While playing with CUDA, I discover something that may help all linux users.
Tying to run a program with too large a matrix for the available memory, the swap space started to be used. It did slow down the program quite a lot and then the program stopped itself due to lack of memory. Because the program was running for a while using the swap space, some of the not too frequenly used processes were move to the swap space and stayed there even after the program had stop running. Therefore, the main free memory available was larger than before the program started running. So more memory space is made available for other tasks to run in main memory. The swap space memory is then slowly freed at a rate of about 4k per minute and then stop decreasing. So if you have a large swap space, it would stay this way for quite a while.
For example:
When I boot my machine, 260kB is used before loading X (4GB available)
After loading X, around 600kB is used.
After doing this trick, It goes down to 300kB again, with X running.
It is possibly better than that, but I did not try too much to play with this.
I don’t know what would happen with the file system if there is a power failure, if the journalling system would still work etc…