What Every Programmer Should Know About Memory (2007) [pdf]
akkadia.org
akkadia.org
Also, if you know Ulrich Drepper at all, either from some of his talks or his mailing list presence, this is just a very fitting title from him. Just pure deadpan, you think its funny, he probably does not, the fact that you think its amusing is just disappointing him like a professor looking out at freshman undergrads wondering how he got stuck teaching this class.
> The title of this paper is an homage to David Goldberg’s classic paper “What Every Computer Scientist Should Know About Floating-Point Arithmetic”
But then I also think that What Every Computer Scientist Should Know About Floating-Point Arithmetic should be updated to include UNUMs. I don't think that will happen either. Also, thank you Mr. Goldberg.
https://yarchive.net/comp/linux/software_prefetching.html
Something I don't think either Drepper or Hennessy + Patterson's books get across is memory banks from a programmer's perspective. How cache organization affects a program is explained well but how banks affect said same isn't. Construction yes. Visibility, no.
Further, they are completely irrelevant for many developers. The next Mars rover for example is not going to be running untrusted code.
However, and I can't remember if it is alreaddy in the book, a section on compensated summation / dot product would give strong time tested tools to attack the problem from within the current framework.
I quickly perused the article and I think this link should be renamed "What 99.9% of programmers don't need to know about memory."
I've managed to go from Associate to Principal without knowing 99% of what's covered in this document, and I'm struggling to understand why the average Java, C#, Python, Rust, <insert language here> programmer would need to know about transistor configurations or voltages, pin configurations, etc. Let alone 114 pages of low level hardware diagrams and jargon!
This document is for someone working on low level drivers for memory, or working on the hardware side. For any normal software engineer, this information is not helpful for doing your job.
But again this is speaking from my experience as an SRE/SDE (or a perf/reliability engineer).
But any high-performance programmer needs to understand the RAS / CAS / PRE cycle, if only to understand WHY the "streaming" of data is efficient, while random-access is very inefficient.
If you are accessing RAM randomly, you better be sure its within L3 cache (or nearer). I've done some experiments, and "streaming" data from beginning to end can be 2x to 3x faster than random access on modern DDR4 RAM.
Understanding the RAS / CAS / PRE cycle helps me understand why streaming data to RAM is faster. And understanding that cells are simply capacitors helps me understand why the RAS / CAS / PRE cycle is necessary in DRAM.
I think what they're saying is that 99.9% of programmers don't need to know why, they just need to know that streaming access is 2-3x faster than random even without any memory stalls.
Now, I'd personally explain things in a far more simple manner than what was described in the PDF. Here are the facts that programmers need to know:
1. DRAM stores data in very tiny capacitors. These tiny capacitors have two properties: they run out of electricity in just 64ms. And second, they run out of electricity after a SINGLE read operation.
2. DRAM has a temporary location called "sense amplifiers" where data is stored during a refresh or a read. These sense amplifiers can hold data permanently.
3. This "temporary read" is called Row-open (or RAS). Reading from an already open row is called a Column-read (CAS). Sending the data back to DRAM proper is called Precharge (PRE). Remember, the sense amplifiers must be clear before they can read from a new row. (The old data in DRAM was destroyed when you read it with the RAS operation)
4. I guess there's a periodic refresh you should know about: instead of trying to fix all RAM every 64ms, you're supposed to do it in small chunks at a time. Every dozen microseconds, RAM will self-read / self-write to refresh another row. Don't be surprised if your memory-reads randomly stall out an extra few hundred nanoseconds because of this refresh.
The end. Not so hard, now is it?
--------
DRAM is faster when you stream, because you open a row once, fill out all the data in a row, and then send the row back to DRAM. In effect, you only have to do a bunch of "column" writes to sense amplifiers, as opposed to opening-and-closing a bunch of different rows.
----------
So yeah, programmers should know it because its really not that hard to learn :-) And if you start measuring your program at the nanosecond level, you'll actually see these effects and start to demand explanations.
-------------
EDIT: Hmmm... the more I think of it, the less its something "programmers" need to know and something "SysAdmins / DevOps need to know". An advanced Sys Admin can use these profiler tools to figure out whether they need that 6x Memory Channel computer or the 8x Memory channel computer on the next purchase.
Is your code memory-bound? Or is it CPU bound? Should you buy more cores? Should you buy more LRDIMMs for higher amounts of RAM? Or is your program latency-bound and actually benefits from the lower latency of RDIMMs or even UDIMMs ?
The programmers kinda don't make those decisions.
I suppose that's a credit to all the engineers who build all the middle layers that allow software engineers to float along at an abstract and more productive level.
* Rough latency timings
* Caching
* Prefetching
* Sequential vs. random access
* N-dimensional layouts (row/column major and arbitrarily strided)
* Design of cache-oblivious algorithms
* SIMD-able access patterns
* False sharing
* Instruction cache & code size
* Branch prediction and speculative execution
I'd be curious to hear what else folks would put on (or remove from) this list.
I'd add "Virtual Memory" to that list. In particular, the TLB cache, memory pages (4kB, 2MB "Large Pages", 2GB "Huge Pages).
Although x86 specific, I'd also add x86-64 has 48-bit physical pointers: the top 16-bits are basically ignored by the current virtual memory system. I dunno if the whole Page Directory / Directory Tables / etc. etc. needs to be fully explained, but programmers should have an overall good idea what they are.
There's lots of things to do with Virtual Memory. And anyone who actually reads profiler data needs to understand what the heck that TLB Cache Hits performance counter means.
> * False sharing
More specifically the MESI Protocol (although that's an abstraction), and cache lines should be taught. False Sharing comes as an understanding after you understand those other two concepts.
A CPU Core holds a cache line in Exclusive state so that it can write to it. A 2nd CPU Core attempts to gain access, but it cannot until the 1st core releases control (by writing data back to memory and setting the line to the Invalid state).
The knowledge of the 64-byte cache line is more general than just false sharing: it helps understand why alignment can be an issue (a load/store across a 64-byte cache line would require 2-reads by the memory controller), etc. etc.
and "the top 16-bits are basically ignored" is a funny way to spell "general protection exception on linear memory reference in non-canonical space" but sure, guess we're just handwaving here
Because no CPU actually implements MESI. All CPUs implement more complicated stuff, like MESIF / MOESI. Instead of going into MESIF (which only Intel CPUs implement) or MOESI (only older AMD CPUs implement), lets just stick with the textbook MESI.
Which is "wrong", but its "correct enough" to explain the concept. That's what I mean by an abstraction, no CPU today actually does MESI, its simply a concept to introduce to solidify the student's understanding of cache coherency. Its close enough to reality without getting into the tricky CPU-specific details of the real world.
> guess we're just handwaving
I mean, you have to set those bits back to 0 before using them as a pointer.
But the system will literally never use those top 16-bits for anything. So some highly optimized code stores data in those top 16-bits and then zeros them out before using. IIRC, Lisp machines and various interpreters.
I think its easier to explain cache-coherence through MESI, rather than to abstractly just say "Caches are coherent". At least personally, I didn't understand cache coherence until I sat down and really tried to understand MESI.
I guess other people learn differently than I do, but I always view cache-coherence through the MESI lens.
Could you elaborate on this? Which pointers exactly?
There's an extension to use 55-bits or 56-bits... I forget exactly. But I don't think its actually been implemented yet on any CPU yet.
EDIT: It was Intel's 57-bit memory: https://en.wikipedia.org/wiki/Intel_5-level_paging . Yeah, I knew it was a weird number. But I guessed wrong earlier.
With 48 bits you can still address 256TB of memory. I guessing that from a practical and financial point of view it probably made little sense for vendors to build a CPU that enabled addressing the full 64 bits. At least for now.
Let's say someone is building a micro-service in C#.Net. Why would any of this stuff matter to them? The company cares about features and moving forward quickly.
It's rather like forcing all your developers to be able to code every type of different sorting algorithm when they are only ever going to call "mylist.Sort()" in reality.
This seems like another of those occasions when HN misjudges what the vast reality of practical day to day software creation looks like. Most engineers are not writing low-level code, they are not working on hardware directly, and they don't need to know exactly how RAM works at the transistor level.
I'm not saying it's not interesting, but it's not going to change how I write my for-loops and if statements or make calls to BigTable, etc.
You might be surprised at how much changing your for loops can be when accessing matrix data.
for(int i=0; i<10000; i++){
for(int j=0; j<10000; j++){
blah[j][i] = whatever(); // Column oriented is very slow.
}
}
for(int j=0; j<10000; j++){
for(int i=0; i<10000; i++){ // Very fast: We move row-wise now. With luck, this is SIMD-vectorized by your compiler.
blah[j][i] = whatever();
}
}
-----------
If someone is writing high performance code, the #1 goal is to be able to read your profiler's output. If you can't read what the profiler says (cache hits, TLB hits, memory stalls, etc. etc.) then you can't make sense of the data.
The PDF goes deeper than I'd personally go, but there's some concepts here that are absolutely necessary if you actually want to understand what a decent profiler gives you these days.
Also, you have to understand that saving 1 or 2 milliseconds isn't important when compared to getting software built faster. For the typical use case, optimizing a for-loop like this for in-memory data isn't going to change the total performance of your program by even 0.5%. This is what so many people just don't understand. For modern web based applications, 99.9% of the time spent is doing IO, either hard-drive, data storage, or across the network. Optimizing those calls is FAR more important, than wasting time trying to save 1 millisecond in a for-loop which is iterating over 100 in-memory objects.
This is where the understanding that the practical reality of building software for a real business is very different to theoretical examples drummed up by computer scientists on discussion boards.
I see so many engineers trying to make their in-memory code faster by parallizing it, saving (if they're lucky) 1 or 2 milliseconds, but making their code far less maintainable and potentially introducing bugs, while the users wouldn't notice any practical difference in the app. Meanwhile, the app really does feel slow, because some numb-nuts made a call to some api and blocked the ui thread.
Our hadoop cluster is relatively small at 5PB right now, but I would be impressed indeed if you were processing PB of data on a smallish AWS instance.
If you were considering using Hadoop for work that can be processed on a "smallish instance" then I'd suggest you maybe don't really understand big-data or the normal workloads for which Hadoop is intended. For example, we run ML models on our Hadoop cluster which take hours to run distributed over around 100 nodes. Let me know when this can be done on a "smallish instance."
In general, you optimize software by starting from the slowest things first, and then working your way to the faster things when you run out of things to optimize. It seems like your problems are I/O bound, and therefore the bulk of optimizations can be made by simply optimizing your I/O (either through async calls, or better understanding of what the frameworks are actually doing, etc. etc.).
And that makes sense for sure.
The thing is: the next level of optimization is not CPU-optimization, but instead memory optimization. In fact, all CPU-based optimization starts at the Main-memory level. Why?
Because memory is slower than virtually everything inside of the main CPU Core.
Which means, memory-level optimization is a far more important skill than any other CPU-based optimization technique. Modern CPU Cores operate at 4GHz easily, but RAM only responds every 100ns on server systems (that's 400 CPU-cycles of RAM Latency!)
-----------
In effect, if there's one low-level optimization any higher-level programmer should know about, it is RAM optimization. Sure, there are CPU-optimizations (SIMD registers, L1 vs L2 cache, and more), but RAM access itself is hundreds of times slower than CPU-speeds and needs to be optimized before lower-level optimizations are sought.
Once RAM access are fully optimized, then you can finally reach for the CPU-core optimizations. Instruction-level parallelism, or SIMD Registers, or what-not.
RAM is the next slowest part of your system after I/O. As such, optimizing RAM access is the most logical next step forward when your programs are running poorly.
Why not? I work with the scripting languages only and regularly hit this scenario so I won't say this is completely useless.
When you analyze a few hundred 50GB files for specific patterns, you have to go line by line in those cases shaving off milliseconds, optimizing how data should be accessed and optimized becomes valuable.
There are tens of thousands teams and thousands of tournaments and his code was taking two days to precalculate the results for every team.
Written with minimal understanding of how slow the memory is, the new code takes less than a minute. There was no point of optimizing it further, the algorithmic changes were enough.
1. L1 cache lines are 64-bytes long. By fetching column-wise, you are wasting the bandwidth between L1 and main-memory. L1 cache will always fetch 64-bytes. By moving "with" the cache, you allow the L1 --> Main Memory data-transfers to be far more efficient.
2. Virtual Memory is translated by the TLB before it returns the actual value. Moving within a 4kB page is more efficient than moving across pages.
The non-reason:
* Hardware prefetcher probably works, even on column-oriented data.
All of these reasons hold even if the SIMD-optimizer fails. If the SIMD-optimizer is actually working, you'll more efficiently load/store to L1 cache. But this is likely a memory-bound problem and optimizing the core isn't as important.
% cat t.c
#include <sys/time.h>
#include <sys/resource.h>
#include <stdio.h>
double get_time()
{
struct timeval t;
struct timezone tzp;
gettimeofday(&t, &tzp);
return t.tv_sec + t.tv_usec*1e-6;
}
int blah[SIZE][SIZE];
int whatever() { static int i=0; return ++i; }
int main()
{
double t0 = get_time();
for(int i=0; i<SIZE; i++){
for(int j=0; j<SIZE; j++){
blah[j][i] = whatever(); // Column oriented is very slow.
}
}
double t1 = get_time();
for(int j=0; j<SIZE; j++){
for(int i=0; i<SIZE; i++){ // Very fast: We move row-wise now. With luck, this is SIMD-vectorized by your compiler.
blah[j][i] = whatever();
}
}
double t2 = get_time();
printf("SIZE=%5d dt1=%4.3e dt2=%4.3e\n",SIZE,t1-t0,t2-t1);
return 0;
}
% for (( i=1; i<100000 ; i*=10 )) do gcc -DSIZE=$i -O t.c; ./a.out; done
SIZE= 1 dt1=4.179e-04 dt2=0.000e+00
SIZE= 10 dt1=5.190e-04 dt2=0.000e+00
SIZE= 100 dt1=5.062e-04 dt2=9.537e-07
SIZE= 1000 dt1=4.014e-03 dt2=1.490e-04
SIZE=10000 dt1=1.347e+00 dt2=4.349e-02
SIMD only really makes a difference at the L1 or L2 cache levels. Its a fancy micro-optimization that compilers do and can improve code speed in those cases...
But the "big" change, going from column-wise traversal into row-wise traversal, is the huge memory optimization that programmers should know about. It just so happens that SIMD-optimizations are also easier for compilers to figure out on row-wise traversal, so you get SIMD-optimization "for free" in many cases.
My list presupposes you've hit a wall and need the best performance you can get. That's not always the case, but I certainly wouldn't say that these optimizations don't make a difference in the "real world of softare."
False sharing alone can be the difference between a 16x parallelization speedup and a 1000x+ slowdown over the naive serial algorithm. 16x can be the difference between actionable results tomorrow or three weeks from today... and if you didn't know about it, the 1000x+ slowdown would be otherwise inscrutable.
I don't know exactly what that number is going to be, but I do know it is going to be a lot less than 100%.
For example, an O(n^2) algorithm can often beat an O(n) one when n is "small". How do you get that "experienced" feel for which to choose when you're designing your program?
This is even important for people doing REST APIs, Web Services and Microservices. Milliseconds matter when the load gets beyond a certain point and your system ends up full of stragglers. How do you get that "experienced" feel for where that point might be and how far you can push a particular architecture without investing in a major rework?
The effect of tail latency on most web servers might surprise you:
http://latencytipoftheday.blogspot.com/
http://latencytipoftheday.blogspot.com/2014/06/latencytipoft...
TL;DR: most of your visitors will experience the bottom 1% of your performance curve on every page load.
Back when I started, both the BASIC and Common Lisp camps had heuristics on helping the compiler generate more efficient code. The reasons are similar to today: increase throughput for scaling or customer experience while simultaneously reducing what you spend on hardware and/or electricity. Lean, easy-to-manage setups can also sometimes let you afford extra personnel to build stuff faster.
There's definitely a cut-off where micro-optimizations wouldn't be necessary. A lot of efficiency gains are simple, though. Just gotta consistently use what you learn.
"web applications/web sites/microservices"
The world is bigger than you think it is. There could be more code in making a single AAA videogame than in the entire Amazon infrastructure.
I'd bet you would suggest to rename it too.
I would argue that sure, if you want to be a PhD, academic, researcher etc, in Computer Science, then reading said article might be useful.
For engineers working at companies whose goal is to get product to market faster, coding ruby on rails, javascript, making web apps, websites, it's far more useful knowing your web frameworks well and being able to iterate fast, than knowing the intimate mathematical details of floating point arithmetic.
Computer Science != engineering. It's like saying knowing physics is the same thing as knowing how to build a building.
You just proved their point about how you'd reply, though.
What actual engineers NEED to know about floating point numbers could be put on a single A4 page. It doesn't have to be 50 pages, where many of those pages are full of equations that 99% of engineers don't understand.
My point is, because you seem to have trouble grasping this. Most real-world programmers, working at startups or bigger companies, making their WPF applications or Spring applications, or react/jquery/angular apps, or Swift IOS apps, etc, do not NEED to know 99% of this document. For most real-life cases all they need to know is "Use the decimal type if you're dealing with money" or something along those lines.
Why do I know I'm correct? Because I work in an organization with 800-1000 engineers, and I know a fair few of them myself, and I bet maybe 1 person in the org has read this doc fully (and even this is a stretch). But the company stills makes billions in revenue every year like clockwork and the world keeps on spinning.
Hence, this document is for theorists and academics, not for the average engineer making enterprise business applications. If it is for an engineer, it's for someone making extremely niche mathematical software or something equally arcane.
Please don't insult people like that.
If someone doesn't follow your reasoning, take responsibility yourself and find another way to explain the point so others may understand it more easily.
"My point is, because you seem to have trouble grasping this, ..."
Ha-ha, the irony.
No, but we can make algorithms that will do things in other ways than standard. Knowing memory layout and behaviour enables programmer to invent better ways to do things. Of course those algorithms are very application specific, but at least it increases solution space. Erlangers mantra here is "profile this" so that you know which one is actually better.
And when necessary the C FFI is useful to have to write small bits of code to take advantage of particular layouts and drive them from code. See [0].
So, of course, you had a wonderful career, but it doesn't prove that you write efficient code
[1] http://www.rntz.net/post/against-software-development.html [2] http://tonsky.me/blog/disenchantment/
https://www.joelonsoftware.com/2001/03/23/strategy-letter-iv...
"In 1993, given the cost of hard drives in those days, Microsoft Excel 5.0 took up about $36 worth of hard drive space. In 2000, given the cost of hard drives in 2000, Microsoft Excel 2000 takes up about $1.03 in hard drive space. In real terms, it’s almost like Excel is actually getting smaller!"
Most programmers are working for a business whose goal is to make money. Usually that involves adding more features or producing the product faster (as in development time). As far as optimal or efficient goes, all the company cares about is: "Is it fast enough so that people still buy/use the product?"
Premature optimization is wasteful, both a waste of time and money. Unless you're doing it for fun, trying to eke out 1ms on a website load at the cost of excessive dev time and risk of bugs, is just silly.
Don't forget, trying to make super optimal code often introduces bugs as you now have code that is more complex and/or uses more low level constructs. I've seen many deadlocks and race conditions due to engineers trying to optimize code unnecessarily.
Studies have shown that is not the case.
Amazon found that every 100ms of latency cost them 1% in sales. Google found that 400 milliseconds means a nearly 0.5% decrease in search sessions.
So yes, a 5% speedup in an application could be an enormous win for a company.
More than premature pessimization? Setting your project performance goals is plain good design. It's not premature optimization to suggest that your web server should be able to respond to queries within an average window of time. Working with any kind of service level object and agreement practically requires you to think about these goals at the design phase regardless.
I've come across teams with engineers who couldn't understand why their application was timing out. They had no idea how to work with data structures and memory layout to optimize access. Their response to business stakeholders was to weakly raise their hands to the ceiling and say, "It can't be fixed." Their customers would not have been happy with that answer; they had precious data and reports to deliver.
It's worth knowing, at least on a conceptual level, how memory architectures and data structures affect performances. Otherwise you end up paying people like me a lot of money. And you don't want that.
Reading a paragraph like this:
> "The CAS signal can be sent after tRCD (RAS-to-CAS Delay) clock cycles. The column address is then transmitted by making it available on the address bus and lowering the CAS line."
is certainly interesting but it offers no additional insight in to how to write an efficient program. Regardless of whether the row and column addresses are sent on the same bus or a different bus, the optimization strategies for programmers are exactly the same. The same goes for almost all of the information here.
Unless you're writing a kernel, 99% of this article is overkill. I mean that literally: it's over 100 pages long, and I think the important and relevant points for most programmers could be summarized in a page or two.
The title says "should," after all. Who says every programmer should knows these things about memory? Obviously not you. Ulrich Drepper does - I bet if you asked him, he would say everyone should know these things, but would concede that almost no programmers do and most programmers don't need to.
But it's a fun title, and an excellent resource.
Why did the production guy even need them in the first place? Why were they hired?
> (true story). And that production guy's name was Albert Einstein.
- If you know this much about memory, you'll know more than 99% of programmers.
- What 99% of programmers don't know about memory.
> What 99% of programmers don't know about memory... number 17 will shock you!
> Besides, if you don't find this interesting why become a programmer in the first place
Who is saying it's not interesting? We're arguing that it's not fundamentally vital knowledge to know the difference in RAS & CAS latency for SDRAM for most programmers.
It has all been from more experienced coworkers, and I would much appreciate any resources anybody could point me to (free or paid) so that I could round out my knowledge
The profiler not only tells you how fast your code is, but also why your code is fast or slow... at least to the best ability of the hardware performance counters.
Is it RAM-bottlenecked? Is it Compute bound? Are your Warps highly utilized? Etc. etc. If you don't know what the profiler is saying, then study some more.
https://docs.nvidia.com/nsight-visual-studio-edition/Content...
> It is expected that, from late 2008 on, every SMP machine will use NUMA.
Outside servers, still not happened.
They are not that exotic anymore and are no longer exclusive to very expensive servers, e.g. Threadripper 2920X is a $650 CPU, but market penetration is still low.