There exists an algorithm for exact summation of a set of floating-point values, famously available in Python. If you use this algorithm for computing both linear and parallel mode, the two of them will show the exact same totals: they will both compute exact sums of your set of values.
This algorithm requires dynamic allocation and may not be appropriate depending on your context, but it is the only solution I see with the information you have provided. Other approaches may be possible if reduced precision is acceptable (one obvious one being to convert all values to fixed-point before including them into the sum. Fixed-point addition, unlike floating-point addition, is associative).