Impact of binary code length on performance of CUDA program - performance

I wrote a CUDA program that has several subroutines. When I disable subroutine A, the runtime improves by amount a. When I disable subroutine B the runtime improves by amount b. When I disable subroutines A and B the runtime improves by amount c > a + b. Both subroutines are completely independent of each other.
This next part may be a naive approach to analyze this, but here is what I did: I compiled each version of the code and ran cuobjdump --dump-sass for each binary. The resulting output was about 1350 lines for the complete binary and around 1100 lines for each binary with one subroutine disabled. If I disabled both subroutines I got 850 lines. It appears that I need 3.1 us per line for the first three and 2.4 us for both subroutines disabled.
Since A and B do not contain anything complicated or use memory more intensively than the rest of the code I do not think that this is caused by commenting out all the time intensive operations and leaving the simple ones active. My guess is that the program code with both A and B disabled still fits in the streaming multiprocessors' instruction caches while the other versions are too large. That might result in global memory accesses so that more program code can be loaded and the latency causes this discrepancy. Unfortunately I could not find any information on the instruction cache size.
Can anyone help me with the interpretation of these results?

There can be several things leading to your results. Some may include:
Lack of A & B leads to some additional optimization of the remaining code. The resulting code might not be shorter but may still execute faster.
The optimizer found that certain code in A exists also in B (A and B may still be independent). As a result having both A & B slows the program less than having separate A and separate B.
Having just A or B (or neither) results in better memory access patterns or more cache hits when reading the data in the remaining code. (I am talking about data, not program code)
Lack of A and B may lead to better occupancy, allowing more threads to be run concurrently.
I highly doubt that the instruction cache might be the issue. Even if global memory reads were necessary, it is never "misaligned" as warp always needs only one instruction. My guess is that they use some dedicated piece of hardware for instructions.

Related

Dynamically executing large volumes of execute-once, straight-line x86 code

Dynamically generating code is pretty well-known technique, for example to speed up interpreted languages, domain-specific languages and so on. Whether you want to work low-level (close to 1:1 with assembly), or high-level you can find libraries you help you out.
Note the distinction between self-modifying code and dynamically-generated code. The former means that some code that has executed will be modified in part and then executed again. The latter means that some code, that doesn't exist statically in the process binary on disk, is written to memory and then executed (but will not necessarily ever be modified). The distinction might be important below or simply because people treat self-modifying code as a smell, but dynamically generated code as a great performance trick.
The usual use-case is that the generated code will be executed many times. This means the focus is usually on the efficiency of the generated code, and to a lesser extent the compilation time, and least of all the mechanics of actually writing the code, making it executable and starting execution.
Imagine however, that your use case was generating code that will execute exactly once and that this is straight-line code without loops. The "compilation" process that generates the code is very fast (close to memcpy speed). In this case, the actual mechanics of writing to the code to memory and executing it once become important for performance.
For example, the total amount of code executed may be 10s of GBs or more. Clearly you don't want to just write all out to a giant buffer without any re-use: this would imply writing 10GB to memory and perhaps also reading 10GB (depending on how generation and execution was interleaved). Instead you'd probably want to use some reasonably sized buffer (say to fit in the L1 or L2 cache): write out a buffer's worth of code, execute it, then overwrite the buffer with the next chunk of code and so on.
The problem is that this seems to raise the spectre of self-modifying code. Although the "overwrite" is complete, you are still overwriting memory that was at one point already executed as instructions. The newly written code has to somehow make its way from the L1D to the L1I, and the associated performance hit is not clear. In particular, there have been reports that simply writing to the code area that has already been executed may suffer penalties of 100s of cycles and that the number of writes may be important.
What's the best way of generating a large about of dynamically generated straight-line code on x86 and executing it?
I think you're worried unnecessarily. Your case is more like when a process exits and its pages are reused for another process (with different code loaded into them), which shouldn't cause self-modifying code penalties. It's not the same as when a process writes into its own code pages.
The self-modifying code penalties are significant when the overwritten instructions have been prefetched or decoded to the trace cache. I think it is highly unlikely that any of the generated code will still be in the prefetch queue or trace cache by the time the code generator starts overwriting it with the next bit (unless the code generator is trivial).
Here's my suggestion: Allocate pages up to some fraction of L2 (as suggested by Peter), fill them with code, and execute them. Then map the same pages at the next higher virtual address and fill them with the next part of the code. You'll get the benefit of cache hits for the reads and the writes but I don't think you'll get any self-modifying code penalty. You'll use 10s of GB of virtual address space, but keep using the same physical pages.
Use a serializing operation such as CPUID before each time you start executing the modified instructions, as described in sections 8.1.3 and 11.6 of the Intel SDM.
I'm not sure you'll stand to gain much performance by using a gigantic amount of straight-line code instead of much smaller code with loops, since there's significant overhead in continually thrashing the instruction cache for so long, and the overhead of conditional jumps has gotten much better over the past several years. I was dubious when Intel made claims along those lines, and some of their statements were rather hyperbolic, but it has improved a lot in common cases. You can still always avoid call instructions if you need to for simplicity, even for tree recursive functions, by effectively simulating "the stack" with "a stack" (possibly itself on "the stack"), in the worst case.
That leaves two reasons I can think of that you'd want to stick with straight-line code that's only executed once on a modern computer: 1) it's too complicated to figure out how to express what needs to be computed with less code using jumps, or 2) it's an extremely heterogeneous problem being solved that actually needs so much code. #2 is quite uncommon in practice, though possible in a computer theoretical sense; I've just never encountered such a problem. If it's #1 and the issue is just how to efficiently encode the jumps as either short or near jumps, there are ways. (I've also just recently gotten back into x86-64 machine code generation in a side project, after years of not touching my assembler/linker, but it's not ready for use yet.)
Anyway, it's a bit hard to know what the stumbling block is, but I suspect that you'll get much better performance if you can figure out a way to avoid generating gigabytes of code, even if it may seem suboptimal on paper. Either way, it's usually best to try several options and see what works best experimentally if it's unclear. I've sometimes found surprising results that way. Best of luck!

Drhystone benchmark on 32 bit micrcontroller

Currently I am doing a performance comparison on two 32bit microcontrollers. I used Dhrystone benchmark to run on both microcontrollers. One microcontroller has 4KB I-cache while second coontroller has 8KB of I-cache. Both microcontrollers are using same tool chain. As much as possible I kept same static and run-time settings onboth microcontrollers. But microcontroller with 4KB cache are faster than 8KB cache microcontroller. Both microcontroller are from same vendor and based on same CPU.
Could anyone provide some information why microcontroller with 4KB cache is faster than other?
Benchmarks are in general useless. dhrystone being one of the oldest ones and it may have had a little bit of value back then, before pipelines and too much compiler optimization. I think I gave up on dhrystone 15 years ago roughly about the time I started using it.
It is trivial to demonstrate that this code
.globl ASMDELAY
ASMDELAY:
sub r0,r0,#1
bne ASMDELAY
bx lr
Which is primarily two INSTRUCTIONS, can vary WIDELY in execution time on the same chip if you understand how modern processors work. The simple trick to seeing this is turn off the caches and prefetchers and such, and place this code at offset 0x0000, call it with some value. place it at 0x0004 repeat, then at 0x0008, repeat. keep doing this. You can put one, two, etc nops between the subtract and branch. try it at various offsets.
THEN, turn on and of caches for each of these alignments, if you have a prefetch outside the processor for the flash, turn that on and off.
THEN, vary your clocks, esp for those MCUs where you have to adjust the wait states based on clock rate.
On a SINGLE MCU, you are going to see those essentially two instructions vary in execution time by a very large amount. Lets just say 20 times longer in some cases than others.
Now take a small program or a small fraction of the dhrystone program. Compile that for your mcu, how many instructions do you see? Make minor to major optimization and other variations on the compile command line, how much does the code change. If two instructions can vary by lets call it 20 times in execution time, how bad can it get for 200 instructions or 2000 instructions? It can get pretty bad.
If you take the dhrystone programs you have right now with the compiler options you have right now, go into your bootstrap, add one nop (causing the entire binary to shift by one instruction in flash) run again. Add two, three, four. You are still not comparing different mcus just running your benchmark on one system.
Run with and without d cache, with and without i cache if you have each of those. turn on and off the flash prefetch if you have one and if you have a write buffer you can turn on and of try that. Still remaining on the same compiler same options same mcu.
Take the different functions in the dhrystone source code, and re-arrange them in the source code. instead of Proc_1, Proc_2, Proc_3 make it Proc_1, Proc_3, Proc_2. Do all of the above again. Re-arrange again, repeat.
Before leaving this mcu you should now see that the execution time of the same source code which is completely unmodified (other than perhaps re-arranging functions) can and will have vastly different execution times.
if you then start changing compiler options or keep the same source and change compilers, you will see even more of a difference in execution time.
How is it possible that dhrystone benchmarks today or from way back had a single result for each platform? Simple, that result was just one of a wide range, not really representing the platform.
So then if you try to compare two different hardware platforms be it the same arm core inside a different mcu from the same vendor or different vendors. the arm core, assuming (which is not a safe assumption) it is the same source with the same compile/build options, even assuming the same verilog compile and synthesis was used. you can have that same core change based on arm provided options. Anyway, how the vendor be it the same vendor in two instances or two different vendors wraps that same core, you will see variations. Then take a completely different core be it another arm or a mips, etc. How could you gain any value comparing those using a program like this that itself varies widely on each platform?
You cant. What you can do is use benchmarks to give the illusion of one thing being better than another, one computer is faster than another, one compiler is faster than another. In order to sell computers or compilers. Sprints coverage is within one percent of Verizons...does that tell us anything useful? Nope.
If you were to eliminate the compiler from the equation, and if these are truly the same "CPU", same rev of source from ARM, built the same way, then they should fetch the same, but the size of the cache is part of that so it may already be a different cpu implementation as the width or depth of the cache can affect things. In software it is like needing a 32 bit pointer rather than a 16 bit pointer (17 bit instead of 16, but you cant have a 17 bit generally in logic you can).
Anyway, if you compile the code under test one time for an address space that is common to both platforms, use that same binary exactly for that space, can attach different bootstrap code as needed, note the C library calls strcpy, etc also have to be the same in the same space between platforms to eliminate the compiler and alignment from messing you up. this may or may not level the playing field.
If you want to believe these are the same cpu, then turn the caches off, eliminate the compiler variations by doing the above. See if they execute the same. Copy the program to ram and run in ram, eliminate the flash issues. I assume you have them both clocked the same with the same wait states in the flash?
if they are the same cpu and the chip vendor has with these two chips made the memory system take the same number of clocks say for ram accesses, and it is really the same cpu, you should be able to get the same time by eliminating optimizations (caching, flash prefetching, alignment).
what you are probably seeing is some form of alignment with how the code lies in memory from the compiler vs cache lines, or it could be much much simpler it could be just the differences in the caches, how the hits and misses work and the 4KB is just more lucky than the 8KB for this particular program compiled a certain way, aligned in memory a certain way, etc.
With the simple two instruction loop above it is easy to see some of the reasons why performance varies on the same system, if your modern cpu fetches 8 instructions at a time and your loop gets too close to the tail end of that fetch, the prefetch may think it needs to fetch another 8 beyond that costing you those clock cycles. Certainly as you exactly straddle two "fetch lines" as I call them with those two instructions it is going to cost you a lot more cycles per loop even with a cache. Same problem happens when those two instructions approach a cache line (as you vary their alignment per test) eventually it takes two cache line reads instead of one to fetch those two instructions. At least the first time through there is an extra cache line read. The extra clocks for that first time through is something you can see using a simple benchmark like that while playing with alignment.
Michael Abrash, Zen of Assembly Language. There is an epub/etc you can build from github of this book. the 8088 was obsolete when this book came out if that is all you see 8088 stuff, then you are completely missing the point. It applies to the most modern processors today, how to view the problem, how to test, how to time the test, how to interpret the results. All the stuff I have mentioned so far, and all the things I know about this that I have not mentioned, all came from that books knowledge applied for however many decades I have been doing this.
So again if you have truly eliminated the compiler, alignment, the cpu, the memory system tied to that cpu, etc and it is down to only the size of the cache varies. Then it is probably related to how the cache line fetches hit and miss differently based on alignment of the code relative to the cache lines for the two caches. One is hitting more and missing less and/or evicting better for this particular binary. You can try rearranging the functions, can add nops, or if you cant get at the bootstrap then add whole functions or more code (another printf, etc) at a lower address in the binary causing the linker to slide the code under test through to different addresses changing how the program lines up with the cache lines. Since the functions in the code under test are so big (more than a few instructions to implement) you would have to start modifying the program in order to get a finer grained adjustment of binary relative to cache lines.
You most definitely should see execution time differences on the two platforms if you adjust alignment and or wholesale re-arrange the binary based on functions being re-arranged.
Bottom line benchmarks dont really tell you much, the results have more of a negative stink to them than a positive joy. Without a re-write a particular benchmark or application may just do better on one platform (be it everything is the same but the size of the cache or two completely different architectures) even when you try to vary the results due to alignment, turning on and off prefetching, write buffering, branch prediction, etc. Pulling out all the tricks you can come up with one may vary from x to y the other from n to m and maybe there is some overlap in the ranges. For these two platforms that you say are similiar except for cache size I would hope you could find a combination where sometimes A is faster than B and at least one combination where B is faster than A, with the same features turned on and off for both in that comparision. If/when you switch to some other application/benchmark this all starts over again, no reason to assume dhrystone results predict any other code under test. The only program that matters, esp on an MCU is the final build of your application. Just remember changing a single line of code, or even adding a single nop in the bootstrap can have dramatic results in performance sometimes several TIMES slower or faster from a single nop.

How to compare two implementations of the same algorithm? (by examine their Assembly code)

Assume I have two implementations of the same algorithm in assembly. I would like to know by examining the two snippets codes which one is faster.
The parameters I thought one might take into account are: number of op-codes, number of branches, number of function frames.
My questions are:
Can I assume each opcode execution is one cycle ?
What is the overhead of branch which break the pipeline ?
What are the effects and overhead of calling a function ?
Is there a difference in the analysis between ARM and x86 ?
The question is theoretical since I have two implementations; one 130 instructions long and one is 184 instructions long.
And I would like to know if it is definitely true to say the 130 instructions long snippet is faster than the 184 instructions long implementation?
"BETTER == FASTER"
Without wanting to be flippant, the answers are
no
that depends on your hardware
that depends on your hardware
yes
You would really need to test things on your target hardware, or have a simulator that understands your hardware fully, in order to answer your question the way you meant to...
For the last part of your question, you need to define "better"…better.
Since you asked about a Cortex A9, the data sheet has instruction cycle counts in appendix B. These counts generally assume that the memory bus is fast enough to keep the CPU busy. In reality this is rarely the case. Many video/audio algorithms will have a big win in how they access memory.
One cycle per op
Of course you can't assume this if you want an exact count. However, if you are deciding which algorithm to choose, you can get a feel for the best algorithm by looking at the instructions in the inner loop. Here, your cache should allow the code to execute as per the instruction counts in the data sheet. If the counts are close, then you probably need to look at each instruction. Load/stores are more expensive and usually multiples, etc. Some algorithms, especially crytographic, will have big wins by using assembler that doesn't map well to C. For example, clz, ror, using the carry for multi-word arithmetic, etc.
Branch overhead
Look in Appendix B, or whatever data sheet has cycle counts for your processor. For an ARM926 it is about 3 cycles. The compiler only generates two conditional opcodes in a row to avoid branching, otherwise, it branches. If the algorithm is large, the branch may disrupt the cache. A hard answer depends on your CPU, cache, and memory. According to the Cortex A9 datasheet (B.5), there is only one cycle overhead to a fixed branch.
Function overhead
This is much the same as the branch overhead. However, the compiler will also have an influence. noted by Jim Does it cache align functions. Does the compiler perform leaf function optimizations, etc. With modern gcc versions, if all the functions are static, the compiler will generally in-line when it is advantageous. If the algorithms are particularly large, a register spill may be advantageous. However, with your example of 130/184 instructions, this seems unlikely. The compiler options will obviously effect the overhead. You can use objdump -S to examine the prologue/epilogue and then determine the number of cycles for your hardware.
ARM verus x86
Of course there is a technical difference in the cycle counts. The CISC x86 also has variable instruction size. This complicates the analysis. It is slightly easier on the ARM.
Normally, you want to ball park things and then actually run them with a profiler. The estimates can help guide development of the algorithms. Loop/memory tuning, etc for your hardware. Something like instruction emulation, page or alignment faults, etc may be dominant and make all the cycle count analysis meaningless. If the algorithm is in user space, per-emption, may negate cache wins from run to run. It is possible that one algorithm will work better in a little loaded system and the other will work better under a higher load.
A note on cycle counts
See the post-process objdump for some complications in getting cycle counts. Basically a typical CPU is several phases (a pipe line) and different conditions can cause stalls. As CPU's become more complex, the pipe line typically gets longer, meaning there are more conditions or phases which can stall. However, cycle count estimates can be helpful in guiding development of an algorithm and evaluating them. Things like memory timing or branch prediction can be just as important, depending on the algorithm. Ie, cycle counts are not completely useless, but they are not complete either. Profiling should confirm actual algorithm times. If they diverge, instruction re-ordering, pre-fetching and other techniques may bring them closer. The fact that cycle counts and active profiling diverge can be helpful in itself.
It is definitely not true to say that the 130 instruction code is faster than the 184 instruction code. it is very easy to have 1000 instructions run faster than 100 and vice versa on either of these platforms.
1 Can I assume each opcode execution is one cycle ?
Start by looking at the advertised mips/mhz, although a marketing number it gives a rough idea of what is possible. If the number is greater than one then more than one instruction per clock is possible.
2 What is the overhead of branch which break the pipeline ?
Anywhere from absolutely no affect to a very dramatic affect, on either system. one clock to hundreds are the potential penalty.
3 What are the effects and overhead of calling a function ?
Depends heavily on the function, and the function calling the function. Depending on the calling convention you might have to save registers to the stack, or rearrange the contents of registers to prepare for the parameters for the function to be called. If passing a struct by value a copy of the struct may need to be made on the stack, the bigger the struct passed the bigger the copy. once in the function a stack frame may need to be prepared, etc, etc. There are many factors involved. This question and answer are also independent of platform.
4 Is there a difference in the analysis between ARM and x86 ?
yes and no, both systems use all the modern tricks of pipelining, branch prediction, etc to keep the mips/mhz up. ARM is going to give a better mips per mhz than x86, x86 being variable instruction length might give more instructions per unit cache. How you analyze the cache, and memory and peripheral systems in the systems side of the analysis is roughly the same. The comparison of the instructions and core are similar and different depending on what aspects you are analyzing. The arm is not microcoded, the x86 likely is so you dont really see how many registers there really are, things like that. at the same time the x86 you can get a better look at the memory system with the arm, since they are generally not system on a chip. Depending on what ARM chip you buy you may lose a lot of the visibility in the boundaries of the chip, might not see all the memory and peripheral busses, for example. (x86 is changing that by putting pcie on chip now for example) in the case of something in the cortex-a class you mentioned you would have similar edge of chip visibility as those would use larger/cheaper dram based memory off chip rather than microcontroller like on chip resources.
Bottom line your final question:
"And I would like to know if it is definitely true to say the 130 instructions long snippet is faster than the 184 instructions long implementation?"
It is definitely NOT TRUE to say the 130 instruction snippet is faster than the 184 instruction snippet. It might be faster it might be slower and it might be about the same. With a lot more information we might be able to make a pretty good statement or it may still be non-deterministic. it is easy to choose 100 instructions that execute faster than 1000 instructions and likewise easy to choose 1000 instructions that execute faster than 100 instructions (even if I were to add no branching and no loops, just linear execution)
Your question is almost entirely meaningless: It probably depends on your input.
Most CPUs have something resembling a branch misprediction penalty (e.g. traditional ARM which throws away an instruction fetch/decode on any taken branch, IIRC). ARM and x86 also allow conditional execution, which can be faster than branching. If either of these are dependent on input data, then different inputs will follow different code paths.
Perhaps one version heavily uses conditional execution, which is wasteful when the condition is false. Perhaps another was compiled using some profiling information that performs no branches (except the return at the end) for a specific case. There are many, many reason why a compiler can take the same source and produce an "optimized" output which is faster for one input and slower for another.
Many optimizations have this characteristic — for example, aligning the start of a loop to 16 bytes helps on some processors, but not when the loop is only executed once.
Some text book answer to this question from Cortex
™
-A Series Programmer’s Guide, chapter 17.
Although cycle timing information can be found in the Technical Reference Manual (TRM) for the processor that you are using, it is very difficult to work out how many cycles even a trivial piece of code will take to execute. The movement of instructions through the pipeline is dependent on the progress of the surrounding instructions and can be significantly affected by memory system
activity. Pending loads or instruction fetches which miss in the cache can stall code for tens of cycles. Standard data processing instructions (logical and arithmetic) will take only one or two cycles to execute, but this does not give the full picture. Instead, we must use profiling tools, or the system performance monitor built-in to the processor, to extract useful information about performance.
Also read under 17.4 Cortex-A9 micro-architecture optimizations which answers your question very very much.

How does one write code that best utilizes the CPU cache to improve performance?

This could sound like a subjective question, but what I am looking for are specific instances, which you could have encountered related to this.
How to make code, cache effective/cache friendly (more cache hits, as few cache misses as possible)? From both perspectives, data cache & program cache (instruction cache),
i.e. what things in one's code, related to data structures and code constructs, should one take care of to make it cache effective.
Are there any particular data structures one must use/avoid, or is there a particular way of accessing the members of that structure etc... to make code cache effective.
Are there any program constructs (if, for, switch, break, goto,...), code-flow (for inside an if, if inside a for, etc ...) one should follow/avoid in this matter?
I am looking forward to hearing individual experiences related to making cache efficient code in general. It can be any programming language (C, C++, Assembly, ...), any hardware target (ARM, Intel, PowerPC, ...), any OS (Windows, Linux,S ymbian, ...), etc..
The variety will help to better to understand it deeply.
The cache is there to reduce the number of times the CPU would stall waiting for a memory request to be fulfilled (avoiding the memory latency), and as a second effect, possibly to reduce the overall amount of data that needs to be transfered (preserving memory bandwidth).
Techniques for avoiding suffering from memory fetch latency is typically the first thing to consider, and sometimes helps a long way. The limited memory bandwidth is also a limiting factor, particularly for multicores and multithreaded applications where many threads wants to use the memory bus. A different set of techniques help addressing the latter issue.
Improving spatial locality means that you ensure that each cache line is used in full once it has been mapped to a cache. When we have looked at various standard benchmarks, we have seen that a surprising large fraction of those fail to use 100% of the fetched cache lines before the cache lines are evicted.
Improving cache line utilization helps in three respects:
It tends to fit more useful data in the cache, essentially increasing the effective cache size.
It tends to fit more useful data in the same cache line, increasing the likelyhood that requested data can be found in the cache.
It reduces the memory bandwidth requirements, as there will be fewer fetches.
Common techniques are:
Use smaller data types
Organize your data to avoid alignment holes (sorting your struct members by decreasing size is one way)
Beware of the standard dynamic memory allocator, which may introduce holes and spread your data around in memory as it warms up.
Make sure all adjacent data is actually used in the hot loops. Otherwise, consider breaking up data structures into hot and cold components, so that the hot loops use hot data.
avoid algorithms and datastructures that exhibit irregular access patterns, and favor linear datastructures.
We should also note that there are other ways to hide memory latency than using caches.
Modern CPU:s often have one or more hardware prefetchers. They train on the misses in a cache and try to spot regularities. For instance, after a few misses to subsequent cache lines, the hw prefetcher will start fetching cache lines into the cache, anticipating the application's needs. If you have a regular access pattern, the hardware prefetcher is usually doing a very good job. And if your program doesn't display regular access patterns, you may improve things by adding prefetch instructions yourself.
Regrouping instructions in such a way that those that always miss in the cache occur close to each other, the CPU can sometimes overlap these fetches so that the application only sustain one latency hit (Memory level parallelism).
To reduce the overall memory bus pressure, you have to start addressing what is called temporal locality. This means that you have to reuse data while it still hasn't been evicted from the cache.
Merging loops that touch the same data (loop fusion), and employing rewriting techniques known as tiling or blocking all strive to avoid those extra memory fetches.
While there are some rules of thumb for this rewrite exercise, you typically have to carefully consider loop carried data dependencies, to ensure that you don't affect the semantics of the program.
These things are what really pays off in the multicore world, where you typically wont see much of throughput improvements after adding the second thread.
I can't believe there aren't more answers to this. Anyway, one classic example is to iterate a multidimensional array "inside out":
pseudocode
for (i = 0 to size)
for (j = 0 to size)
do something with ary[j][i]
The reason this is cache inefficient is because modern CPUs will load the cache line with "near" memory addresses from main memory when you access a single memory address. We are iterating through the "j" (outer) rows in the array in the inner loop, so for each trip through the inner loop, the cache line will cause to be flushed and loaded with a line of addresses that are near to the [j][i] entry. If this is changed to the equivalent:
for (i = 0 to size)
for (j = 0 to size)
do something with ary[i][j]
It will run much faster.
The basic rules are actually fairly simple. Where it gets tricky is in how they apply to your code.
The cache works on two principles: Temporal locality and spatial locality.
The former is the idea that if you recently used a certain chunk of data, you'll probably need it again soon. The latter means that if you recently used the data at address X, you'll probably soon need address X+1.
The cache tries to accomodate this by remembering the most recently used chunks of data. It operates with cache lines, typically sized 128 byte or so, so even if you only need a single byte, the entire cache line that contains it gets pulled into the cache. So if you need the following byte afterwards, it'll already be in the cache.
And this means that you'll always want your own code to exploit these two forms of locality as much as possible. Don't jump all over memory. Do as much work as you can on one small area, and then move on to the next, and do as much work there as you can.
A simple example is the 2D array traversal that 1800's answer showed. If you traverse it a row at a time, you're reading the memory sequentially. If you do it column-wise, you'll read one entry, then jump to a completely different location (the start of the next row), read one entry, and jump again. And when you finally get back to the first row, it will no longer be in the cache.
The same applies to code. Jumps or branches mean less efficient cache usage (because you're not reading the instructions sequentially, but jumping to a different address). Of course, small if-statements probably won't change anything (you're only skipping a few bytes, so you'll still end up inside the cached region), but function calls typically imply that you're jumping to a completely different address that may not be cached. Unless it was called recently.
Instruction cache usage is usually far less of an issue though. What you usually need to worry about is the data cache.
In a struct or class, all members are laid out contiguously, which is good. In an array, all entries are laid out contiguously as well. In linked lists, each node is allocated at a completely different location, which is bad. Pointers in general tend to point to unrelated addresses, which will probably result in a cache miss if you dereference it.
And if you want to exploit multiple cores, it can get really interesting, as usually, only one CPU may have any given address in its L1 cache at a time. So if both cores constantly access the same address, it will result in constant cache misses, as they're fighting over the address.
I recommend reading the 9-part article What every programmer should know about memory by Ulrich Drepper if you're interested in how memory and software interact. It's also available as a 104-page PDF.
Sections especially relevant to this question might be Part 2 (CPU caches) and Part 5 (What programmers can do - cache optimization).
Apart from data access patterns, a major factor in cache-friendly code is data size. Less data means more of it fits into the cache.
This is mainly a factor with memory-aligned data structures. "Conventional" wisdom says data structures must be aligned at word boundaries because the CPU can only access entire words, and if a word contains more than one value, you have to do extra work (read-modify-write instead of a simple write). But caches can completely invalidate this argument.
Similarly, a Java boolean array uses an entire byte for each value in order to allow operating on individual values directly. You can reduce the data size by a factor of 8 if you use actual bits, but then access to individual values becomes much more complex, requiring bit shift and mask operations (the BitSet class does this for you). However, due to cache effects, this can still be considerably faster than using a boolean[] when the array is large. IIRC I once achieved a speedup by a factor of 2 or 3 this way.
The most effective data structure for a cache is an array. Caches work best, if your data structure is laid out sequentially as CPUs read entire cache lines (usually 32 bytes or more) at once from main memory.
Any algorithm which accesses memory in random order trashes the caches because it always needs new cache lines to accomodate the randomly accessed memory. On the other hand an algorithm, which runs sequentially through an array is best because:
It gives the CPU a chance to read-ahead, e.g. speculatively put more memory into the cache, which will be accessed later. This read-ahead gives a huge performance boost.
Running a tight loop over a large array also allows the CPU to cache the code executing in the loop and in most cases allows you to execute an algorithm entirely from cache memory without having to block for external memory access.
One example I saw used in a game engine was to move data out of objects and into their own arrays. A game object that was subject to physics might have a lot of other data attached to it as well. But during the physics update loop all the engine cared about was data about position, speed, mass, bounding box, etc. So all of that was placed into its own arrays and optimized as much as possible for SSE.
So during the physics loop the physics data was processed in array order using vector math. The game objects used their object ID as the index into the various arrays. It was not a pointer because pointers could become invalidated if the arrays had to be relocated.
In many ways this violated object-oriented design patterns but it made the code a lot faster by placing data close together that needed to be operated on in the same loops.
This example is probably out of date because I expect most modern games use a prebuilt physics engine like Havok.
A remark to the "classic example" by user 1800 INFORMATION (too long for a comment)
I wanted to check the time differences for two iteration orders ( "outter" and "inner"), so I made a simple experiment with a large 2D array:
measure::start();
for ( int y = 0; y < N; ++y )
for ( int x = 0; x < N; ++x )
sum += A[ x + y*N ];
measure::stop();
and the second case with the for loops swapped.
The slower version ("x first") was 0.88sec and the faster one, was 0.06sec. That's the power of caching :)
I used gcc -O2 and still the loops were not optimized out. The comment by Ricardo that "most of the modern compilers can figure this out by itselves" does not hold
Only one post touched on it, but a big issue comes up when sharing data between processes. You want to avoid having multiple processes attempting to modify the same cache line simultaneously. Something to look out for here is "false" sharing, where two adjacent data structures share a cache line and modifications to one invalidates the cache line for the other. This can cause cache lines to unnecessarily move back and forth between processor caches sharing the data on a multiprocessor system. A way to avoid it is to align and pad data structures to put them on different lines.
I can answer (2) by saying that in the C++ world, linked lists can easily kill the CPU cache. Arrays are a better solution where possible. No experience on whether the same applies to other languages, but it's easy to imagine the same issues would arise.
Cache is arranged in "cache lines" and (real) memory is read from and written to in chunks of this size.
Data structures that are contained within a single cache-line are therefore more efficient.
Similarly, algorithms which access contiguous memory blocks will be more efficient than algorithms which jump through memory in a random order.
Unfortunately the cache line size varies dramatically between processors, so there's no way to guarantee that a data structure that's optimal on one processor will be efficient on any other.
To ask how to make a code, cache effective-cache friendly and most of the other questions , is usually to ask how to Optimize a program, that's because the cache has such a huge impact on performances that any optimized program is one that is cache effective-cache friendly.
I suggest reading about Optimization, there are some good answers on this site.
In terms of books, I recommend on Computer Systems: A Programmer's Perspective which has some fine text about the proper usage of the cache.
(b.t.w - as bad as a cache-miss can be, there is worse - if a program is paging from the hard-drive...)
There has been a lot of answers on general advices like data structure selection, access pattern, etc. Here I would like to add another code design pattern called software pipeline that makes use of active cache management.
The idea is borrow from other pipelining techniques, e.g. CPU instruction pipelining.
This type of pattern best applies to procedures that
could be broken down to reasonable multiple sub-steps, S[1], S[2], S[3], ... whose execution time is roughly comparable with RAM access time (~60-70ns).
takes a batch of input and do aforementioned multiple steps on them to get result.
Let's take a simple case where there is only one sub-procedure.
Normally the code would like:
def proc(input):
return sub-step(input))
To have better performance, you might want to pass multiple inputs to the function in a batch so you amortize function call overhead and also increases code cache locality.
def batch_proc(inputs):
results = []
for i in inputs:
// avoids code cache miss, but still suffer data(inputs) miss
results.append(sub-step(i))
return res
However, as said earlier, if the execution of the step is roughly the same as RAM access time you can further improve the code to something like this:
def batch_pipelined_proc(inputs):
for i in range(0, len(inputs)-1):
prefetch(inputs[i+1])
# work on current item while [i+1] is flying back from RAM
results.append(sub-step(inputs[i-1]))
results.append(sub-step(inputs[-1]))
The execution flow would look like:
prefetch(1) ask CPU to prefetch input[1] into cache, where prefetch instruction takes P cycles itself and return, and in the background input[1] would arrive in cache after R cycles.
works_on(0) cold miss on 0 and works on it, which takes M
prefetch(2) issue another fetch
works_on(1) if P + R <= M, then inputs[1] should be in the cache already before this step, thus avoid a data cache miss
works_on(2) ...
There could be more steps involved, then you can design a multi-stage pipeline as long as the timing of the steps and memory access latency matches, you would suffer little code/data cache miss. However, this process needs to be tuned with many experiments to find out right grouping of steps and prefetch time. Due to its required effort, it sees more adoption in high performance data/packet stream processing. A good production code example could be found in DPDK QoS Enqueue pipeline design:
http://dpdk.org/doc/guides/prog_guide/qos_framework.html Chapter 21.2.4.3. Enqueue Pipeline.
More information could be found:
https://software.intel.com/en-us/articles/memory-management-for-optimal-performance-on-intel-xeon-phi-coprocessor-alignment-and
http://infolab.stanford.edu/~ullman/dragon/w06/lectures/cs243-lec13-wei.pdf
Besides aligning your structure and fields, if your structure if heap allocated you may want to use allocators that support aligned allocations; like _aligned_malloc(sizeof(DATA), SYSTEM_CACHE_LINE_SIZE); otherwise you may have random false sharing; remember that in Windows, the default heap has a 16 bytes alignment.
Write your program to take a minimal size. That is why it is not always a good idea to use -O3 optimisations for GCC. It takes up a larger size. Often, -Os is just as good as -O2. It all depends on the processor used though. YMMV.
Work with small chunks of data at a time. That is why a less efficient sorting algorithms can run faster than quicksort if the data set is large. Find ways to break up your larger data sets into smaller ones. Others have suggested this.
In order to help you better exploit instruction temporal/spatial locality, you may want to study how your code gets converted in to assembly. For example:
for(i = 0; i < MAX; ++i)
for(i = MAX; i > 0; --i)
The two loops produce different codes even though they are merely parsing through an array. In any case, your question is very architecture specific. So, your only way to tightly control cache use is by understanding how the hardware works and optimising your code for it.

Does JIT performance suffer due to writeable pages?

In the book Linkers and Loaders, it's mentioned that one of the reasons for executables to have a separate code section is that the code section can be kept in read only pages, which results in a performance increase. Is this still true for a modern OS? Seeing as Just in Time compilers are generating code on the fly, I assume they need writable pages. Does this mean that JIT generated code will always suffer a performance hit in comparison? If so, how significant a hit is it?
Effects of memory management aside (which are explained in other answers), CPU doesn't need to continuously check if the current stream of instructions are modified and the intermediate results in the pipeline should be thrown away and new code needs to be read. In the case of jit compilation, this scenario may occur often depending on the design of the compiler, depth of CPU pipeline, size of code cache on CPU and number of other CPUs which may modify that code. It isn't normally allowed to occur in well designed modern systems where code is generated to a writeable page and marked as executable and readonly afterwards. This is not unique to jit of course. It can happen in all kinds of self modifying code.
Yes, it should suffer some sort of hit because the code in memory is not backed directly by the executable, so it would have to be paged out instead of just dropped. Having said that, various forms of linking can also dirty up regular code pages so that they no longer match the disk image, with the same consequences, so I'm not sure that this is a big deal.
The performance increase is not because of whether the pages are read only or not. The advantage is that read only pages can be shared between processes, so you use less memory which means less swapping (both to L1/L2/L3 caches as well as to disk in extreme cases).
JIT tries to mitigate this by not needlessly JITting, but only JITting the hot functions. This will result in only a modest increase in memory since the number of hot functions are relatively small.
A JIT compiler could also be smart and cache the result of the JITting so it could (theoretically) be shared. But I don't know whether this is done in practice.

Resources