In my relentless pursuit of trying to coax more performance out of my Lemmy instance I read that PostgreSQL heavily relies on the OSs disk cache for read performance. I’ve got 16 GB of RAM and two hdds in RAID 1. I’ve PostgreSQL configured to use 12 GB of RAM and I’ve zram swap set up with 8 GB.
But according to htop PostgreSQL is using only about 4 GB. My swap gets hardly touched. And read performance is awful. Opening my profile regularly times out. Only when it’s worked once does it load quickly until I don’t touch it again for half an hour or so.
Now, my theory is that the zram actually takes available RAM away from the disk cache, thus slowing the whole system down. My googling couldn’t bring me the answer because it only showed me how to set up zram in the first place.
Does anyone know if my theory is correct?
Linux has kind of two forms of memory pages (entries in RAM), one is a file cache (page cache) and the other is “memory allocated by programs for work” (anonymous pages).
When you look at memory consumed by a process you are looking at RSS, page/file cache is part of kernel and for example in btop corresponds to
Cached.Page cache can never be moved into swap - that would be the same as duplicating the file from one place on a disk to another place on a (possibly different) disk.
If more memory is needed, page cache is evicted (written back into the respective file, if changed). Only anonymous pages (not backed by anything permanent) can be moved into swap.
So what does “PostgreSQL heavily relies on the OSs disk cache” mean? The more free memory there is, the more files can be kept cached in RAM and the faster postgres can then retrieve these files.
When you add zram, you dedicate part of actual RAM to a compressed swap device which, as I said above, will never contain page cache.
In theory this still increases the total available memory but in reality that is only true if you configure the kernel to aggressively “swap” anonymous pages into the zram backed swap.
Notes: I tried to simplify this a bit so it might not be exact, also if you look at a process, the memory consumed by it is called RSS and it contains multiple different things not just memory directly allocated by the code of the program.