Do we need FORCE_FLOAT? about gemma HOT 7 CLOSED

The float version was mainly introduced for the bslmm model. I tried the float version a while ago on our previous cluster (pps cluster), and found that it only uses half the memory required by the double version (if I remember correctly, for 5k idv and 0.5million snps, the float version uses 8gb while the double version uses 16gb). But that was about 5 years ago. I am not sure either whether there is any benefit nowadays. It would be great if you could do a quick test on this, and we can certainly remove this option if there is no longer benefits.

from gemma.

@xiangzhou I will test to see if FORCE_FLOAT reduces memory usage in 64-bit systems.

from gemma.

Don't use float if you have double. You have to really know what you are doing.

https://wiki.haskell.org/Performance/Floating_point

from gemma.

@xiangzhou An update on this:

I wasn't able to build the binary executable using FORCE_FLOAT = 1; it generated several errors which did not seem straightforward to resolve.

Upon further experimentation in MATLAB, I think it is possible to reduce memory requirements with single-precision. I wasn't able to find any documentation about this in MATLAB, but this webpage briefly discusses storage of <64-bit numbers in Julia.

In short, there is a potential benefit to keeping the FORCE_FLOAT option.

However, perhaps a larger issue is that debugging & maintaining single-precision and double-precision versions of GEMMA will involve substantially more effort, and I'm not sure if the effort is justified by the benefit (which, in my opinion, is relatively small---there are other more effective ways to reduce memory, e.g., by LD pruning).

I've created a new branch force-float that preserves the state of the current repository with the FORCE_FORCE option, and my plan is to remove this option in the master branch unless there are any objections.

from gemma.

This is one of the best talks I ever heard. If you want to know about float/double, spend half an hour here. It is worth it. Entertaining too.

http://dconf.org/2016/talks/clugston

from gemma.

I also had issues with FORCE_FLOAT in Linux.

from gemma.

@tseemann Thanks for sharing.

from gemma.