Originally Posted by
nkedel
In the general case, you're right.
In the specific case I mentioned (the zram driver on Linux), exchanging some main memory for a compressible memory area makes a lot of sense for some workloads. The exchange, due to compression, is nowhere near 1:1.
There are loadfull of memory techniques. For example, ESXi uses something called Transient Page Sharing, which in other words is memory dedup. It also uses memory compression (caching of pages to be swapped), which outperforms swapping since decompressing is faster than loading from storage. However, the swap file is still, and no matter what memory management techniques you use, the swap is the last resort and is meant to be stored AWAY from the memory.