2 ms·
Feels like there might have been a middle ground short of a scratch rewrite. LRU caches are always suspect. LRU is pretty much the worst replacement policy you
by jeffbee 6y ago
Feels like there might have been a middle ground short of a scratch rewrite. LRU caches are always suspect. LRU is pretty much the worst replacement policy you can choose. I wonder if a smaller cache with a better replacement policy (2Q, ARC, ...) would have solved this specific problem.
Another approach I've used with caches in Go is to just use a huge byte slice as the cache without any pointers, so the GC can ignore it.
- Jtsummers 6y ago> Another approach I've used with caches in Go is to just use a huge byte slice as the cache without any pointers, so the GC can ignore it. Just to check that I understand what you're suggesting here, this is comparable to how we embedded systems folks might pre-allocate (either statically or at the very start of a program execution) a large chunk of memory and use it instead of using traditional malloc/free (or language equivalent), correct? So you bypass the GC and avoid any potential performance penalties introduced by having a smaller number of massive allocations that never get moved, freed, or reallocated.
- jeffbee 6y agoYes, the idea is similar. If you make a huge map in Go with millions of items, it will stress the GC. If you take the same information and just put it into a big byte slice, and index into it in a way that the GC doesn't think you're pointing to addresses, then it will be much faster. Of course now you are writing a little more code.