A more interesting example might be if each slot in the target array has the next index to go to in addition to the value, then you will introduce a dependency chain preventing this from happening.
Another thing I noticed is that the spike on the left hand side of his graphs is the overhead of file access.
Without this overhead, small array random access should have a lot better per-element cost.
However, on some processors there's a data-dependent prefetcher that will notice the pointer-like value and start prefetching that address before the CPU requests it.
At this point I'm kinda expecting CPU vendors to stop putting as many Spectre mitigations in the main core, and just have a small crypto core with full-fat arithmetic, less hardware for memory access, less speculation, and careful side-channel hardening. You still have to block Meltdown and other large vulnerabilities on the main cores, but if someone wants to protect elliptic curves from weird attacks? Try to set the DIT bit, trap into the OS, and get sent to the hardened core.
Well in a language that allows you to generate compile time specialized serde code like Nim or Zig. Maybe C++ has enough compile time reflection to do it now as well?
I don't know enough about Rust's serde, but it seems like there'd be a lot of performance overhead with it's design and the limits of Rust's macro and compiler system.
*maddeningly, with mutually incompatible sorting rules.
Does x86 64 actually do this data dependent single deref prefetech? Because in that case I have a some design assumptions I have to reevaluate.
CPU implementation has become too complex to grasp. The only sure way to know how a CPU will behave for a given workload is to run the workload. It's good to have some basic expectations of performance, instructions/cycle, memory bandwidth, to detect if something is off. I guess I'm trying to say it's hard to keep in your head all the details of what ~1B transistors are doing together to run your code. It's just too big.
You can avoid this with two arrays. One contains random query positions, and the target array is also filled with random values. The next index is then a function of the next query position and the previous value read from the target array.
Eg start with every element referencing the next element (i.e. i+1 with wrap-around), and then use a random shuffle. That way, you preserve the full cycle.
Basically, pick a large prime number p as the size of your array and a number 0 < x < p. Then visit your array in the order of (i*x) modulo p.
You can also do something with (x^i) modulo p, if your processor is smart enough to figure out your additive pattern.
Basically, the idea is to look into the same theory they use to produce PRNG with long cycles.
This is a very small Nim program to demonstrate for "show me the code" and "it must just not be 'random enough'!" skeptics: https://github.com/c-blake/bu/blob/main/memlat.nim It uses the exact dependency idea @andersa mentions of a random cycle of `x[i] = i` that others else-sub-thread say some CPUs these days are smart enough to "see through". On Intel CPUs I have, the dependency makes things 12x slower at the gigabyte scale (DIMMs).
EDIT: This same effect makes many a naive hash table microbenchmark { e.g., `for key in keys: lookup(key)` } unrepresentative of performance in real programs where each `key` is often not speculatively pre-computable.
Traversing a contiguous list of pointers in L1 is also slower than accessing those pointers by generating their address sequentially, so adding a load-load dependency is not a good way to benchmarking random access vs sequential access (it is a good way to benchmark vector traversal vs list traversal of course).
At the end of the day you have to accept that like caching and prefetching speedup sequential access, OoO execution[1] will speedup (to a lesser extent) random access. Instead of memory latency, in this case the bottleneck would be the OoO queue depth, or more likely the maximum number of outstanding L1/L2/L3 (and potentially TLB) misses. As long as the maximum number of outstanding misses is lower than the memory latency for that cache level, then, in first approximation, the cpu can effectively hide the sequential vs random access cost for independent accesses.
Benchmarking is hard. Making sure that that a microbenchmark represents your load effectively, doubly so.
[1] Even many in-order CPUs have some run-ahead capabilities for memory.
I still suspect @kortilla is one of today's lucky 10,000 (https://xkcd.com/1053/) or just read/replied too quickly. :-)
There is a lot written that indicates that the complexity of modern CPUs is ill-disseminated. But there is also wonderful stuff like https://gamozolabs.github.io/metrology/2019/08/19/sushi_roll... { To add a couple links to underwrite my reply in full agreeance. :-) }
The only explanation I see it the OS implementation then, i.e. cpu does perfect job of prefetching ram into cache, but OS does not perfect job of prefetching ssd to ram?
https://www.forrestthewoods.com/blog/memory-bandwidth-napkin...
I’m not sure I agree with the data presentation format. “time per element” doesn’t seem like the right metric.
Using something like the overall run length would have such large variations making only the shape of the graph particularly useful (to me) less so much the values themselves.
If I was showing a chart like this to "leadership" I'd show with the overall run length. As I'd care more about them realizing the "real world" impact rather than the per unit impact. But this is written for engineers, so I'd expect it to also be focused on per unit impacts for a blog like this.
However, having said all that, I'd love to hear what your reservations are using it as a metric.
Perhaps it’s a long time inspiration from this post: https://randomascii.wordpress.com/2018/02/04/what-we-talk-ab...
I also just don’t know what to do with “1 ns per element”. The scale of 1 to 4 ns per element is remarkably imprecise. Discussing 1 to 250 million to 1 billion elements per second feels like a much wider range. Even if it’s mathematically identical.
Your graphs have a few odd spikes that weren’t deeply discussed. If it’s under 2ns per element who cares!
The logarithmic scale also made it really hard to interpret. Should have drawn clearer lines at L1/L2/L3/ram limits.
On skim I don’t think there’s anything wrong. But as presented it’s a little hard for me as an engineer to extract lessons or use this information for good (or evil).
There shouldn’t be a Linux vs Mac issue. Ignoring mmap this should be HW.
I dunno. Those are all just surface level reactions.
Agreed that the odd spikes don't matter, that's why I didn't bother discussing them; I was more interested in the data after the array got large enough that random access was actually slower. It looked like all those weird spikes were for arrays small enough to fit in cache anyways.
I agree that it could have been helpful if I'd drawn lines at L1/L2/L3/RAM limits, but I didn't do that because I don't think it's entirely clear where those lines should have been drawn. Specifically because there are two arrays. Should the line show just where the floating-point array is small enough to fit in cache, or where both arrays together are?
Not sure I quite follow what you're saying about mmap on Linux vs Mac; only one of the three sets of experiments used mmap, and the third was explicitly to try to tease out that effect. Especially for the first experiment, I agree that there should be no difference for arrays small enough to fit in RAM, since the whole file gets read into memory first.
That was sarcasm =P Those spikes are very curious and the choice of presentation makes them seem like noise but there is something there that should be investigated further imho. In the graph it looks like noise. I mean it’s just 1ns. But a 2x throughput difference isn’t noise! Thats huge! Very curious.
> Not sure I quite follow what you're saying about mmap on Linux vs Mac
Your 4th conclusion is “On Linux, random order starts getting even slower for arrays over a gigabyte, becoming more than 50x slower than first-to-last order; in contrast, random order on the MacBook seems to just level out as long as everything fits in RAM.”. That doesn’t make sense. There shouldn’t be any OS difference here.
Could you clarify why there shouldn't be an OS difference? I was under the impression that it's the OS that handles how swap space is implemented (which was used by the first set of experiments), as well as how memory-mapped files are implemented (which was used by the second set of experiments). Am I mistaken about that?
Ignoring mmap. But why on a 16gb system why would performance degrade at 1Gb? You shouldn’t be hitting the swap. So any differences should be hardware architecture related. M1 unified vs Ryzen. And I wouldn’t expect 1Gb to be a magic threshold.
I would definitely expect a threshold beyond 16gb. And I’d expect the swap to come into play at maybe 12gb. I wouldn’t expect a huge difference between 500mb and 8gb. Ok there’s probability difference of hitting the L3 cache. But most of those random accesses will be hitting system RAM so it should be the same.
Could be wrong! But that’s what I’d expect.
Yes, I agree that it doesn't make sense for that to be due to the OS. That's just poor writing on my part: in that case I wasn't actually trying to imply that it was due to the operating system specifically, I was just using "Linux" and "the MacBook" as shorthand to refer to my two different computers, which differ in more ways than just the OS. In this case, another commenter suggested it might be due to my physical setup of having three RAM sticks: https://news.ycombinator.com/item?id=44397214
So yes, in that sentence I should have written "on my desktop" instead of just "on Linux".
> Random access from the cache is remarkably quick. It's comparable to sequential RAM performance
That's actually expected once you think about it, it's a natural consequence of prefetching.
Lots of things are expected when you deeply understand a complex system and think about it. But, like, not everyone knows the system that deeply nor have they thought about it!
The reason I chose "time per element" then follows from that different goal, because I was comparing across vastly different array sizes, so no other metric I could think of would have really worked for the charts I was drawing.
There were proofs of concept by 2010 that the latency-hiding mechanics could be implemented on CPUs in software, which while not as efficient had the advantage of cost and performance, which was a death knell for the MTA. A few attempts to revive that style of architecture have come and gone. It is very difficult to compete with the economics of mass-scale commodity silicon.
I hold out hope that a modern barrel processor will become available at some point but I’m not sanguine about it.
Presumably if you'd split the elements into 16 shares (one for each CPU), summed with 16 threads, and then summed the lot at the end, then random would be faster than sorted?
Although, it looks like that chip has a 1MB L2 cache for each core. If these are 4 Bytes ints, then I guess they won’t all fit in one core’s L2, but maybe they can all start out in their respective cores’ L2 if it is parallelized (well, depends on how you set it up).
Maybe it will be closer. Contiguous should still win.
And of course the headline of getting a cpu you can’t fully use.
One hint in the same article that random access is not cheap, in contrast with the conclusion, was noticing that the shuffle was unacceptably slow on large data sets.
Still, good to see peformance measurements, especially where the curves look roughly like you'd hope them to.
This means: The system likely uses 3x 8 GB modules. As a result, one channel has two modules with 16 GB total, while the other channel has only a single 8 GB module.
Not sure how big this impact is with the given memory access patterns and assuming [mostly] exclusive single-threaded access. It's just something I noted, and could be a source of unexpected artifacts.
At least once the tests become big enough to have some data in both partitions, the bandwidth will start to matter.
Out of curiosity, what do you see when running the same code on your machine?
If I get around to fixing my dual boot machine, I'll try to remember running the benchmark and dropping you some results by mail.
Adhyyan1252•7mo ago
Nevermark•7mo ago