Concerned about precision...

[dawe]

Hi all,
I've been thinking recently about softwares for NGS... On the samtools mailing list I've discussed about some variables that were implemented as "long double" as the developer needed more precision than "double" numbers.
NGS systems output gigabytes of data, millions of sequences to be analyzed. Before this, software using "float" or "double" variables might be enough precise when hundreds or thousands of sequences had to be handled.
I'm afraid that we are possibly using software that, in a certain way, doesn't support huge amounts of data.
Arbitrary precision libraries are available (i.e. GMP or MPFR) but, AFAIK, they are not used in bioinformatics tools...
What do you think about this?

d

[westerman]

This makes me want to read the samtools list. It is hard to imagine needing more arithmetic precision in bioinformatics than 'double' (aka, 64-bit). Yes there a need for 64-bit aware programs for data storage and manipulation. Perhaps you could give us an example of the need for 'long double'. (BTW: 'long double' itself is an imprecise definition/implementation, AFIK.)

[dawe]

Perhaps you could give us an example of the need for 'long double'. (BTW: 'long double' itself is an imprecise definition/implementation, AFIK.)

I'm not a fan of long double (which is implemented in a architecture dependent way and also avoids SSE optimizations), that's why I wonder why MPFR is not included in most of the bioinformatics software...
BTW, do you think today's results are affected by an incorrect rounding issue?

[nilshomer]

Hi all,
I've been thinking recently about softwares for NGS... On the samtools mailing list I've discussed about some variables that were implemented as "long double" as the developer needed more precision than "double" numbers.
NGS systems output gigabytes of data, millions of sequences to be analyzed. Before this, software using "float" or "double" variables might be enough precise when hundreds or thousands of sequences had to be handled.
I'm afraid that we are possibly using software that, in a certain way, doesn't support huge amounts of data.
Arbitrary precision libraries are available (i.e. GMP or MPFR) but, AFAIK, they are not used in bioinformatics tools...
What do you think about this?

d

Consider that a lot of the time probabilities are computed in log space, adding further precision.

Nils