關(guān)于在SMP模式下同樣的輸入文件每次提交計(jì)算結(jié)果不同問題的說明

2016-11-15 by:CAE仿真在線來源:互聯(lián)網(wǎng)

剛才看到一個(gè)帖子,講了在SMP模式下每次提交計(jì)算結(jié)果不同的情況,我看了一下dyna手冊(cè)中的×CONTROL_PARALLEL關(guān)鍵字,特將手冊(cè)翻譯如下,翻譯不準(zhǔn)的地方請(qǐng)同志們指正。
對(duì)于任何給定的問題,當(dāng)結(jié)果一致性選項(xiàng)關(guān)閉時(shí)(此選項(xiàng)在*control_parallel中),如CONST=2,同樣的任務(wù)在不同的時(shí)間提交,即使CPU數(shù)是一樣的或者CPU數(shù)是不一樣的,我們會(huì)看到結(jié)果會(huì)有細(xì)微的差別,對(duì)照節(jié)點(diǎn)加速度常常會(huì)顯示出較大范圍的差異;然而,需要指出的是,加速度計(jì)結(jié)果常常顯示出不明顯的變化是因?yàn)榧铀俣扔?jì)的平滑效應(yīng)通常是與剛體節(jié)點(diǎn)相聯(lián)系的,精度問題,并不是數(shù)字模擬中的新問題,而是碰撞仿真的固有問題,因?yàn)樵谶@一類問題中在壓縮載荷的作用下結(jié)構(gòu)的分歧是很普遍的。這個(gè)問題可以很輕易的用一個(gè)完美的截面為四邊形的薄壁管來演示,在一個(gè)壓縮載荷下,典型情況是,每次在一個(gè)CPU上運(yùn)行,且每次只做一個(gè)很小的修改(如單元或者沙漏公式),每次的運(yùn)行結(jié)果都會(huì)有很多的不同,這一點(diǎn)可以從管子的最終形狀上看出,同樣的,如果同樣的問題運(yùn)行在不同品牌的電腦上也會(huì)出現(xiàn)。如果同樣的問題在多處理器及其上提交時(shí),即使提交計(jì)算的文件沒有修改,但是所得出的結(jié)果卻變化很大。這個(gè)問題是由于隨機(jī)的數(shù)字截?cái)?或者說圓整)作為“完美”梁問題的引發(fā)器,因?yàn)槊看斡?jì)算匯總(CONST=2)是按照不同的次序的,數(shù)字截?cái)嘁彩请S機(jī)的,一致性判據(jù)CONST=1提供了每次同一(或相近)結(jié)果,無論是在SMP模式下使用一個(gè)、兩個(gè)或者更多的處理器,因?yàn)樵谶@種情況下,要求所有相關(guān)的全局向量計(jì)算在一個(gè)精確的順序下而與處理器無關(guān)。當(dāng)檢測結(jié)果一致性時(shí),應(yīng)該比較節(jié)點(diǎn)位移或單元應(yīng)力。(不同批次提交結(jié)果)NODOUT和ELOUT文件中的數(shù)字應(yīng)該是一致的,然后,GLSTAT,SEFORC和許多其他ASCII文件應(yīng)該是不一致的,因?yàn)楣こ塘吭谶@些文件中總計(jì)是并行的因?yàn)樾实脑?并且總計(jì)操作的順序并不是強(qiáng)制的。使用這個(gè)選項(xiàng)的最大缺點(diǎn)是至少多15%CPU消耗,如果PARA=1并且使用2個(gè)或更多處理器時(shí)CPU消耗會(huì)少許多。除非PARA判據(jù)是打開的(對(duì)于非向量處理器),并行縮放比例是起反作用的。一致性判據(jù)不作用于MPP模式。

下面手手冊(cè)上的原文。
For any given problem with the consistency option off, i.e., CONST=2, slight differences in results are seen when running the same job multiple times with the same number of processors and also when varying the number of processors. Comparisons of nodal accelerations often show wide discrepancies; however, it is worth noting that the results of accelerometers often show insignificant variations due to the smoothing effect of the accelerometers which are generally attached to nodal rigid bodies. The accuracy issues are not new and are inherent in numerical simulations of automotive crash and impact problems where structural bifurcations under compressive loads are common. This problem can be easily demonstrated by using a perfectly square thin-walled tubular beam of uniform cross section under a compressive load. Typically, every run on one processor that includes a minor input change (i.e., element or hourglass formulation) will produces dramatically different results in terms of the final shape, and, likewise, if the same problem is again run on a different brand of computer. If the same problem is run on multiple processors the results can vary dramatically from run to run WITH NO INPUT CHANGE. The problem here is due to the randomness of numerical round-off which acts as a trigger in a “perfect” beam. Since summations with (CONST=2) occur in a different order from run to run, the round-off is also random. The consistency flag,CONST=1, provides for identical results (or nearly so) whether one, two, or more processors are used while running in the shared memory parallel (SMP) mode. This is done by requiring that all contributions to global vectors be summed in a precise order independently of the number of processors used. When checking for consistent results,nodal displacements or element stresses should be compared. The NODOUT and ELOUT files should be digit to digit identical. However, the GLSTAT, SECFORC, and
many of the other ASCII files will not be identical since the quantities in these files are
summed in parallel for efficiency reasons and the ordering of summation operations are not enforced. The biggest drawback of this option is the CPU cost penalty which is at least 15 percent if PARA=0 and is much less if PARA=1 and 2 or more processors are used. Unless the PARA flag is on (for non-vector processors), parallel scaling is
adversely affected. The consistency flag does not apply to MPP parallel.

注:在971R5.0中已經(jīng)廢棄了這個(gè)卡片,

開放分享：優(yōu)質(zhì)有限元技術(shù)文章,助你自學(xué)成才