えないが、少なくとも nb=1024 よりも大きい方が良いのは確か。
◯ソフトウェア SDPARA 7.6.0
◯問題 tai18a.dat-s
47142 -= mDIM
2 = nBLOCK
-69946 290 = bLOCKsTRUCT
[gpdpotrf] ### END n=47142, nb=1024, 1x2 procs, ver 40: 32.312202sec --> 1080.777267GFlops ###
[gpdpotrf] ### END n=47142, nb=1280, 1x2 procs, ver 40: 28.777453sec --> 1213.529672GFlops ###
[gpdpotrf] ### END n=47142, nb=1536, 1x2 procs, ver 40: 26.506123sec --> 1317.517965GFlops ###
[gpdpotrf] ### END n=47142, nb=1792, 1x2 procs, ver 40: 27.553626sec --> 1267.430038GFlops ###
[gpdpotrf] ### END n=47142, nb=2048, 1x2 procs, ver 40: 27.989612sec --> 1247.687649GFlops ###
![](https://blogimg.goo.ne.jp/thumbnail/7d/13/b48e7f053cf2c8d7adeecbb9070465f6_s.jpg)
◯計算サーバ
CPU : Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz x 1
メモリ : 32GB
GeForce GTX TITAN x 2枚
![](https://blogimg.goo.ne.jp/thumbnail/7e/93/bc5d91d20a1ca1f1e2db96c9d31df479_s.jpg)
◯ソフトウェア SDPARA 7.6.0
◯問題 tai18a.dat-s
47142 -= mDIM
2 = nBLOCK
-69946 290 = bLOCKsTRUCT
[gpdpotrf] ### END n=47142, nb=1024, 1x2 procs, ver 40: 32.312202sec --> 1080.777267GFlops ###
[gpdpotrf] ### END n=47142, nb=1280, 1x2 procs, ver 40: 28.777453sec --> 1213.529672GFlops ###
[gpdpotrf] ### END n=47142, nb=1536, 1x2 procs, ver 40: 26.506123sec --> 1317.517965GFlops ###
[gpdpotrf] ### END n=47142, nb=1792, 1x2 procs, ver 40: 27.553626sec --> 1267.430038GFlops ###
[gpdpotrf] ### END n=47142, nb=2048, 1x2 procs, ver 40: 27.989612sec --> 1247.687649GFlops ###
![](https://blogimg.goo.ne.jp/thumbnail/7d/13/b48e7f053cf2c8d7adeecbb9070465f6_s.jpg)
◯計算サーバ
CPU : Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz x 1
メモリ : 32GB
GeForce GTX TITAN x 2枚
![](https://blogimg.goo.ne.jp/thumbnail/7e/93/bc5d91d20a1ca1f1e2db96c9d31df479_s.jpg)