Comment by wrs
9 hours ago
Indeed, as I used to tell my ops colleagues when they pointed to RAM utilization graphs, "we paid for all of that RAM, why aren't we using it?"
9 hours ago
Indeed, as I used to tell my ops colleagues when they pointed to RAM utilization graphs, "we paid for all of that RAM, why aren't we using it?"
Because OoM errors are oh so fun.
I write algorithms that operate on predictable amounts of data. It's very easy to work out the maximum amount of things we need to have and then allocate it all in fixed size arrays. If you allocate all your memory at startup you can never OOM at runtime. Some containers need over 100GB but like the parent comment said we've already bought the RAM.
Caches are automatically released by the OS when demand for memory increases.
You eventually run out of caches to evict.
4 replies →
do you also say that about hdd space? about money in the bank?
It’s counterintuitive but I learned this best by playing RTS games. If you don’t spend money your opponent can outdo you on the map by simply spending their money. But the principle extends, everything you have doing nothing (buildings units etc) is losing. The most efficient process is to have all your resources working for you at all times.
If you don't have savings to spend for a potential change of tactics, larger players, groups or players with different strategies can easily overtake you as your perfectly efficient economy collapses.
Going to also echo the comment that this isn't an RTS
It's why I wake up at 3am to make sure my agents aren't waiting on me :D
> It’s counterintuitive but I learned this best by playing RTS games. If you don’t spend money your opponent can outdo you on the map by simply spending their money.
OK, hear me out over here:
We are not in an RTS.
Edit: in real-world settings lacking redundancy tends to make systems incredibly fragile, in a way that just rarely matters in an RTS. Which we are _not in_.
Why he wouldn't say it about HDD space? You buy HDD to keep them empty?
And as for the money analogy, what's the idea there, that memory grows interest? Or that it's better to put your money in the bank and leave it there, as opposed to buy assets or stocks, and of course, pay for food, rent, and stuff you enjoy?
Money analogy could better be put as one of:
1. Store your money in a 0% interest account—leave RAM totally unused—or put it in an account that actually generates some interest—fill the RAM with something, anything that might be useful.
2. Store your money buried in your backyard or put it in a bank account? If you want to actually use your money, it's already loaded into the bank.
Imperfect analogies because money is fungible. In either case though, money getting spent day-to-day (e.g. the memory being used by running programs) is separate.
> about money in the bank?
Yes, generally. That's the entire idea behind the stock market.
because memory access performance is not O(1) but depends on the size of what's in memory (https://www.ilikebigbits.com/2014_04_21_myth_of_ram_1.html). Every byte used makes the whole thing slower.
I am not following, isn't this just a graph that shows that how fast operations happen is largely dependent on the odds that it is in cache at various levels (CPU/Ram/Disk)?
The memory operation itself is O(1), around 100 ns, where at a certain point we are doing full ram fetches each time because the odds of it being in CPU cache are low?
Typically O notation is an upper bound, and it holds well there.
That said, due to cache hits, the lower bound is much lower than that.
You see similar performance degradation if you iterate in a double sided array the in the wrong index first.
O notation is technically meaningless for systems with bounded resources. That said, yes the performance is depending on the probability of cache hits, notably also the TLB. For large amounts of memory used and random access patterns, assuming logarithmic costs for memory access tends to model reality better.
Memory access performance depends on the _maximum size of memory you need to address_. You can clearly see it in the graph of that article where L1, L2, L3 and RAM are no longer enough to fit the linked list. However while the working set fits in them the performance scales much better. So as long as you give priority to the working set, you can fill the rest of the biggest memory with whatever you want without affecting performance.
> Every byte used makes the whole thing slower.
This is an incorrect conclusion to make from the link you posted in the context of this discussion. That post is a very long-winded way of saying that the average speed of addressing N elements depends on N and the size of the caches, which isn't news to anyone. Key word: addressing.
RAM is always storing something, it’s just sometimes zeros or garbage. Nothing in how DRAM timings work is sensitive to what bits are encoded in each cell.
why is it not O(1)? It has to service within a deadline time, so it is still constant.
Huh? There is nothing called "empty memory". There is always something being stored in the memory, the important thing is whether you care about that specific bits or not.
And no, the articles you linked is about caching, not RAM access. Hardware-wise, it doesn't matter what you have in the cells, access latency is the same. There is gonna be some degradation with #read/write cycles, but that is besides the point.
The author of that post effectively re-defines "memory"/"RAM" as "data", and uses that to say "accessing data in the limit scales to N x sqrt(N) as N increases". Which, like, yeah? Duh, I can't fit 200PB of data into the physical RAM of my computer and the more data I have to access the slower it'll be to access any part of it without working harder at other abstraction layers to bring the time taken down. That's true. It's also unrelated to what people are talking about when they say "memory access is O(1)". When people say "memory access is O(1)" they are talking about cases where their data fits in memory (RAM).
Their experimental results would in fact be a flat line IF they could disable all the CPU caches, even though performance would be slow.