go-perfbook/performance.md

Writing and Optimizing Go code

This document outlines best practices for writing high-performance Go code.

At the moment, it's a collection of links to videos, slides, and blog posts ("awesome-golang-performance"), but I would like this to evolve into a longer book format where the content is here instead of external. The links should be sorted into categories.

While some discussions will be made for individual services faster (caching, etc), designing performant distributed systems is beyond the scope of this work.

All the content will be licensed under CC-BY-SA.

This book is split into different sections:

1. basic tips for writing not-slow software * CS 101-level stuff
2. tips for writing fast software * Go-specific sections on how to get the best from Go
3. advanced tips for writing really fast software * For when your optimized code isn't fast enough

When and Where to Optimize

I'm putting this first because it's really the most important step. Should you even be doing this at all?

Every optimization has a cost. Generally this cost is expressed in terms of code complexity or cognitive load -- optimized code is rarely simpler than the unoptimized version.

But there's another side that I'll call the economics of optimization. As a programmer, your time is valuable. There's the opportunity cost of what else you could be working on for your project, which bugs to fix, which features to add. Optimizing things is fun, but it's not always the right task to choose. Performance is a feature, but so is shipping, and so is correctness.

Choosing the most important thing to work on. Sometimes this isn't an optimization at all. Sometimes it's not an actual CPU optimization, but a user-experience one. Making something start up faster by doing computation in the background after drawing the main window, for example.

Some times this will be obvious: an hourly report that completes in three hours is probably less useful that one that completes in less than one.

Just because something is easy to optimize doesn't mean it's worth optimizing. Ignoring low-hanging fruit is a valid development strategy.

Think of this as optimizing your time.

Choosing what to optimize. Choosing when to optimize.

Clarify "Premature optimization" quote. 97% of the time. But work on those important 3%.

TPOP: Should you optimize? "Yes, but only if the problem is important, the program is genuinely too slow, and there is some expectation that it can be made faster while maintaining correctness, robustness, and clarity."

Fast software or fast deployment.

http://bitfunnel.org/strangeloop . has numbers. Hypothetical search engine needing 30k machines @ $1k USD / year. Doubling the speed of your software can save $15M/year. Even a developer spending an entire year to shave off 1% will pay for itself

In the vast majority of cases, the size and speed of a program is not a concern. Easiest optimization is not having to do it. The second easiest optimization is just buying faster hardware.

Once you've decided you're going to change your program, keep reading.

How to Optimize

Optimization Workflow

Before we get into the specifics, lets talk about the general process of optimization.

Optimization is a form of refactoring. But each step, rather than improving some aspect of the source code (code duplication, clarity, etc), improves some aspect of the performance: lower CPU, memory usage, latency, etc. This means that in addition to a comprehensive set of unit tests (to ensure your changes haven't broken anything), you also need a good set of benchmarks to ensure your changes are having the desired effect on performance. You must be able to verify that your change really is lowering CPU. Sometimes a change you thought would improve will actually turn out to have a zero or negative change. Always make sure you undo your fix in these cases.

The benchmarks you are using must be correct and provide reproducible numbers on representative workloads. If individual runs have too high a variance, it will make small improvements more difficult to spot. You will need to use benchstat or equivalent statistical tests and won't be able just eyeball it. (Note that using statistical tests is a good idea anyways.) The steps to run the benchmarks should be documented, and any custom scripts and tooling should be committed to the repository with instructions for how to run them. Be mindful of large benchmark suites that take a long time to run: it will make the development iterations slower.

(Note also that anything that can be measured can be optimized. Make sure you're measuring the right thing.)

The next step is to decide what you are optimizing for. If the goal is to improve CPU, what is an acceptable speed. Do you want to improve the current performance by 2x? 10x? Can you state it as "problem of size N in less than time T"? Are you trying to reduce memory usage? By how much? How much slower is acceptable for what change in memory usage? What are you willing to give up in exchange for lower space?

Optimizing for service latency is a trickier proposition. Entire books have been written on how to performance test web servers. The primary issue is that for single-threaded code, the performance is fairly consistent for a given problem size. For webservices, you don't have a single number. A proper web-service benchmark suite will provide a latency distribution for a given reqs/second level. ...

The performance goals must be specific. You will (almost) always be able to make something faster. Optimizing is frequently a game of diminishing returns. You need to know when to stop.

Dan Luu's talk also points out the advantage of rough calculations to determine if your target performance figures are reasonable.

For greenfield development, you shouldn't leave all benchmarking and performance numbers until the end. It's easy to say "we'll fix it later", but if performance is really important it will be a design consideration from the start. Any significant architectural changes required to fix performance issues will be too risky near the deadline. Note that during development, the focus should be on reasonable program design, algorithms, and data structures. Optimizing at lower-levels of the stack should wait until later in the development cycle when a more complete view of the system performance is available. Any full-system profiles you do while the system is incomplete will give a skewed view of where the bottlenecks will be in the finished system.

The difference between what your target is and the current performance will also give you an idea of where to start. If you need only a 10%-20% performance improvement, you can probably get that with some implementation tweaks and smaller fixes. If you need a factor of 10x or more, then just replacing a multiplication with a left-shift isn't going to cut it. That's probably going to call for changes up and down your stack.

Good performance work requires knowledge at many different levels, from system design, networking, hardware (CPU, caches, storage), algorithms, tuning, and debugging. With limited time and resources, consider which level will give the most improvement: it won't always be algorithm or program tuning.

In general, optimizations should proceed from top to bottom. Optimizations at the system level will have more impact than expression-level ones. Make sure you're solving the problem at the appropriate level.

This book is mostly going to talk about reducing CPU usage, reducing memory usage, and reducing latency. It's good to point out that you can very rarely do all three. Maybe CPU time is faster, but now your program uses more memory. Maybe you need to reduce memory space, but now the program will take longer.

Amdahl's Law tells us to focus on the bottlenecks. If you double the speed of routine that only takes 5% of the runtime, that's only a 2.5% speedup in total wall-clock. On the other hand, speeding up routine that takes 80% of the time by only 10% will improve runtime by almost 8%. Profiles will help identify where time is actually spent.

When optimizing, you want to reduce the amount of work the CPU has to do.

A profiler might show you that lots of time is spent in a particular routine. It could be this is an expensive routine, or it could be a cheap routine that is just called many many times. Rather than immediately trying to speed up that one routine, see if you can reduce the number of times it's called or eliminate it completely.

The Three Optimization Questions:

• Do we have to do this at all? The fastest code is the code that's not there.
• If yes, is this the best algorithm.
• If yes, is this the best implementation of this algorithm.

Concrete optimization tips

Jon Bentley's 1982 work "Writing Efficient Programs" approached program optimization as an engineering problem: Benchmark. Analyze. Improve. Verify. Iterate. A number of his tips are now done automatically by compilers. A programmers job is to use the transformations compilers can't do.

There's a summary of this book: http://www.crowl.org/lawrence/programming/Bentley82.html

When thinking changes you can make to your program, there are two basic options: you can either change your data or you can change your code.

Changing your data means either adding to or altering the representation of the data you're processing.

Ideas for augmenting your data structure:

• extra fields: For example, store the size of a linked lists rather than iterating when asked for it. Or storing additional pointers to frequently needed other nodes to multiple searches (for example, "backwards" links in a doubly-linked list to make removal O(1) ). These sorts of changes are useful when the data you need is cheap to store and keep up-to-date.

• extra search indexes: Most data structures are designed for a single type of query. If you need two different query types, having an additional "view" onto your data can be large improvement. For example, []struct, referenced by ID but sometimes string -> map[string]id (or *struct)

• extra information about elements: for example, a bloom filter. These need to be small and fast to not overwhelm the rest of the data structure.

• if queries are expensive, add a cache. We're all familiar with memcache, but there are in-process caches.

• over the wire, the network + cost of serialization will hurt
• in-process caches, but now you need to worry about expiration
• even a single item can help (logfile time parse example)

TODO: "cache" might not even be key-value, just a pointer to where you were working. This can be as simple as a "search finger"

These are all clear examples of "do less work" at the data structure level. They all cost space. Most of the time if you're optimizing for CPU, your program will use more memory. This is the classic space-time trade-off: https://en.wikipedia.org/wiki/Space%E2%80%93time_tradeoff

If your program uses too much memory, it's also possible to go the other way. Reduce space usage in exchange for increased computation. Rather than storing things, calculate them every time. You can also compress the data in memory and decompress it on the fly when you need it.

There's a book available on line covering techniques for reducing the space used by your programs. While it was originally written targetting embedded developers, the ideas are applicable for programs on modern hardware dealing with huge amounts of data. http://www.smallmemory.com/

Rearrange your data: Eliminate padding. Remove extra fields. Change to a slower data structure. Skip pointer-heavy tree structure and use slice and linear search instead. Custom compression format for your data: floating point (go-tsz), integers (delta, xor + huffman)

We will talk more about data layouts later.

Modern computers and the memory hierarchy make the space/time trade-off less clear. It's very easy for lookup tables to be "far away" in memory (and therefore expensive to access) making it faster to just recompute a value every time it's needed.

This also means that benchmarking will frequently show improvements that are not realized in the production system due to cache contention (e.g., lookup tables are in the processor cache during benchmarking but always flushed by "real data" when used in a real system. Google's Jump Hash paper in fact addressed this directly, comparing performance on both a contented and uncontended processor cache. See graphs 4 and 5 in the Jump Hash paper: https://arxiv.org/pdf/1406.2294.pdf )

TODO: how to simulate a contented cache, show incremental cost

Another aspect to consider is data-transfer time. Generally network and disk access is very slow, and so being able to load a compressed chunk will be much faster than the extra CPU time required to decompress the data once it has been fetched. As always, benchmark. A binary format will generally be smaller and faster to parse than a text one, but at the cost of no longer being as human readable.

If you're not changing the data, the other main option is to change the code.

The biggest improvement is likely to come from an algorithmic changes. This is the equivalent of replacing bubble sort with quicksort to from O(n^2) sort to O(n log n), or replacing a linear scan through an array that used to be small O(n) with a map lookup (O(1)).

This is how software becomes slow. Structures originally designed for one use is repurposed for something it wasn't designed for. This happens gradually.

It's important to have an intuitive grasp of the different big-O levels. Choose the right data structure for your problem. You don't have to alway shave cycles, but this just prevents dumb performance issues that might not be noticed until much later.

The basic classes of complexity are:

• O(1): a field access, array or map lookup
• O(log n): binary search
• O(n): simple loop
• O(n*m): nested loop
• O(n log n): divide-and-conquer
• combinatoric - look out!!

Link: bigocheatsheet.com

Let's say you need to search through of an unsorted set of data. "I should use a binary search" you think, knowing that a binary search O(log n) which is faster than the O(n) linear scan. However, a binary search requires that the data is sorted, which means you'll need to sort it first, which will take O(n log n) time. If you're doing lots of searches, then the upfront cost of sorting will pay off. On the other hand, if you're mostly doing lookups, maybe having an array was the wrong choice and you'd be better off paying the O(1) lookup cost for a map instead.

Choose the simplest reasonable data structure and move on. CS 101, writing "not-slow software". Don't be dumb. This should be your default development mode. If you know you need random access, don't choose a linked-list. If you know you need in-order traversal, don't use a map.

When writing a package to be used to by others, avoid the temptation to optimize up front for every single use case. This will result in unreadable code. Data structures by design are effectively single-purpose. You can neither read minds nor predict the future. If a user says "Your package is too slow for this use case", a reasonable answer might be "Then use this other package over here". A package should "do one this well".

Sometimes hybrid data structures will provide the performance improvement you need. For example, by bucketing your data you can limit your search to a single bucket. This still pays the theoretical cost of O(n), but the constant will be smaller. We'll revisit these kinds of tweaks when we get to program tuning.

Two things that people forget when discussion big-O notation

One: there's a constant factor involved. Two algorithms which have the same algorithmic complexity can have different constant factors. Imagine running a looping over a list 100 times vs just looping over it once Even though both are O(n), one has a constant factor that's 100 times higher.

These constant factors are why even though merge sort, quicksort, and heapsort all O(n log n), everybody uses quicksort because it's the fastest. It has the smallest constant factor.

The second thing is that big-O only says "as n grows to infinity". It says nothing about small n. "As the numbers get big, this is the growth factor that will dominate the run time."

There's frequently a cut-off point below which a dumber algorithm is faster. A nice example from the Go standard library's sort package. Most of the time it's using quicksort, but it has a shell-sort pass then insertion sort when the partition size drops below 12 elements.

Know how big each of your input sizes is likely to be in production.

Your benchmarks must use appropriately-sized inputs. As we've seen, different algorithms make sense at different input sizes. If your expected input range in <100, then your benchmarks should reflect that. Otherwise, choosing an algorithm which is optimal for n=10^6 might not be the fastest.

Be able to generate representative test data. Different distributions of data can provoke different behaviours in your algorithm: think of the classic "quicksort is O(n^2) when the data is sorted" example. Similarly, interpolation search is O(log log n) for uniform random data, but O(n) worst case. Knowing what your inputs look like is the key to both representative benchmarks and for choosing the best algorithm. If the data you're using to test isn't representative of real workloads, you can easily end up optimizing for one particular data set, "overfitting" your code to work best with one specific set of inputs.

The memory hierarchy in modern computers confuses the issue here a little bit, in that caches prefer the predictable access of scanning a slice to the effectively random access of chasing a pointer. Still, it's best to begin with a good algorithm. We will talk about this in the hardware-specific section.

"The fight may not always go to the strongest, nor the race to the fastest, but that's the way to bet." -- Rudyard Kipling.

Sometimes the best algorithm for a particular problem is not a single algorithm, but a collection of algorithms specialized for slightly different input classes. This "polyalgorithm" quickly detects what kind of input it needs to deal with and then dispatches to the appropriate code path. This is what the sorting package mentioned above does: determine the problem size and choose a different algorithm. The string and bytes packages do something similar, detecting and specializing for different cases. As with data compression, the more you know about what your input looks like, the better your custom solution can be. Even if an optimization is not always applicable, complicating your code by determining that it's safe to use and executing different logic can be worth it.

Keep comments. If something doesn't need to be done, explain why. Frequently when optimizing an algorithm you'll discover steps that don't need to be performed under some circumstances. Document them. Somebody else might think it's a bug and needs to be put back.

Empty program gives the wrong answer in no time at all. It's easy to be fast if you don't have to be correct.

Tips for implementing papers: (For algorithm read also data structure)

• Don't. Start with the obvious solution and reasonable data structures.
• "Modern" algorithms tend to have lower theoretical complexities but high constants and lots of implementation complexity.
• Look for the paper their algorithm claims to beat and implement that.
• Make sure you understand the algorithm. This sounds obvious, but it will be impossible to debug otherwise.
• The original paper for a data structure or algorithm isn't always the best. Later papers may have better explanations.
• Make sure the assumptions the algorithm makes about your data hold.
• Some papers release reference source code which you can compare against, but
  • 1. academic code is almost universally terrible
  • 2. beware licensing restrictions
  • 3. beware bugs Also look out for other implementations on GitHub: they may have the same (or different!) bugs as yours.

Cache common cases: Your cache doesn't even need to be huge. Optimized a log processing script to cache the previous time passed to time.parse() for significant speedup But beware cache invalidation, thread issues, etc Random cache eviction is fast and sufficiently effective. - only put "some" items in cache (probabilistically) to limit cache size to popular items with minimal logic Compare cost of cache logic to cost of refetching the data.

The standard library implementations need to be "fast enough" for most cases. If you have higher performance needs you will probably need specialized implementations.

This also means your benchmark data needs to be representative of the real world. If repeated requests are sufficiently rare, it's more expensive to keep them around than to recompute them. If your benchmark data consists of only the same repeated request, your cache will give an inaccurate view of the performance.

Program tuning used to be an art form, but then compilers got better. So now it turns out that compilers can optimize straight-forward code better than complicated code. The Go compiler still has a long way to go to match gcc and clang, but it does mean that you need to be careful when tuning and especially when upgrading that your code doesn't become "worse". There are definitely cases where tweaks to work around the lack of a particular compiler optimization became slower once the compiler was improved.

If you are working around a specific runtime or compiler code generation issue, always document your change with a link to the upstream issue. This will allow you to quickly revisit your optimization once the bug is fixed.

Fight the temptation to cargo cult folklore-based "performance tips".

Iterative program improvements:

• ensure progress at each step
• but frequently one improvement will enable others
• which means you need to keep looking at the entire picture

Once you've settled on the right algorithm, program tuning is the process of improving the implementation of that algorithm. In Big-O notation, this is the process of reducing the constants associated with your program.

Find cheaper ways of doing the same thing. Replace SHA1 or hash/fnv1 with a faster hash function.

program tuning:

if possible, keep the old implementation around for testing if not possible, generate sufficient golden test cases to compare output exploit a mathematical identity: https://go-review.googlesource.com/c/go/+/85477 just clearing the parts you used, rather than an entire array best done in tiny steps, a few statements at a time moving from floating point math to integer math or mandelbrot removing sqrt, or lttb removing abs cheap checks before more expensive checks: e.g., strcmp before regexp, (q.v., bloom filter before query)

Program tuning improvements are cumulative. 5x 3% improvements is a 15% improvement. Making optimizations it's worth it to think about the expected performance improvement. Replacing a hash function with a faster one is a constant factor improvement.

Log parsing example:

• time parsing is slow
• adding a single item cache is good
• removing time parsing and doing some integer math by hand is against faster
• general algorithm is slow, you can be faster because you know more about your problem
• but the code is more closely tied to exactly what you need; harder to change

Optimization is specialization, and specialized code is more fragile to change than general purpose code.

Profile regularly to ensure the track the performance characteristics of your system and be prepared to re-optimize as your traffic changes. Know the limits of your system and have good metrics that allow you to predict when you will hit those limits.

When the usage of your application changes, different pieces may become hotspots. Revisit previous optimizations and decide if they're still worth it, and revert to more readable code when possible. I had one system that I had optimized process startup time with a complex set of mmap, reflect, and unsafe. Once we changed how the system was deployed, this code was no longer required and I replaced it with much more readable regular file operations.

Optimization workflow summary

• All optimizations should follow these steps:

  1. determine your performance goals and confirm you are not meeting them
  2. profile to identify the areas to improve. This can be CPU, heap allocations, or goroutine blocking.
  3. benchmark to determine the speed up your solution will provide using the built-in benchmarking framework (http://golang.org/pkg/testing/) Make sure you're benchmarking the right thing on your target operating system and architecture.
  4. profile again afterwards to verify the issue is gone
  5. use https://godoc.org/golang.org/x/perf/benchstat or https://github.com/codahale/tinystat to verify that a set of timings are 'sufficiently' different for an optimization to be worth the added code complexity.
  6. use https://github.com/tsenart/vegeta for load testing http services
  7. make sure your latency numbers make sense: https://youtu.be/lJ8ydIuPFeU

The first step is important. It tells you when and where to start optimizing. More importantly, it also tells you when to stop. Pretty much all optimizations add code complexity in exchange for speed. And you can always make code faster. It's a balancing act.

Tooling

Introductory Profiling

Techniques applicable to source code in general

1. introduction to pprof

• go tool pprof (and https://github.com/google/pprof)

1. Writing and running (micro)benchmarks

• profile, extract hot code to benchmark, optimize benchmark, profile.
• -cpuprofile / -memprofile / -benchmem
• 0.5 ns/op means it was optimized away -> how to avoid
• tips for writing good microbenchmarks (remove unnecessary work, but add baselines)

1. How to read it pprof output
2. What are the different pieces of the runtime that show up
3. Macro-benchmarks (Profiling in production)

• net/http/pprof

Tracer

Advanced Techniques

• Techniques specific to the architecture running the code

• introduction to CPU caches

  • performance cliffs
  • building intuition around cache-lines: sizes, padding, alignment
  • false-sharing
  • true sharing -> sharding
  • OS tools to view cache-misses
  • maps vs. slices
  • SOA vs AOS layouts
  • reducing pointer chasing

• branch prediction

• function call overhead

• Comment about Jeff Dean's 2002 numbers (plus updates)

  • cpus have gotten faster, but memory hasn't kept up

Garbage Collection

• Stack vs. heap allocations
• What causes heap allocations?
• Understanding escape analysis (and the current limitation)
• API design to limit allocations: allow passing in buffers so caller can reuse rather than forcing an allocation
  • you can even modify a slice in place carefully while you scan over it
• reducing pointers to reduce gc scan times
• GOGC
• buffer reuse (sync.Pool vs or custom via go-slab, etc)

Runtime and compiler

• cost of calls via interfaces (indirect calls on the CPU level)
• runtime.convT2E / runtime.convT2I
• type assertions vs. type switches
• defer
• special-case map implementations for ints, strings
• bounds check elimination
• []byte <-> string copies, map optimizations

Common gotchas with the standard library

• time.After() leaks until it fires
• Reusing HTTP connections...
• ....
• rand.Int() and friends are 1) mutex protected and 2) expensive to create
  • consider alternate random number generation (go-pcgr, xorshift)

Unsafe

• And all the dangers that go with it
• Common uses for unsafe
• mmap'ing data files
  • struct padding
• speedy de-serialization
• string <-> slice conversion, []byte <-> []uint32, ...

cgo

• Performance characteristics of cgo calls
• Tricks to reduce the costs: batching
• Rules on passing pointers between Go and C
• syso files

Assembly

• Stuff about writing assembly code for Go
• replace as little as possible to make an impact
• very important to benchmark: improvements can be huge (10x for go-highway) zero (go-speck), or even slower (no inlining)
• always have pure-Go version (noasm build tag): testing,
• brief intro to syntax
• calling convention
• using opcodes unsupported by the asm
• notes about why intrinsics are hard
• all the tooling to make this easier: asmfmt, peachpy, c2goasm, ...

Alternate implementations

• Popular replacements for standard library packages:
  • encoding/json -> ffjson
  • net/http -> fasthttp (but incompatible API)
  • regexp -> ragel (or other regular expression package)
  • serialization
    • encoding/gob -> https://github.com/alecthomas/go_serialization_benchmarks
    • protobuf -> https://github.com/gogo/protobuf
    • all formats have trade-offs: choose one that matches what you need encoded space, decoding speed, language/tooling compatibility, ...
  • database/sql -> jackx/pgx, ...
  • gccgo

Tooling

Look at some more interesting/advanced tooling

• perf (perf2pprof)