Related
I have a piece of code in my full code:
const unsigned int GL=8000000;
const int cuba=8;
const int cubn=cuba+cuba;
const int cub3=cubn*cubn*cubn;
int Length[cub3];
int Begin[cub3];
int Counter[cub3];
int MIndex[GL];
struct Particle{
int ix,jy,kz;
int ip;
};
Particle particles[GL];
int GetIndex(const Particle & p){return (p.ix+cuba+cubn*(p.jy+cuba+cubn*(p.kz+cuba)));}
...
#pragma omp parallel for
for(int i=0; i<cub3; ++i) Length[i]=Counter[i]=0;
#pragma omp parallel for
for(int i=0; i<N; ++i)
{
int ic=GetIndex(particles[i]);
#pragma omp atomic update
Length[ic]++;
}
Begin[0]=0;
#pragma omp single
for(int i=1; i<cub3; ++i) Begin[i]=Begin[i-1]+Length[i-1];
#pragma omp parallel for
for(int i=0; i<N; ++i)
{
if(particles[i].ip==3)
{
int ic=GetIndex(particles[i]);
if(ic>cub3 || ic<0) printf("ic=%d out of range!\n",ic);
int cnt=0;
#pragma omp atomic capture
cnt=Counter[ic]++;
MIndex[Begin[ic]+cnt]=i;
}
}
If to remove
#pragma omp parallel for
the code works properly and the output results are always the same.
But with this pragma there is some undefined behaviour/race condition in the code, because each time it gives different output results.
How to fix this issue?
Update: The task is the following. Have lots of particles with some random coordinates. Need to output to the array MIndex the indices in the array particles of the particles, which are in each cell (cartesian cube, for example, 1×1×1 cm) of the coordinate system. So, in the beginning of MIndex there should be the indices in the array particles of the particles in the 1st cell of the coordinate system, then - in the 2nd, then - in the 3rd and so on. The order of indices within given cell in the area MIndex is not important, may be arbitrary. If it is possible, need to make this in parallel, may be using atomic operations.
There is a straight way: to traverse across all the coordinate cells in parallel and in each cell check the coordinates of all the particles. But for large number of cells and particles this seems to be slow. Is there a faster approach? Is it possible to travel across the particles array only once in parallel and fill MIndex array using atomic operations, something like written in the code piece above?
You probably can't get a compiler to auto-parallelize scalar code for you if you want an algorithm that can work efficiently (without needing atomic RMWs on shared counters which would be a disaster, see below). But you might be able to use OpenMP as a way to start threads and get thread IDs.
Keep per-thread count arrays from the initial histogram, use in 2nd pass
(Update: this might not work: I didn't notice the if(particles[i].ip==3) in the source before. I was assuming that Count[ic] will go as high as Length[ic] in the serial version. If that's not the case, this strategy might leave gaps or something.
But as Laci points out, perhaps you want that check when calculating Length in the first place, then it would be fine.)
Manually multi-thread the first histogram (into Length[]), with each thread working on a known range of i values. Keep those per-thread lengths around, even as you sum across them and prefix-sum to build Begin[].
So Length[thread][ic] is the number of particles in that cube, out of the range of i values that this thread worked on. (And will loop over again in the 2nd loop: the key is that we divide the particles between threads the same way twice. Ideally with the same thread working on the same range, so things may still be hot in L1d cache.)
Pre-process that into a per-thread Begin[][] array, so each thread knows where in MIndex to put data from each bucket.
// pseudo-code, fairly close to actual C
for(ic < cub3) {
// perhaps do this "vertical" sum into a temporary array
// or prefix-sum within Length before combining across threads?
int pos = sum(Length[0..nthreads-1][ic-1]) + Begin[0][ic-1];
Begin[0][ic] = pos;
for (int t = 1 ; t<nthreads ; t++) {
pos += Length[t][ic]; // prefix-sum across threads for this cube bucket
Begin[t][ic] = pos;
}
}
This has a pretty terrible cache access pattern, especially with cuba=8 making Length[t][0] and Length[t+1][0] 4096 bytes apart from each other. (So 4k aliasing is a possible problem, as are cache conflict misses).
Perhaps each thread can prefix-sum its own slice of Length into that slice of Begin, 1. for cache access pattern (and locality since it just wrote those Lengths), and 2. to get some parallelism for that work.
Then in the final loop with MIndex, each thread can do int pos = --Length[t][ic] to derive a unique ID from the Length. (Like you were doing with Count[], but without introducing another per-thread array to zero.)
Each element of Length will return to zero, because the same thread is looking at the same points it just counted. With correctly-calculated Begin[t][ic] positions, MIndex[...] = i stores won't conflict. False sharing is still possible, but it's a large enough array that points will tend to be scattered around.
Don't overdo it with number of threads, especially if cuba is greater than 8. The amount of Length / Begin pre-processing work scales with number of threads, so it may be better to just leave some CPUs free for unrelated threads or tasks to get some throughput done. OTOH, with cuba=8 meaning each per-thread array is only 4096 bytes (too small to parallelize the zeroing of, BTW), it's really not that much.
(Previous answer before your edit made it clearer what was going on.)
Is this basically a histogram? If each thread has its own array of counts, you can sum them together at the end (you might need to do that manually, not have OpenMP do it for you). But it seems you also need this count to be unique within each voxel, to have MIndex updated properly? That might be a showstopper, like requiring adjusting every MIndex entry, if it's even possible.
After your update, you are doing a histogram into Length[], so that part can be sped up.
Atomic RMWs would be necessary for your code as-is, performance disaster
Atomic increments of shared counters would be slower, and on x86 might destroy the memory-level parallelism too badly. On x86, every atomic RMW is also a full memory barrier, draining the store buffer before it happens, and blocking later loads from starting until after it happens.
As opposed to a single thread which can have cache misses to multiple Counter, Begin and MIndex elements outstanding, using non-atomic accesses. (Thanks to out-of-order exec, the next iteration's load / inc / store for Counter[ic]++ can be doing the load while there are cache misses outstanding for Begin[ic] and/or for Mindex[] stores.)
ISAs that allow relaxed-atomic increment might be able to do this efficiently, like AArch64. (Again, OpenMP might not be able to do that for you.)
Even on x86, with enough (logical) cores, you might still get some speedup, especially if the Counter accesses are scattered enough they cores aren't constantly fighting over the same cache lines. You'd still get a lot of cache lines bouncing between cores, though, instead of staying hot in L1d or L2. (False sharing is a problem,
Perhaps software prefetch can help, like prefetchw (write-prefetching) the counter for 5 or 10 i iterations later.
It wouldn't be deterministic which point went in which order, even with memory_order_seq_cst increments, though. Whichever thread increments Counter[ic] first is the one that associates that cnt with that i.
Alternative access patterns
Perhaps have each thread scan all points, but only process a subset of them, with disjoint subsets. So the set of Counter[] elements that any given thread touches is only touched by that thread, so the increments can be non-atomic.
Filtering by p.kz ranges maybe makes the most sense since that's the largest multiplier in the indexing, so each thread "owns" a contiguous range of Counter[].
But if your points aren't uniformly distributed, you'd need to know how to break things up to approximately equally divide the work. And you can't just divide it more finely (like OMP schedule dynamic), since each thread is going to scan through all the points: that would multiply the amount of filtering work.
Maybe a couple fixed partitions would be a good tradeoff to gain some parallelism without introducing a lot of extra work.
Re: your edit
You already loop over the whole array of points doing Length[ic]++;? Seems redundant to do the same histogramming work again with Counter[ic]++;, but not obvious how to avoid it.
The count arrays are small, but if you don't need both when you're done, you could maybe just decrement Length to assign unique indices to each point in a voxel. At least the first histogram could benefit from parallelizing with different count arrays for each thread, and just vertically adding at the end. Should scale perfectly with threads since the count array is small enough for L1d cache.
BTW, for() Length[i]=Counter[i]=0; is too small to be worth parallelizing. For cuba=8, it's 8*8*16 * sizeof(int) = 4096 bytes, just one page, so it's just two small memsets.
(Of course if each thread has their own separate Length array, they each need to zero it). That's small enough to even consider unrolling with maybe 2 count arrays per thread to hide store/reload serial dependencies if a long sequence of points are all in the same bucket. Combining count arrays at the end is a job for #pragma omp simd or just normal auto-vectorization with gcc -O3 -march=native since it's integer work.
For the final loop, you could split your points array in half (assign half to each thread), and have one thread get unique IDs by counting down from --Length[i], and another counting up from 0 in Counter[i]++. With different threads looking at different points, this could give you a factor of 2 speedup. (Modulo contention for MIndex stores.)
To do more than just count up and down, you'd need info you don't have from just the overall Length array... but which you did have temporarily. See the section at the top
You are right to make the update Counter[ic]++ atomic, but there is an additional problem on the next line: MIndex[Begin[ic]+cnt]=i; Different iterations can write into the same location here, unless you have mathematical proof that this is never the case from the structure of MIndex. So you have to make that line atomic too. And then there is almost no parallel work left in your loop, so your speed up if probably going to be abysmal.
EDIT the second line however is not of the right form for an atomic operation, so you have to make it critical. Which is going to make performance even worse.
Also, #Laci is correct that since this is an overwrite statement, the order of parallel scheduling is going to influence the outcome. So either live with that fact, or accept that this can not be parallelized.
I am searching for a high performance C++ structure for a table. The table will have void* as keys and uint32 as values.
The table itself is very small and will not change after creation. The first idea that came to my mind is using something like ska::flat_hash_map<void*, int32_t> or std::unordered_map<void*, int32_t>. However that will be overkill and will not provide me the performance I want (those tables are suited for high number of items too).
So I thought about using std::vector<std::pair<void*, int32_t>>, sorting it upon creation and linear probing it. The next ideas will be using SIMD instructions but it is possible with the current structure.
Another solution which I will shortly evaluate is like that:
struct Group
{
void* items[5]; // search using SIMD
int32_t items[5];
}; // fits in cache line
struct Table
{
Group* groups;
size_t capacity;
};
Are there any better options? I need only 1 operation: finding values by keys, not modifying them, not anything. Thanks!
EDIT: another thing I think I should mention are the access patterns: suppose I have an array of those hash tables, each time I will look up from a random one in the array.
Linear probing is likely the fastest solution in this case on common mainstream architectures, especially since the number of element is very small and bounded (ie. <10). Sorting the items should not speed up the probing with so few items (it would be only useful for a binary search which is much more expensive in this case).
If you want to use SIMD instruction, then you need to use structure of arrays instead of array of structures for the sake of performance. This means you should use std::pair<std::vector<void*>, std::vector<int32_t>> instead of std::vector<std::pair<void*, int32_t>> (which alternates void* types and int32_t values in memory with some padding overhead due to the alignment constraints of void* on 64-bit architectures). Having two std::vector is not great too because you pay its overhead twice. As mentioned by #JorgeBellon
in the comments, you can simply use a std::array instead of std::vector assuming the number of items is known or bounded.
A possible optimization with SIMD instructions is to compact the key pointers on 64-bit architectures by splitting them in 32-bit lower/upper part. Indeed, it is very unlikely that two pointers have the same lower part (least significant bits) while having a different upper part. This tricks help you to check 2 times more pointers at a time.
Note that using SIMD instructions may not be so great in this case in practice. This is especially true if the number of items is smaller than the one fitting in a SIMD vector. For example, with AVX2 (on 86-64 processors), you can work on 4 64-bit values at a time (or 8 32-bit values) but if you have less than 8 values, then you need to mask the unwanted values to check (or even not load them if the memory buffer do not contain some padding). This introduces an additional overhead. This is not much a problem with AVX-512 and SVE (only available on a small fraction of processors yet) since they provides advanced masking operations. Moreover, some processors lower they frequency when they execute SIMD instructions (especially with AVX-512 although the down-clocking is not so strong with integer instructions). SIMD instructions also introduce some additional latency compared to scalar version (which can be better pipelined) and modern processors tends to be able to execute more scalar instructions in parallel than SIMD ones. For all these reasons, it is certainly a good idea to try to write a scalar branchless implementation (possibly unrolled for better performance if the number of items is known at compile time).
You may want to look into perfect hashing -- not too difficult, and can provide simple constant time lookups. It can take technically unbounded time to create the table, though, and it's not as fast as a regular hash table when the regular hash table gets lucky.
I think a nice alternative is an optimization of your simple linear probing idea.
Your lookup procedure would look like this:
Slot *s = &table[hash(key)];
Slot *e = s + s->max_extent;
for (;s<e; ++s) {
if (s->key == key) {
return s->value;
}
}
return NOT_FOUND;
table[h].max_extent is the maximum number of elements you may have to look at if you're looking for an element with hash code h. You would pre-calculate this when you generate the table, so your lookup doesn't have to iterate until it gets a null. This greatly reduces the amount of probing you have to do for misses.
Of course you want max_extent to be as small as possible. Pick a hash result size (at least 2n) to make it <= 1 in most cases, and try a few different hash functions before picking the one that produces the best results by whatever metric you like. You hash can be as simple as key % P, where trying different hashes means trying different P values. Fill your hash table in hash(key) order to produce the best result.
NOTE that we do not wrap around from the end to the start of the table while probing. Just allocate however many extra slots you need to avoid it.
What is the difference between "cache unfriendly code" and the "cache friendly" code?
How can I make sure I write cache-efficient code?
Preliminaries
On modern computers, only the lowest level memory structures (the registers) can move data around in single clock cycles. However, registers are very expensive and most computer cores have less than a few dozen registers. At the other end of the memory spectrum (DRAM), the memory is very cheap (i.e. literally millions of times cheaper) but takes hundreds of cycles after a request to receive the data. To bridge this gap between super fast and expensive and super slow and cheap are the cache memories, named L1, L2, L3 in decreasing speed and cost. The idea is that most of the executing code will be hitting a small set of variables often, and the rest (a much larger set of variables) infrequently. If the processor can't find the data in L1 cache, then it looks in L2 cache. If not there, then L3 cache, and if not there, main memory. Each of these "misses" is expensive in time.
(The analogy is cache memory is to system memory, as system memory is to hard disk storage. Hard disk storage is super cheap but very slow).
Caching is one of the main methods to reduce the impact of latency. To paraphrase Herb Sutter (cfr. links below): increasing bandwidth is easy, but we can't buy our way out of latency.
Data is always retrieved through the memory hierarchy (smallest == fastest to slowest). A cache hit/miss usually refers to a hit/miss in the highest level of cache in the CPU -- by highest level I mean the largest == slowest. The cache hit rate is crucial for performance since every cache miss results in fetching data from RAM (or worse ...) which takes a lot of time (hundreds of cycles for RAM, tens of millions of cycles for HDD). In comparison, reading data from the (highest level) cache typically takes only a handful of cycles.
In modern computer architectures, the performance bottleneck is leaving the CPU die (e.g. accessing RAM or higher). This will only get worse over time. The increase in processor frequency is currently no longer relevant to increase performance. The problem is memory access. Hardware design efforts in CPUs therefore currently focus heavily on optimizing caches, prefetching, pipelines and concurrency. For instance, modern CPUs spend around 85% of die on caches and up to 99% for storing/moving data!
There is quite a lot to be said on the subject. Here are a few great references about caches, memory hierarchies and proper programming:
Agner Fog's page. In his excellent documents, you can find detailed examples covering languages ranging from assembly to C++.
If you are into videos, I strongly recommend to have a look at Herb Sutter's talk on machine architecture (youtube) (specifically check 12:00 and onwards!).
Slides about memory optimization by Christer Ericson (director of technology # Sony)
LWN.net's article "What every programmer should know about memory"
Main concepts for cache-friendly code
A very important aspect of cache-friendly code is all about the principle of locality, the goal of which is to place related data close in memory to allow efficient caching. In terms of the CPU cache, it's important to be aware of cache lines to understand how this works: How do cache lines work?
The following particular aspects are of high importance to optimize caching:
Temporal locality: when a given memory location was accessed, it is likely that the same location is accessed again in the near future. Ideally, this information will still be cached at that point.
Spatial locality: this refers to placing related data close to each other. Caching happens on many levels, not just in the CPU. For example, when you read from RAM, typically a larger chunk of memory is fetched than what was specifically asked for because very often the program will require that data soon. HDD caches follow the same line of thought. Specifically for CPU caches, the notion of cache lines is important.
Use appropriate c++ containers
A simple example of cache-friendly versus cache-unfriendly is c++'s std::vector versus std::list. Elements of a std::vector are stored in contiguous memory, and as such accessing them is much more cache-friendly than accessing elements in a std::list, which stores its content all over the place. This is due to spatial locality.
A very nice illustration of this is given by Bjarne Stroustrup in this youtube clip (thanks to #Mohammad Ali Baydoun for the link!).
Don't neglect the cache in data structure and algorithm design
Whenever possible, try to adapt your data structures and order of computations in a way that allows maximum use of the cache. A common technique in this regard is cache blocking (Archive.org version), which is of extreme importance in high-performance computing (cfr. for example ATLAS).
Know and exploit the implicit structure of data
Another simple example, which many people in the field sometimes forget is column-major (ex. fortran,matlab) vs. row-major ordering (ex. c,c++) for storing two dimensional arrays. For example, consider the following matrix:
1 2
3 4
In row-major ordering, this is stored in memory as 1 2 3 4; in column-major ordering, this would be stored as 1 3 2 4. It is easy to see that implementations which do not exploit this ordering will quickly run into (easily avoidable!) cache issues. Unfortunately, I see stuff like this very often in my domain (machine learning). #MatteoItalia showed this example in more detail in his answer.
When fetching a certain element of a matrix from memory, elements near it will be fetched as well and stored in a cache line. If the ordering is exploited, this will result in fewer memory accesses (because the next few values which are needed for subsequent computations are already in a cache line).
For simplicity, assume the cache comprises a single cache line which can contain 2 matrix elements and that when a given element is fetched from memory, the next one is too. Say we want to take the sum over all elements in the example 2x2 matrix above (lets call it M):
Exploiting the ordering (e.g. changing column index first in c++):
M[0][0] (memory) + M[0][1] (cached) + M[1][0] (memory) + M[1][1] (cached)
= 1 + 2 + 3 + 4
--> 2 cache hits, 2 memory accesses
Not exploiting the ordering (e.g. changing row index first in c++):
M[0][0] (memory) + M[1][0] (memory) + M[0][1] (memory) + M[1][1] (memory)
= 1 + 3 + 2 + 4
--> 0 cache hits, 4 memory accesses
In this simple example, exploiting the ordering approximately doubles execution speed (since memory access requires much more cycles than computing the sums). In practice, the performance difference can be much larger.
Avoid unpredictable branches
Modern architectures feature pipelines and compilers are becoming very good at reordering code to minimize delays due to memory access. When your critical code contains (unpredictable) branches, it is hard or impossible to prefetch data. This will indirectly lead to more cache misses.
This is explained very well here (thanks to #0x90 for the link): Why is processing a sorted array faster than processing an unsorted array?
Avoid virtual functions
In the context of c++, virtual methods represent a controversial issue with regard to cache misses (a general consensus exists that they should be avoided when possible in terms of performance). Virtual functions can induce cache misses during look up, but this only happens if the specific function is not called often (otherwise it would likely be cached), so this is regarded as a non-issue by some. For reference about this issue, check out: What is the performance cost of having a virtual method in a C++ class?
Common problems
A common problem in modern architectures with multiprocessor caches is called false sharing. This occurs when each individual processor is attempting to use data in another memory region and attempts to store it in the same cache line. This causes the cache line -- which contains data another processor can use -- to be overwritten again and again. Effectively, different threads make each other wait by inducing cache misses in this situation.
See also (thanks to #Matt for the link): How and when to align to cache line size?
An extreme symptom of poor caching in RAM memory (which is probably not what you mean in this context) is so-called thrashing. This occurs when the process continuously generates page faults (e.g. accesses memory which is not in the current page) which require disk access.
In addition to #Marc Claesen's answer, I think that an instructive classic example of cache-unfriendly code is code that scans a C bidimensional array (e.g. a bitmap image) column-wise instead of row-wise.
Elements that are adjacent in a row are also adjacent in memory, thus accessing them in sequence means accessing them in ascending memory order; this is cache-friendly, since the cache tends to prefetch contiguous blocks of memory.
Instead, accessing such elements column-wise is cache-unfriendly, since elements on the same column are distant in memory from each other (in particular, their distance is equal to the size of the row), so when you use this access pattern you are jumping around in memory, potentially wasting the effort of the cache of retrieving the elements nearby in memory.
And all that it takes to ruin the performance is to go from
// Cache-friendly version - processes pixels which are adjacent in memory
for(unsigned int y=0; y<height; ++y)
{
for(unsigned int x=0; x<width; ++x)
{
... image[y][x] ...
}
}
to
// Cache-unfriendly version - jumps around in memory for no good reason
for(unsigned int x=0; x<width; ++x)
{
for(unsigned int y=0; y<height; ++y)
{
... image[y][x] ...
}
}
This effect can be quite dramatic (several order of magnitudes in speed) in systems with small caches and/or working with big arrays (e.g. 10+ megapixels 24 bpp images on current machines); for this reason, if you have to do many vertical scans, often it's better to rotate the image of 90 degrees first and perform the various analysis later, limiting the cache-unfriendly code just to the rotation.
Optimizing cache usage largely comes down to two factors.
Locality of Reference
The first factor (to which others have already alluded) is locality of reference. Locality of reference really has two dimensions though: space and time.
Spatial
The spatial dimension also comes down to two things: first, we want to pack our information densely, so more information will fit in that limited memory. This means (for example) that you need a major improvement in computational complexity to justify data structures based on small nodes joined by pointers.
Second, we want information that will be processed together also located together. A typical cache works in "lines", which means when you access some information, other information at nearby addresses will be loaded into the cache with the part we touched. For example, when I touch one byte, the cache might load 128 or 256 bytes near that one. To take advantage of that, you generally want the data arranged to maximize the likelihood that you'll also use that other data that was loaded at the same time.
For just a really trivial example, this can mean that a linear search can be much more competitive with a binary search than you'd expect. Once you've loaded one item from a cache line, using the rest of the data in that cache line is almost free. A binary search becomes noticeably faster only when the data is large enough that the binary search reduces the number of cache lines you access.
Time
The time dimension means that when you do some operations on some data, you want (as much as possible) to do all the operations on that data at once.
Since you've tagged this as C++, I'll point to a classic example of a relatively cache-unfriendly design: std::valarray. valarray overloads most arithmetic operators, so I can (for example) say a = b + c + d; (where a, b, c and d are all valarrays) to do element-wise addition of those arrays.
The problem with this is that it walks through one pair of inputs, puts results in a temporary, walks through another pair of inputs, and so on. With a lot of data, the result from one computation may disappear from the cache before it's used in the next computation, so we end up reading (and writing) the data repeatedly before we get our final result. If each element of the final result will be something like (a[n] + b[n]) * (c[n] + d[n]);, we'd generally prefer to read each a[n], b[n], c[n] and d[n] once, do the computation, write the result, increment n and repeat 'til we're done.2
Line Sharing
The second major factor is avoiding line sharing. To understand this, we probably need to back up and look a little at how caches are organized. The simplest form of cache is direct mapped. This means one address in main memory can only be stored in one specific spot in the cache. If we're using two data items that map to the same spot in the cache, it works badly -- each time we use one data item, the other has to be flushed from the cache to make room for the other. The rest of the cache might be empty, but those items won't use other parts of the cache.
To prevent this, most caches are what are called "set associative". For example, in a 4-way set-associative cache, any item from main memory can be stored at any of 4 different places in the cache. So, when the cache is going to load an item, it looks for the least recently used3 item among those four, flushes it to main memory, and loads the new item in its place.
The problem is probably fairly obvious: for a direct-mapped cache, two operands that happen to map to the same cache location can lead to bad behavior. An N-way set-associative cache increases the number from 2 to N+1. Organizing a cache into more "ways" takes extra circuitry and generally runs slower, so (for example) an 8192-way set associative cache is rarely a good solution either.
Ultimately, this factor is more difficult to control in portable code though. Your control over where your data is placed is usually fairly limited. Worse, the exact mapping from address to cache varies between otherwise similar processors. In some cases, however, it can be worth doing things like allocating a large buffer, and then using only parts of what you allocated to ensure against data sharing the same cache lines (even though you'll probably need to detect the exact processor and act accordingly to do this).
False Sharing
There's another, related item called "false sharing". This arises in a multiprocessor or multicore system, where two (or more) processors/cores have data that's separate, but falls in the same cache line. This forces the two processors/cores to coordinate their access to the data, even though each has its own, separate data item. Especially if the two modify the data in alternation, this can lead to a massive slowdown as the data has to be constantly shuttled between the processors. This can't easily be cured by organizing the cache into more "ways" or anything like that either. The primary way to prevent it is to ensure that two threads rarely (preferably never) modify data that could possibly be in the same cache line (with the same caveats about difficulty of controlling the addresses at which data is allocated).
Those who know C++ well might wonder if this is open to optimization via something like expression templates. I'm pretty sure the answer is that yes, it could be done and if it was, it would probably be a pretty substantial win. I'm not aware of anybody having done so, however, and given how little valarray gets used, I'd be at least a little surprised to see anybody do so either.
In case anybody wonders how valarray (designed specifically for performance) could be this badly wrong, it comes down to one thing: it was really designed for machines like the older Crays, that used fast main memory and no cache. For them, this really was a nearly ideal design.
Yes, I'm simplifying: most caches don't really measure the least recently used item precisely, but they use some heuristic that's intended to be close to that without having to keep a full time-stamp for each access.
Welcome to the world of Data Oriented Design. The basic mantra is to Sort, Eliminate Branches, Batch, Eliminate virtual calls - all steps towards better locality.
Since you tagged the question with C++, here's the obligatory typical C++ Bullshit. Tony Albrecht's Pitfalls of Object Oriented Programming is also a great introduction into the subject.
Just piling on: the classic example of cache-unfriendly versus cache-friendly code is the "cache blocking" of matrix multiply.
Naive matrix multiply looks like:
for(i=0;i<N;i++) {
for(j=0;j<N;j++) {
dest[i][j] = 0;
for( k=0;k<N;k++) {
dest[i][j] += src1[i][k] * src2[k][j];
}
}
}
If N is large, e.g. if N * sizeof(elemType) is greater than the cache size, then every single access to src2[k][j] will be a cache miss.
There are many different ways of optimizing this for a cache. Here's a very simple example: instead of reading one item per cache line in the inner loop, use all of the items:
int itemsPerCacheLine = CacheLineSize / sizeof(elemType);
for(i=0;i<N;i++) {
for(j=0;j<N;j += itemsPerCacheLine ) {
for(jj=0;jj<itemsPerCacheLine; jj+) {
dest[i][j+jj] = 0;
}
for( k=0;k<N;k++) {
for(jj=0;jj<itemsPerCacheLine; jj+) {
dest[i][j+jj] += src1[i][k] * src2[k][j+jj];
}
}
}
}
If the cache line size is 64 bytes, and we are operating on 32 bit (4 byte) floats, then there are 16 items per cache line. And the number of cache misses via just this simple transformation is reduced approximately 16-fold.
Fancier transformations operate on 2D tiles, optimize for multiple caches (L1, L2, TLB), and so on.
Some results of googling "cache blocking":
http://stumptown.cc.gt.atl.ga.us/cse6230-hpcta-fa11/slides/11a-matmul-goto.pdf
http://software.intel.com/en-us/articles/cache-blocking-techniques
A nice video animation of an optimized cache blocking algorithm.
http://www.youtube.com/watch?v=IFWgwGMMrh0
Loop tiling is very closely related:
http://en.wikipedia.org/wiki/Loop_tiling
Processors today work with many levels of cascading memory areas. So the CPU will have a bunch of memory that is on the CPU chip itself. It has very fast access to this memory. There are different levels of cache each one slower access ( and larger ) than the next, until you get to system memory which is not on the CPU and is relatively much slower to access.
Logically, to the CPU's instruction set you just refer to memory addresses in a giant virtual address space. When you access a single memory address the CPU will go fetch it. in the old days it would fetch just that single address. But today the CPU will fetch a bunch of memory around the bit you asked for, and copy it into the cache. It assumes that if you asked for a particular address that is is highly likely that you are going to ask for an address nearby very soon. For example if you were copying a buffer you would read and write from consecutive addresses - one right after the other.
So today when you fetch an address it checks the first level of cache to see if it already read that address into cache, if it doesn't find it, then this is a cache miss and it has to go out to the next level of cache to find it, until it eventually has to go out into main memory.
Cache friendly code tries to keep accesses close together in memory so that you minimize cache misses.
So an example would be imagine you wanted to copy a giant 2 dimensional table. It is organized with reach row in consecutive in memory, and one row follow the next right after.
If you copied the elements one row at a time from left to right - that would be cache friendly. If you decided to copy the table one column at a time, you would copy the exact same amount of memory - but it would be cache unfriendly.
It needs to be clarified that not only data should be cache-friendly, it is just as important for the code. This is in addition to branch predicition, instruction reordering, avoiding actual divisions and other techniques.
Typically the denser the code, the fewer cache lines will be required to store it. This results in more cache lines being available for data.
The code should not call functions all over the place as they typically will require one or more cache lines of their own, resulting in fewer cache lines for data.
A function should begin at a cache line-alignment-friendly address. Though there are (gcc) compiler switches for this be aware that if the the functions are very short it might be wasteful for each one to occupy an entire cache line. For example, if three of the most often used functions fit inside one 64 byte cache line, this is less wasteful than if each one has its own line and results in two cache lines less available for other usage. A typical alignment value could be 32 or 16.
So spend some extra time to make the code dense. Test different constructs, compile and review the generated code size and profile.
As #Marc Claesen mentioned that one of the ways to write cache friendly code is to exploit the structure in which our data is stored. In addition to that another way to write cache friendly code is: change the way our data is stored; then write new code to access the data stored in this new structure.
This makes sense in the case of how database systems linearize the tuples of a table and store them. There are two basic ways to store the tuples of a table i.e. row store and column store. In row store as the name suggests the tuples are stored row wise. Lets suppose a table named Product being stored has 3 attributes i.e. int32_t key, char name[56] and int32_t price, so the total size of a tuple is 64 bytes.
We can simulate a very basic row store query execution in main memory by creating an array of Product structs with size N, where N is the number of rows in table. Such memory layout is also called array of structs. So the struct for Product can be like:
struct Product
{
int32_t key;
char name[56];
int32_t price'
}
/* create an array of structs */
Product* table = new Product[N];
/* now load this array of structs, from a file etc. */
Similarly we can simulate a very basic column store query execution in main memory by creating an 3 arrays of size N, one array for each attribute of the Product table. Such memory layout is also called struct of arrays. So the 3 arrays for each attribute of Product can be like:
/* create separate arrays for each attribute */
int32_t* key = new int32_t[N];
char* name = new char[56*N];
int32_t* price = new int32_t[N];
/* now load these arrays, from a file etc. */
Now after loading both the array of structs (Row Layout) and the 3 separate arrays (Column Layout), we have row store and column store on our table Product present in our memory.
Now we move on to the cache friendly code part. Suppose that the workload on our table is such that we have an aggregation query on the price attribute. Such as
SELECT SUM(price)
FROM PRODUCT
For the row store we can convert the above SQL query into
int sum = 0;
for (int i=0; i<N; i++)
sum = sum + table[i].price;
For the column store we can convert the above SQL query into
int sum = 0;
for (int i=0; i<N; i++)
sum = sum + price[i];
The code for the column store would be faster than the code for the row layout in this query as it requires only a subset of attributes and in column layout we are doing just that i.e. only accessing the price column.
Suppose that the cache line size is 64 bytes.
In the case of row layout when a cache line is read, the price value of only 1(cacheline_size/product_struct_size = 64/64 = 1) tuple is read, because our struct size of 64 bytes and it fills our whole cache line, so for every tuple a cache miss occurs in case of a row layout.
In the case of column layout when a cache line is read, the price value of 16(cacheline_size/price_int_size = 64/4 = 16) tuples is read, because 16 contiguous price values stored in memory are brought into the cache, so for every sixteenth tuple a cache miss ocurs in case of column layout.
So the column layout will be faster in the case of given query, and is faster in such aggregation queries on a subset of columns of the table. You can try out such experiment for yourself using the data from TPC-H benchmark, and compare the run times for both the layouts. The wikipedia article on column oriented database systems is also good.
So in database systems, if the query workload is known beforehand, we can store our data in layouts which will suit the queries in workload and access data from these layouts. In the case of above example we created a column layout and changed our code to compute sum so that it became cache friendly.
Be aware that caches do not just cache continuous memory. They have multiple lines (at least 4) so discontinous and overlapping memory can often be stored just as efficiently.
What is missing from all the above examples is measured benchmarks. There are many myths about performance. Unless you measure it you do not know. Do not complicate your code unless you have a measured improvement.
Cache-friendly code is code that has been optimized to make efficient use of the CPU cache. This typically involves organizing data in a way that takes advantage of spatial and temporal locality, which refers to the idea that data that is accessed together is likely to be stored together in memory, and that data that is accessed frequently is likely to be accessed again in the near future.
There are several ways to make code cache-friendly, including:
Using contiguous memory layouts: By storing data in contiguous
blocks in memory, you can take advantage of spatial locality and
reduce the number of cache misses.
Using arrays: Arrays are a good choice for data structures when you
need to access data sequentially, as they allow you to take
advantage of temporal locality and keep hot data in the cache.
Using pointers carefully: Pointers can be used to access data that
is not stored contiguously in memory, but they can also lead to
cache misses if they are used excessively. If you need to use
pointers, try to use them in a way that takes advantage of spatial
and temporal locality to minimize cache misses.
Using compiler optimization flags: Most compilers have optimization
flags that can be used to optimize the use of the CPU cache. These
flags can help to minimize the number of cache misses and improve
the overall performance of your code.
It is important to note that the specific techniques that work best for optimizing the use of the CPU cache will depend on the specific requirements and constraints of your system. It may be necessary to experiment with different approaches to find the best solution for your needs.
general question: the number of threads must be equal to the size of the elements i want to deal with? exmaple: if i have matrix M[a][b]. i must allocate (aXb) threads or i can allocate more threads than i need(more than ab)? because the thread that will focus on element aXb+1 will throw us out, doesnt he? or the solution is to put a condition(only if in range(ab))?
specific question: let be M[x][y] matrix with x rows and y columns. consider that 1000 <= x <= 300000 and y <= 100. how can i organize the threads in that way that it will be general for each input for x and y. i want that each thread will focus on one element in the matrix. CC = 2.1 thanks!
General answer: It depends on a problem.
In most cases natural one-to-one mapping of the problem to the grid of threads is fine to start with, but what you want to keep in mind is:
Achieving high occupancy.
Maximizing GPU resources usage and memory throughput.
Working with valid data.
Sometimes it may require using single thread to process many elements or many threads to process single element.
For instance, you can imagine an series of independent operations A,B and C that need to be applied on array of elements. You could run three different kernels, but it might be better choice to allocate the grid to contain three times more threads than there is elements and distinguish operations by one of the dimensions of the grid (or anything else). On the other side - you might have a problem that could use maximizing the usage of shared memory (e.g transforming the image) - you could use block of 16 threads to process 5x5 image window where each thread would calculate some statistics of each 2x2 slice.
The choice is yours - the best advice is not always go with the obvious. Try different approaches and choose what works best.
What is the difference between "cache unfriendly code" and the "cache friendly" code?
How can I make sure I write cache-efficient code?
Preliminaries
On modern computers, only the lowest level memory structures (the registers) can move data around in single clock cycles. However, registers are very expensive and most computer cores have less than a few dozen registers. At the other end of the memory spectrum (DRAM), the memory is very cheap (i.e. literally millions of times cheaper) but takes hundreds of cycles after a request to receive the data. To bridge this gap between super fast and expensive and super slow and cheap are the cache memories, named L1, L2, L3 in decreasing speed and cost. The idea is that most of the executing code will be hitting a small set of variables often, and the rest (a much larger set of variables) infrequently. If the processor can't find the data in L1 cache, then it looks in L2 cache. If not there, then L3 cache, and if not there, main memory. Each of these "misses" is expensive in time.
(The analogy is cache memory is to system memory, as system memory is to hard disk storage. Hard disk storage is super cheap but very slow).
Caching is one of the main methods to reduce the impact of latency. To paraphrase Herb Sutter (cfr. links below): increasing bandwidth is easy, but we can't buy our way out of latency.
Data is always retrieved through the memory hierarchy (smallest == fastest to slowest). A cache hit/miss usually refers to a hit/miss in the highest level of cache in the CPU -- by highest level I mean the largest == slowest. The cache hit rate is crucial for performance since every cache miss results in fetching data from RAM (or worse ...) which takes a lot of time (hundreds of cycles for RAM, tens of millions of cycles for HDD). In comparison, reading data from the (highest level) cache typically takes only a handful of cycles.
In modern computer architectures, the performance bottleneck is leaving the CPU die (e.g. accessing RAM or higher). This will only get worse over time. The increase in processor frequency is currently no longer relevant to increase performance. The problem is memory access. Hardware design efforts in CPUs therefore currently focus heavily on optimizing caches, prefetching, pipelines and concurrency. For instance, modern CPUs spend around 85% of die on caches and up to 99% for storing/moving data!
There is quite a lot to be said on the subject. Here are a few great references about caches, memory hierarchies and proper programming:
Agner Fog's page. In his excellent documents, you can find detailed examples covering languages ranging from assembly to C++.
If you are into videos, I strongly recommend to have a look at Herb Sutter's talk on machine architecture (youtube) (specifically check 12:00 and onwards!).
Slides about memory optimization by Christer Ericson (director of technology # Sony)
LWN.net's article "What every programmer should know about memory"
Main concepts for cache-friendly code
A very important aspect of cache-friendly code is all about the principle of locality, the goal of which is to place related data close in memory to allow efficient caching. In terms of the CPU cache, it's important to be aware of cache lines to understand how this works: How do cache lines work?
The following particular aspects are of high importance to optimize caching:
Temporal locality: when a given memory location was accessed, it is likely that the same location is accessed again in the near future. Ideally, this information will still be cached at that point.
Spatial locality: this refers to placing related data close to each other. Caching happens on many levels, not just in the CPU. For example, when you read from RAM, typically a larger chunk of memory is fetched than what was specifically asked for because very often the program will require that data soon. HDD caches follow the same line of thought. Specifically for CPU caches, the notion of cache lines is important.
Use appropriate c++ containers
A simple example of cache-friendly versus cache-unfriendly is c++'s std::vector versus std::list. Elements of a std::vector are stored in contiguous memory, and as such accessing them is much more cache-friendly than accessing elements in a std::list, which stores its content all over the place. This is due to spatial locality.
A very nice illustration of this is given by Bjarne Stroustrup in this youtube clip (thanks to #Mohammad Ali Baydoun for the link!).
Don't neglect the cache in data structure and algorithm design
Whenever possible, try to adapt your data structures and order of computations in a way that allows maximum use of the cache. A common technique in this regard is cache blocking (Archive.org version), which is of extreme importance in high-performance computing (cfr. for example ATLAS).
Know and exploit the implicit structure of data
Another simple example, which many people in the field sometimes forget is column-major (ex. fortran,matlab) vs. row-major ordering (ex. c,c++) for storing two dimensional arrays. For example, consider the following matrix:
1 2
3 4
In row-major ordering, this is stored in memory as 1 2 3 4; in column-major ordering, this would be stored as 1 3 2 4. It is easy to see that implementations which do not exploit this ordering will quickly run into (easily avoidable!) cache issues. Unfortunately, I see stuff like this very often in my domain (machine learning). #MatteoItalia showed this example in more detail in his answer.
When fetching a certain element of a matrix from memory, elements near it will be fetched as well and stored in a cache line. If the ordering is exploited, this will result in fewer memory accesses (because the next few values which are needed for subsequent computations are already in a cache line).
For simplicity, assume the cache comprises a single cache line which can contain 2 matrix elements and that when a given element is fetched from memory, the next one is too. Say we want to take the sum over all elements in the example 2x2 matrix above (lets call it M):
Exploiting the ordering (e.g. changing column index first in c++):
M[0][0] (memory) + M[0][1] (cached) + M[1][0] (memory) + M[1][1] (cached)
= 1 + 2 + 3 + 4
--> 2 cache hits, 2 memory accesses
Not exploiting the ordering (e.g. changing row index first in c++):
M[0][0] (memory) + M[1][0] (memory) + M[0][1] (memory) + M[1][1] (memory)
= 1 + 3 + 2 + 4
--> 0 cache hits, 4 memory accesses
In this simple example, exploiting the ordering approximately doubles execution speed (since memory access requires much more cycles than computing the sums). In practice, the performance difference can be much larger.
Avoid unpredictable branches
Modern architectures feature pipelines and compilers are becoming very good at reordering code to minimize delays due to memory access. When your critical code contains (unpredictable) branches, it is hard or impossible to prefetch data. This will indirectly lead to more cache misses.
This is explained very well here (thanks to #0x90 for the link): Why is processing a sorted array faster than processing an unsorted array?
Avoid virtual functions
In the context of c++, virtual methods represent a controversial issue with regard to cache misses (a general consensus exists that they should be avoided when possible in terms of performance). Virtual functions can induce cache misses during look up, but this only happens if the specific function is not called often (otherwise it would likely be cached), so this is regarded as a non-issue by some. For reference about this issue, check out: What is the performance cost of having a virtual method in a C++ class?
Common problems
A common problem in modern architectures with multiprocessor caches is called false sharing. This occurs when each individual processor is attempting to use data in another memory region and attempts to store it in the same cache line. This causes the cache line -- which contains data another processor can use -- to be overwritten again and again. Effectively, different threads make each other wait by inducing cache misses in this situation.
See also (thanks to #Matt for the link): How and when to align to cache line size?
An extreme symptom of poor caching in RAM memory (which is probably not what you mean in this context) is so-called thrashing. This occurs when the process continuously generates page faults (e.g. accesses memory which is not in the current page) which require disk access.
In addition to #Marc Claesen's answer, I think that an instructive classic example of cache-unfriendly code is code that scans a C bidimensional array (e.g. a bitmap image) column-wise instead of row-wise.
Elements that are adjacent in a row are also adjacent in memory, thus accessing them in sequence means accessing them in ascending memory order; this is cache-friendly, since the cache tends to prefetch contiguous blocks of memory.
Instead, accessing such elements column-wise is cache-unfriendly, since elements on the same column are distant in memory from each other (in particular, their distance is equal to the size of the row), so when you use this access pattern you are jumping around in memory, potentially wasting the effort of the cache of retrieving the elements nearby in memory.
And all that it takes to ruin the performance is to go from
// Cache-friendly version - processes pixels which are adjacent in memory
for(unsigned int y=0; y<height; ++y)
{
for(unsigned int x=0; x<width; ++x)
{
... image[y][x] ...
}
}
to
// Cache-unfriendly version - jumps around in memory for no good reason
for(unsigned int x=0; x<width; ++x)
{
for(unsigned int y=0; y<height; ++y)
{
... image[y][x] ...
}
}
This effect can be quite dramatic (several order of magnitudes in speed) in systems with small caches and/or working with big arrays (e.g. 10+ megapixels 24 bpp images on current machines); for this reason, if you have to do many vertical scans, often it's better to rotate the image of 90 degrees first and perform the various analysis later, limiting the cache-unfriendly code just to the rotation.
Optimizing cache usage largely comes down to two factors.
Locality of Reference
The first factor (to which others have already alluded) is locality of reference. Locality of reference really has two dimensions though: space and time.
Spatial
The spatial dimension also comes down to two things: first, we want to pack our information densely, so more information will fit in that limited memory. This means (for example) that you need a major improvement in computational complexity to justify data structures based on small nodes joined by pointers.
Second, we want information that will be processed together also located together. A typical cache works in "lines", which means when you access some information, other information at nearby addresses will be loaded into the cache with the part we touched. For example, when I touch one byte, the cache might load 128 or 256 bytes near that one. To take advantage of that, you generally want the data arranged to maximize the likelihood that you'll also use that other data that was loaded at the same time.
For just a really trivial example, this can mean that a linear search can be much more competitive with a binary search than you'd expect. Once you've loaded one item from a cache line, using the rest of the data in that cache line is almost free. A binary search becomes noticeably faster only when the data is large enough that the binary search reduces the number of cache lines you access.
Time
The time dimension means that when you do some operations on some data, you want (as much as possible) to do all the operations on that data at once.
Since you've tagged this as C++, I'll point to a classic example of a relatively cache-unfriendly design: std::valarray. valarray overloads most arithmetic operators, so I can (for example) say a = b + c + d; (where a, b, c and d are all valarrays) to do element-wise addition of those arrays.
The problem with this is that it walks through one pair of inputs, puts results in a temporary, walks through another pair of inputs, and so on. With a lot of data, the result from one computation may disappear from the cache before it's used in the next computation, so we end up reading (and writing) the data repeatedly before we get our final result. If each element of the final result will be something like (a[n] + b[n]) * (c[n] + d[n]);, we'd generally prefer to read each a[n], b[n], c[n] and d[n] once, do the computation, write the result, increment n and repeat 'til we're done.2
Line Sharing
The second major factor is avoiding line sharing. To understand this, we probably need to back up and look a little at how caches are organized. The simplest form of cache is direct mapped. This means one address in main memory can only be stored in one specific spot in the cache. If we're using two data items that map to the same spot in the cache, it works badly -- each time we use one data item, the other has to be flushed from the cache to make room for the other. The rest of the cache might be empty, but those items won't use other parts of the cache.
To prevent this, most caches are what are called "set associative". For example, in a 4-way set-associative cache, any item from main memory can be stored at any of 4 different places in the cache. So, when the cache is going to load an item, it looks for the least recently used3 item among those four, flushes it to main memory, and loads the new item in its place.
The problem is probably fairly obvious: for a direct-mapped cache, two operands that happen to map to the same cache location can lead to bad behavior. An N-way set-associative cache increases the number from 2 to N+1. Organizing a cache into more "ways" takes extra circuitry and generally runs slower, so (for example) an 8192-way set associative cache is rarely a good solution either.
Ultimately, this factor is more difficult to control in portable code though. Your control over where your data is placed is usually fairly limited. Worse, the exact mapping from address to cache varies between otherwise similar processors. In some cases, however, it can be worth doing things like allocating a large buffer, and then using only parts of what you allocated to ensure against data sharing the same cache lines (even though you'll probably need to detect the exact processor and act accordingly to do this).
False Sharing
There's another, related item called "false sharing". This arises in a multiprocessor or multicore system, where two (or more) processors/cores have data that's separate, but falls in the same cache line. This forces the two processors/cores to coordinate their access to the data, even though each has its own, separate data item. Especially if the two modify the data in alternation, this can lead to a massive slowdown as the data has to be constantly shuttled between the processors. This can't easily be cured by organizing the cache into more "ways" or anything like that either. The primary way to prevent it is to ensure that two threads rarely (preferably never) modify data that could possibly be in the same cache line (with the same caveats about difficulty of controlling the addresses at which data is allocated).
Those who know C++ well might wonder if this is open to optimization via something like expression templates. I'm pretty sure the answer is that yes, it could be done and if it was, it would probably be a pretty substantial win. I'm not aware of anybody having done so, however, and given how little valarray gets used, I'd be at least a little surprised to see anybody do so either.
In case anybody wonders how valarray (designed specifically for performance) could be this badly wrong, it comes down to one thing: it was really designed for machines like the older Crays, that used fast main memory and no cache. For them, this really was a nearly ideal design.
Yes, I'm simplifying: most caches don't really measure the least recently used item precisely, but they use some heuristic that's intended to be close to that without having to keep a full time-stamp for each access.
Welcome to the world of Data Oriented Design. The basic mantra is to Sort, Eliminate Branches, Batch, Eliminate virtual calls - all steps towards better locality.
Since you tagged the question with C++, here's the obligatory typical C++ Bullshit. Tony Albrecht's Pitfalls of Object Oriented Programming is also a great introduction into the subject.
Just piling on: the classic example of cache-unfriendly versus cache-friendly code is the "cache blocking" of matrix multiply.
Naive matrix multiply looks like:
for(i=0;i<N;i++) {
for(j=0;j<N;j++) {
dest[i][j] = 0;
for( k=0;k<N;k++) {
dest[i][j] += src1[i][k] * src2[k][j];
}
}
}
If N is large, e.g. if N * sizeof(elemType) is greater than the cache size, then every single access to src2[k][j] will be a cache miss.
There are many different ways of optimizing this for a cache. Here's a very simple example: instead of reading one item per cache line in the inner loop, use all of the items:
int itemsPerCacheLine = CacheLineSize / sizeof(elemType);
for(i=0;i<N;i++) {
for(j=0;j<N;j += itemsPerCacheLine ) {
for(jj=0;jj<itemsPerCacheLine; jj+) {
dest[i][j+jj] = 0;
}
for( k=0;k<N;k++) {
for(jj=0;jj<itemsPerCacheLine; jj+) {
dest[i][j+jj] += src1[i][k] * src2[k][j+jj];
}
}
}
}
If the cache line size is 64 bytes, and we are operating on 32 bit (4 byte) floats, then there are 16 items per cache line. And the number of cache misses via just this simple transformation is reduced approximately 16-fold.
Fancier transformations operate on 2D tiles, optimize for multiple caches (L1, L2, TLB), and so on.
Some results of googling "cache blocking":
http://stumptown.cc.gt.atl.ga.us/cse6230-hpcta-fa11/slides/11a-matmul-goto.pdf
http://software.intel.com/en-us/articles/cache-blocking-techniques
A nice video animation of an optimized cache blocking algorithm.
http://www.youtube.com/watch?v=IFWgwGMMrh0
Loop tiling is very closely related:
http://en.wikipedia.org/wiki/Loop_tiling
Processors today work with many levels of cascading memory areas. So the CPU will have a bunch of memory that is on the CPU chip itself. It has very fast access to this memory. There are different levels of cache each one slower access ( and larger ) than the next, until you get to system memory which is not on the CPU and is relatively much slower to access.
Logically, to the CPU's instruction set you just refer to memory addresses in a giant virtual address space. When you access a single memory address the CPU will go fetch it. in the old days it would fetch just that single address. But today the CPU will fetch a bunch of memory around the bit you asked for, and copy it into the cache. It assumes that if you asked for a particular address that is is highly likely that you are going to ask for an address nearby very soon. For example if you were copying a buffer you would read and write from consecutive addresses - one right after the other.
So today when you fetch an address it checks the first level of cache to see if it already read that address into cache, if it doesn't find it, then this is a cache miss and it has to go out to the next level of cache to find it, until it eventually has to go out into main memory.
Cache friendly code tries to keep accesses close together in memory so that you minimize cache misses.
So an example would be imagine you wanted to copy a giant 2 dimensional table. It is organized with reach row in consecutive in memory, and one row follow the next right after.
If you copied the elements one row at a time from left to right - that would be cache friendly. If you decided to copy the table one column at a time, you would copy the exact same amount of memory - but it would be cache unfriendly.
It needs to be clarified that not only data should be cache-friendly, it is just as important for the code. This is in addition to branch predicition, instruction reordering, avoiding actual divisions and other techniques.
Typically the denser the code, the fewer cache lines will be required to store it. This results in more cache lines being available for data.
The code should not call functions all over the place as they typically will require one or more cache lines of their own, resulting in fewer cache lines for data.
A function should begin at a cache line-alignment-friendly address. Though there are (gcc) compiler switches for this be aware that if the the functions are very short it might be wasteful for each one to occupy an entire cache line. For example, if three of the most often used functions fit inside one 64 byte cache line, this is less wasteful than if each one has its own line and results in two cache lines less available for other usage. A typical alignment value could be 32 or 16.
So spend some extra time to make the code dense. Test different constructs, compile and review the generated code size and profile.
As #Marc Claesen mentioned that one of the ways to write cache friendly code is to exploit the structure in which our data is stored. In addition to that another way to write cache friendly code is: change the way our data is stored; then write new code to access the data stored in this new structure.
This makes sense in the case of how database systems linearize the tuples of a table and store them. There are two basic ways to store the tuples of a table i.e. row store and column store. In row store as the name suggests the tuples are stored row wise. Lets suppose a table named Product being stored has 3 attributes i.e. int32_t key, char name[56] and int32_t price, so the total size of a tuple is 64 bytes.
We can simulate a very basic row store query execution in main memory by creating an array of Product structs with size N, where N is the number of rows in table. Such memory layout is also called array of structs. So the struct for Product can be like:
struct Product
{
int32_t key;
char name[56];
int32_t price'
}
/* create an array of structs */
Product* table = new Product[N];
/* now load this array of structs, from a file etc. */
Similarly we can simulate a very basic column store query execution in main memory by creating an 3 arrays of size N, one array for each attribute of the Product table. Such memory layout is also called struct of arrays. So the 3 arrays for each attribute of Product can be like:
/* create separate arrays for each attribute */
int32_t* key = new int32_t[N];
char* name = new char[56*N];
int32_t* price = new int32_t[N];
/* now load these arrays, from a file etc. */
Now after loading both the array of structs (Row Layout) and the 3 separate arrays (Column Layout), we have row store and column store on our table Product present in our memory.
Now we move on to the cache friendly code part. Suppose that the workload on our table is such that we have an aggregation query on the price attribute. Such as
SELECT SUM(price)
FROM PRODUCT
For the row store we can convert the above SQL query into
int sum = 0;
for (int i=0; i<N; i++)
sum = sum + table[i].price;
For the column store we can convert the above SQL query into
int sum = 0;
for (int i=0; i<N; i++)
sum = sum + price[i];
The code for the column store would be faster than the code for the row layout in this query as it requires only a subset of attributes and in column layout we are doing just that i.e. only accessing the price column.
Suppose that the cache line size is 64 bytes.
In the case of row layout when a cache line is read, the price value of only 1(cacheline_size/product_struct_size = 64/64 = 1) tuple is read, because our struct size of 64 bytes and it fills our whole cache line, so for every tuple a cache miss occurs in case of a row layout.
In the case of column layout when a cache line is read, the price value of 16(cacheline_size/price_int_size = 64/4 = 16) tuples is read, because 16 contiguous price values stored in memory are brought into the cache, so for every sixteenth tuple a cache miss ocurs in case of column layout.
So the column layout will be faster in the case of given query, and is faster in such aggregation queries on a subset of columns of the table. You can try out such experiment for yourself using the data from TPC-H benchmark, and compare the run times for both the layouts. The wikipedia article on column oriented database systems is also good.
So in database systems, if the query workload is known beforehand, we can store our data in layouts which will suit the queries in workload and access data from these layouts. In the case of above example we created a column layout and changed our code to compute sum so that it became cache friendly.
Be aware that caches do not just cache continuous memory. They have multiple lines (at least 4) so discontinous and overlapping memory can often be stored just as efficiently.
What is missing from all the above examples is measured benchmarks. There are many myths about performance. Unless you measure it you do not know. Do not complicate your code unless you have a measured improvement.
Cache-friendly code is code that has been optimized to make efficient use of the CPU cache. This typically involves organizing data in a way that takes advantage of spatial and temporal locality, which refers to the idea that data that is accessed together is likely to be stored together in memory, and that data that is accessed frequently is likely to be accessed again in the near future.
There are several ways to make code cache-friendly, including:
Using contiguous memory layouts: By storing data in contiguous
blocks in memory, you can take advantage of spatial locality and
reduce the number of cache misses.
Using arrays: Arrays are a good choice for data structures when you
need to access data sequentially, as they allow you to take
advantage of temporal locality and keep hot data in the cache.
Using pointers carefully: Pointers can be used to access data that
is not stored contiguously in memory, but they can also lead to
cache misses if they are used excessively. If you need to use
pointers, try to use them in a way that takes advantage of spatial
and temporal locality to minimize cache misses.
Using compiler optimization flags: Most compilers have optimization
flags that can be used to optimize the use of the CPU cache. These
flags can help to minimize the number of cache misses and improve
the overall performance of your code.
It is important to note that the specific techniques that work best for optimizing the use of the CPU cache will depend on the specific requirements and constraints of your system. It may be necessary to experiment with different approaches to find the best solution for your needs.