Monthly Archives: June 2014

Paranormal trivia

Like all children who grew up in the 80’s, I was, and still am a huge fan of Ghostbusters.

While recently re watching Ivan Reitman’s homage to New York, I spotted something which has gone unnoticed on the IMDB trivia page.

Shortly after being evicted from the University, Ray Stantz (Aykroyd) and Peter Venkman (Murray) are commiserating their fall from grace, when Stantz opines:

aykroyd

In which movie did Dan Aykroyd star a year before Ghostbusters ? Continue reading

Ice cream makers and data races

This is a post about data races. The code for this post lives on Github, github.com/davecheney/benandjerry.

The example program simulates two Ice cream makers, Ben and Jerry, who greet their customers randomly.

package main

import "fmt"

type IceCreamMaker interface {
        // Hello greets a customer
        Hello()
}

type Ben struct {
        name string
}

func (b *Ben) Hello() {
        fmt.Printf("Ben says, \"Hello my name is %s\"\n", b.name)
}

type Jerry struct {
        name string
}

func (j *Jerry) Hello() {
        fmt.Printf("Jerry says, \"Hello my name is %s\"\n", j.name)
}

func main() {
        var ben = &Ben{"Ben"}
        var jerry = &Jerry{"Jerry"}
        var maker IceCreamMaker = ben

        var loop0, loop1 func()

        loop0 = func() {
                maker = ben
                go loop1()
        }

        loop1 = func() {
                maker = jerry
                go loop0()
        }

        go loop0()

        for {
                maker.Hello()
        }
}

It’s a data race, silly

Most programmers should easily spot the data race in this program.

The loop functions are changing the value of maker without using a lock, so it is undefined which implementation of Hello will be called when maker.Hello() is executed by the for loop in the main function.

Some programmers appear to be happy with this data race; either Ben or Jerry will greet the customer, it doesn’t matter which.

Lets run this code, and see what happens.

% env GOMAXPROCS=2 go run main.go
...
Ben says, "Hello my name is Ben"
Jerry says, "Hello my name is Jerry"
Jerry says, "Hello my name is Jerry"
Ben says, "Hello my name is Jerry"
Ben says, "Hello my name is Ben"
...

What! Hold up. Ben sometimes thinks that he is Jerry. How is this possible?

Interface values

The key to understanding this race is to understand how interface values are represented in memory.

An interface is conceptually a struct with two fields.
Interface
If we were to describe an interface in Go, it would look something like this.

type interface struct {
       Type uintptr     // points to the type of the interface implementation
       Data uintptr     // holds the data for the interface's receiver
}

Type points to a structure that describes the type of the value that implements this interface. Data points to the value of the implementation itself. The contents of Data are passed as the receiver of any method called via the interface.

For the statement var maker IceCreamMaker = ben, the compiler will generate code that does the following.
var marker IceCreamMaker = ben
The interface’s Type field is set to point to the definition of the *Ben type, and the Data field contains a copy of ben, that is, a pointer to a Ben value.

When loop1() executes the statement, maker = jerry, both fields of the interface value must be updated.
marker = jerry

Type now points to the definition of a *Jerry and Data contains a pointer to an instance of Jerry.

The Go memory model says that writes to a single machine word will be atomic, but interfaces are two word values. It is possible that another goroutine may observe the contents of the interface value while it is being changed. In this case it may see something like this
Data race in progress
And so Jerry‘s Hello() function is called with ben as the receiver.

Conclusion

There is no such thing as a safe data race. Your program either has no data races, or its operation is undefined.

In this example, the layout of the Ben and Jerry structs were identical in memory, so they were in some sense compatible. Imagine the chaos that would occur if they had different memory representations (this is left as an exercise to the reader).

The Go race detector will spot this error, and many others, and is as simple to use as adding the -race flag to your go test, build, or install command.

Bonus question

In the example code, the Hello method is declared on Ben or Jerry‘s pointer receiver. If this were instead declared as a method on the Ben or Jerry value, would this solve the data race ?

Futher reading

That wild ass recruiting idea

A few days ago, after reading yet another article on the critical importance of only hiring the best people — yet being unable to offer any concrete suggestions on how to do this, save slavishly repeating the “best people” homily — I posted the following

I find twitter to be a hopeless medium for communication. I am continually frustrated with its requirement to summarise a thought into its most abridged form, so please read that tweet in the light of a half baked thought forced through an uncooperative medium.

Groupthink

Despite the brevity of that tweet, it did contain several mistakes.

The first mistake, is I should have not said that this should be the only way that companies hire.

What I think many missed in that short message was my suggestion to send your staff to a conference to meet more people, which brings me to my second mistake. If I had more space in the tweet I would have said

If they are tapped out, send them to a conference, or a meetup, or to visit another company, or on training, or encourage them to give a brown bag lunch, or give them the time to contribute to an open source project.

The key in that message is your staff are your most important resource. If you want to lean on them to provide you with the best people, then you should invest in them and their careers.


Discuss this on Hacker News

Five things that make Go fast

Anthony Starks has remixed my original Google Present based slides using his fantastic Deck presentation tool. You can check out his remix over on his blog, mindchunk.blogspot.com.au/2014/06/remixing-with-deck.


I was recently invited to give a talk at Gocon, a fantastic Go conference held semi-annually in Tokyo, Japan. Gocon 2014 was an entirely community-run one day event combining training and an afternoon of presentations surrounding the theme of Go in production.

The following is the text of my presentation. The original text was structured to force me to speak slowly and clearly, so I have taken the liberty of editing it slightly to be more readable.

I want to thank Bill Kennedy, Minux Ma, and especially Josh Bleecher Snyder, for their assistance in preparing this talk.


Good afternoon.

My name is David.

I am delighted to be here at Gocon today. I have wanted to come to this conference for two years and I am very grateful to the organisers for extending me the opportunity to present to you today.

Gocon 2014
I want to begin my talk with a question.

Why are people choosing to use Go ?

When people talk about their decision to learn Go, or use it in their product, they have a variety of answers, but there always three that are at the top of their list

Gocon 2014
These are the top three.

The first, Concurrency.

Go’s concurrency primitives are attractive to programmers who come from single threaded scripting languages like Nodejs, Ruby, or Python, or from languages like C++ or Java with their heavyweight threading model.

Ease of deployment.

We have heard today from experienced Gophers who appreciate the simplicity of deploying Go applications.

Gocon 2014

This leaves Performance.

I believe an important reason why people choose to use Go is because it is fast.

Gocon 2014 (4)

For my talk today I want to discuss five features that contribute to Go’s performance.

I will also share with you the details of how Go implements these features.

Gocon 2014 (5)

The first feature I want to talk about is Go’s efficient treatment and storage of values.

Gocon 2014 (6)

This is an example of a value in Go. When compiled, gocon consumes exactly four bytes of memory.

Let’s compare Go with some other languages

Gocon 2014 (7)

Due to the overhead of the way Python represents variables, storing the same value using Python consumes six times more memory.

This extra memory is used by Python to track type information, do reference counting, etc

Let’s look at another example:

Gocon 2014 (8)

Similar to Go, the Java int type consumes 4 bytes of memory to store this value.

However, to use this value in a collection like a List or Map, the compiler must convert it into an Integer object.

Gocon 2014 (9)

So an integer in Java frequently looks more like this and consumes between 16 and 24 bytes of memory.

Why is this important ? Memory is cheap and plentiful, why should this overhead matter ?

Gocon 2014 (10)

This is a graph showing CPU clock speed vs memory bus speed.

Notice how the gap between CPU clock speed and memory bus speed continues to widen.

The difference between the two is effectively how much time the CPU spends waiting for memory.

Gocon 2014 (11)

Since the late 1960’s CPU designers have understood this problem.

Their solution is a cache, an area of smaller, faster memory which is inserted between the CPU and main memory.

Gocon 2014 (12)

This is a Location type which holds the location of some object in three dimensional space. It is written in Go, so each Location consumes exactly 24 bytes of storage.

We can use this type to construct an array type of 1,000 Locations, which consumes exactly 24,000 bytes of memory.

Inside the array, the Location structures are stored sequentially, rather than as pointers to 1,000 Location structures stored randomly.

This is important because now all 1,000 Location structures are in the cache in sequence, packed tightly together.

Gocon 2014 (13)

Go lets you create compact data structures, avoiding unnecessary indirection.

Compact data structures utilise the cache better.

Better cache utilisation leads to better performance.

Gocon 2014 (14)

Function calls are not free.

Gocon 2014 (15)

Three things happen when a function is called.

A new stack frame is created, and the details of the caller recorded.

Any registers which may be overwritten during the function call are saved to the stack.

The processor computes the address of the function and executes a branch to that new address.

Gocon 2014 (16)

Because function calls are very common operations, CPU designers have worked hard to optimise this procedure, but they cannot eliminate the overhead.

Depending on what the function does, this overhead may be trivial or significant.

A solution to reducing function call overhead is an optimisation technique called Inlining.

Gocon 2014 (17)

The Go compiler inlines a function by treating the body of the function as if it were part of the caller.

Inlining has a cost; it increases binary size.

It only makes sense to inline when the overhead of calling a function is large relative to the work the function does, so only simple functions are candidates for inlining.

Complicated functions are usually not dominated by the overhead of calling them and are therefore not inlined.

Gocon 2014 (18)

This example shows the function Double calling util.Max.

To reduce the overhead of the call to util.Max, the compiler can inline util.Max into Double, resulting in something like this

Gocon 2014 (19)

After inlining there is no longer a call to util.Max, but the behaviour of Double is unchanged.

Inlining isn’t exclusive to Go. Almost every compiled or JITed language performs this optimisation. But how does inlining in Go work?

The Go implementation is very simple. When a package is compiled, any small function that is suitable for inlining is marked and then compiled as usual.

Then both the source of the function and the compiled version are stored.

Gocon 2014 (20)

This slide shows the contents of util.a. The source has been transformed a little to make it easier for the compiler to process quickly.

When the compiler compiles Double it sees that util.Max is inlinable, and the source of util.Max is available.

Rather than insert a call to the compiled version of util.Max, it can substitute the source of the original function.

Having the source of the function enables other optimizations.

Gocon 2014 (21)

In this example, although the function Test always returns false, Expensive cannot know that without executing it.

When Test is inlined, we get something like this

Gocon 2014 (22)

The compiler now knows that the expensive code is unreachable.

Not only does this save the cost of calling Test, it saves compiling or running any of the expensive code that is now unreachable.

The Go compiler can automatically inline functions across files and even across packages. This includes code that calls inlinable functions from the standard library.

Gocon 2014 (23)

Mandatory garbage collection makes Go a simpler and safer language.

This does not imply that garbage collection makes Go slow, or that garbage collection is the ultimate arbiter of the speed of your program.

What it does mean is memory allocated on the heap comes at a cost. It is a debt that costs CPU time every time the GC runs until that memory is freed.

Gocon 2014 (24)

There is however another place to allocate memory, and that is the stack.

Unlike C, which forces you to choose if a value will be stored on the heap, via malloc, or on the stack, by declaring it inside the scope of the function, Go implements an optimisation called escape analysis.

Gocon 2014 (25)

Escape analysis determines whether any references to a value escape the function in which the value is declared.

If no references escape, the value may be safely stored on the stack.

Values stored on the stack do not need to be allocated or freed.

Lets look at some examples

Gocon 2014 (26)

Sum adds the numbers between 1 and 100 and returns the result. This is a rather unusual way to do this, but it illustrates how Escape Analysis works.

Because the numbers slice is only referenced inside Sum, the compiler will arrange to store the 100 integers for that slice on the stack, rather than the heap.

There is no need to garbage collect numbers, it is automatically freed when Sum returns.

Gocon 2014 (27)

This second example is also a little contrived. In CenterCursor we create a new Cursor and store a pointer to it in c.

Then we pass c to the Center() function which moves the Cursor to the center of the screen.

Then finally we print the X and Y locations of that Cursor.

Even though c was allocated with the new function, it will not be stored on the heap, because no reference c escapes the CenterCursor function.

Gocon 2014 (28)

Go’s optimisations are always enabled by default. You can see the compiler’s escape analysis and inlining decisions with the -gcflags=-m switch.

Because escape analysis is performed at compile time, not run time, stack allocation will always be faster than heap allocation, no matter how efficient your garbage collector is.

I will talk more about the stack in the remaining sections of this talk.

Gocon 2014 (30)

Go has goroutines. These are the foundations for concurrency in Go.

I want to step back for a moment and explore the history that leads us to goroutines.

In the beginning computers ran one process at a time. Then in the 60’s the idea of multiprocessing, or time sharing became popular.

In a time-sharing system the operating systems must constantly switch the attention of the CPU between these processes by recording the state of the current process, then restoring the state of another.

This is called process switching.

Gocon 2014 (29)

There are three main costs of a process switch.

First is the kernel needs to store the contents of all the CPU registers for that process, then restore the values for another process.

The kernel also needs to flush the CPU’s mappings from virtual memory to physical memory as these are only valid for the current process.

Finally there is the cost of the operating system context switch, and the overhead of the scheduler function to choose the next process to occupy the CPU.

Gocon 2014 (31)

There are a surprising number of registers in a modern processor. I have difficulty fitting them on one slide, which should give you a clue how much time it takes to save and restore them.

Because a process switch can occur at any point in a process’ execution, the operating system needs to store the contents of all of these registers because it does not know which are currently in use.

Gocon 2014 (32)

This lead to the development of threads, which are conceptually the same as processes, but share the same memory space.

As threads share address space, they are lighter than processes so are faster to create and faster to switch between.

Gocon 2014 (33)

Goroutines take the idea of threads a step further.

Goroutines are cooperatively scheduled, rather than relying on the kernel to manage their time sharing.

The switch between goroutines only happens at well defined points, when an explicit call is made to the Go runtime scheduler.

The compiler knows the registers which are in use and saves them automatically.

Gocon 2014 (34)

While goroutines are cooperatively scheduled, this scheduling is handled for you by the runtime.

Places where Goroutines may yield to others are:

  • Channel send and receive operations, if those operations would block.
  • The Go statement, although there is no guarantee that new goroutine will be scheduled immediately.
  • Blocking syscalls like file and network operations.
  • After being stopped for a garbage collection cycle.

Gocon 2014 (35)

This an example to illustrate some of the scheduling points described in the previous slide.

The thread, depicted by the arrow, starts on the left in the ReadFile function. It encounters os.Open, which blocks the thread while waiting for the file operation to complete, so the scheduler switches the thread to the goroutine on the right hand side.

Execution continues until the read from the c chan blocks, and by this time the os.Open call has completed so the scheduler switches the thread back the left hand side and continues to the file.Read function, which again blocks on file IO.

The scheduler switches the thread back to the right hand side for another channel operation, which has unblocked during the time the left hand side was running, but it blocks again on the channel send.

Finally the thread switches back to the left hand side as the Read operation has completed and data is available.

Gocon 2014 (36)

This slide shows the low level runtime.Syscall function which is the base for all functions in the os package.

Any time your code results in a call to the operating system, it will go through this function.

The call to entersyscall informs the runtime that this thread is about to block.

This allows the runtime to spin up a new thread which will service other goroutines while this current thread blocked.

This results in relatively few operating system threads per Go process, with the Go runtime taking care of assigning a runnable Goroutine to a free operating system thread.

Gocon 2014 (37)

In the previous section I discussed how goroutines reduce the overhead of managing many, sometimes hundreds of thousands of concurrent threads of execution.

There is another side to the goroutine story, and that is stack management, which leads me to my final topic.

Gocon 2014 (39)

This is a diagram of the memory layout of a process. The key thing we are interested is the location of the heap and the stack.

Traditionally inside the address space of a process, the heap is at the bottom of memory, just above the program (text) and grows upwards.

The stack is located at the top of the virtual address space, and grows downwards.

Gocon 2014 (40)

Because the heap and stack overwriting each other would be catastrophic, the operating system usually arranges to place an area of unwritable memory between the stack and the heap to ensure that if they did collide, the program will abort.

This is called a guard page, and effectively limits the stack size of a process, usually in the order of several megabytes.

Gocon 2014 (41)

We’ve discussed that threads share the same address space, so for each thread, it must have its own stack.

Because it is hard to predict the stack requirements of a particular thread, a large amount of memory is reserved for each thread’s stack along with a guard page.

The hope is that this is more than will ever be needed and the guard page will never be hit.

The downside is that as the number of threads in your program increases, the amount of available address space is reduced.

Gocon 2014 (42)

We’ve seen that the Go runtime schedules a large number of goroutines onto a small number of threads, but what about the stack requirements of those goroutines ?

Instead of using guard pages, the Go compiler inserts a check as part of every function call to check if there is sufficient stack for the function to run. If there is not, the runtime can allocate more stack space.

Because of this check, a goroutines initial stack can be made much smaller, which in turn permits Go programmers to treat goroutines as cheap resources.

Gocon 2014 (43)

This is a slide that shows how stacks are managed in Go 1.2.

When G calls to H there is not enough space for H to run, so the runtime allocates a new stack frame from the heap, then runs H on that new stack segment. When H returns, the stack area is returned to the heap before returning to G.

Gocon 2014 (44)

This method of managing the stack works well in general, but for certain types of code, usually recursive code, it can cause the inner loop of your program to straddle one of these stack boundaries.

For example, in the inner loop of your program, function G may call H many times in a loop,

Each time this will cause a stack split. This is known as the hot split problem.

Gocon 2014 (45)

To solve hot splits, Go 1.3 has adopted a new stack management method.

Instead of adding and removing additional stack segments, if the stack of a goroutine is too small, a new, larger, stack will be allocated.

The old stack’s contents are copied to the new stack, then the goroutine continues with its new larger stack.

After the first call to H the stack will be large enough that the check for available stack space will always succeed.

This resolves the hot split problem.

Gocon 2014 (46)

Values, Inlining, Escape Analysis, Goroutines, and segmented/copying stacks.

These are the five features that I chose to speak about today, but they are by no means the only things that makes Go a fast programming language, just as there more that three reasons that people cite as their reason to learn Go.

As powerful as these five features are individually, they do not exist in isolation.

For example, the way the runtime multiplexes goroutines onto threads would not be nearly as efficient without growable stacks.

Inlining reduces the cost of the stack size check by combining smaller functions into larger ones.

Escape analysis reduces the pressure on the garbage collector by automatically moving allocations from the heap to the stack.

Escape analysis is also provides better cache locality.

Without growable stacks, escape analysis might place too much pressure on the stack.

Gocon 2014 (47)

* Thank you to the Gocon organisers for permitting me to speak today
* twitter / web / email details
* thanks to @offbymany, @billkennedy_go, and Minux for their assistance in preparing this talk.

What does go build build ?

For packages

  • go build builds your package then discards the results.
  • go install builds then installs the package in your $GOPATH/pkg directory.

For commands

    • go build builds the command and leaves the result the current working directory.
    • go install builds the command in a temporary directory then moves it to $GOPATH/bin.

If you liked this post you may also find Using go test, build and install helpful.