18 Performance
Alan Perlis famously quipped “Lisp programmers know the value of
everything and the cost of nothing.” A Racket programmer knows, for
example, that a lambda anywhere in a program produces a value
that is closed over its lexical environment—
In this chapter, we narrow the gap by explaining details of the Racket compiler and run-time system and how they affect the run-time and memory performance of Racket code.
18.1 Performance in DrRacket
By default, DrRacket instruments programs for debugging, and debugging instrumentation can significantly degrade performance for some programs. Even when debugging is disabled through the Choose Language... dialog’s Show Details panel, the Preserve stacktrace checkbox is clicked by default, which also affects performance. Disabling debugging and stacktrace preservation provides performance results that are more consistent with running in plain racket.
Even so, DrRacket and programs developed within DrRacket use the same Racket virtual machine, so garbage collection times (see Memory Management) may be longer in DrRacket than when a program is run by itself, and DrRacket threads may impede execution of program threads. For the most reliable timing results for a program, run in plain racket instead of in the DrRacket development environment. Non-interactive mode should be used instead of the REPL to benefit from the module system. See Modules and Performance for details.
18.2 The Bytecode and Just-in-Time (JIT) Compilers
Every definition or expression to be evaluated by Racket is compiled to an internal bytecode format. In interactive mode, this compilation occurs automatically and on-the-fly. Tools like raco make and raco setup marshal compiled bytecode to a file, so that you do not have to compile from source every time that you run a program. (Most of the time required to compile a file is actually in macro expansion; generating bytecode from fully expanded code is relatively fast.) See Compilation and Configuration for more information on generating bytecode files.
The bytecode compiler applies all standard optimizations, such as constant propagation, constant folding, inlining, and dead-code elimination. For example, in an environment where + has its usual binding, the expression (let ([x 1] [y (lambda () 4)]) (+ 1 (y))) is compiled the same as the constant 5.
On some platforms, bytecode is further compiled to native code via a just-in-time or JIT compiler. The JIT compiler substantially speeds programs that execute tight loops, arithmetic on small integers, and arithmetic on inexact real numbers. Currently, JIT compilation is supported for x86, x86_64 (a.k.a. AMD64), and 32-bit PowerPC processors. The JIT compiler can be disabled via the eval-jit-enabled parameter or the --no-jit/-j command-line flag for racket.
The JIT compiler works incrementally as functions are applied, but the JIT compiler makes only limited use of run-time information when compiling procedures, since the code for a given module body or lambda abstraction is compiled only once. The JIT’s granularity of compilation is a single procedure body, not counting the bodies of any lexically nested procedures. The overhead for JIT compilation is normally so small that it is difficult to detect.
18.3 Modules and Performance
The module system aids optimization by helping to ensure that identifiers have the usual bindings. That is, the + provided by racket/base can be recognized by the compiler and inlined, which is especially important for JIT-compiled code. In contrast, in a traditional interactive Scheme system, the top-level + binding might be redefined, so the compiler cannot assume a fixed + binding (unless special flags or declarations are used to compensate for the lack of a module system).
Even in the top-level environment, importing with require enables some inlining optimizations. Although a + definition at the top level might shadow an imported +, the shadowing definition applies only to expressions evaluated later.
Within a module, inlining and constant-propagation optimizations take additional advantage of the fact that definitions within a module cannot be mutated when no set! is visible at compile time. Such optimizations are unavailable in the top-level environment. Although this optimization within modules is important for performance, it hinders some forms of interactive development and exploration. The compile-enforce-module-constants parameter disables the JIT compiler’s assumptions about module definitions when interactive exploration is more important. See Assignment and Redefinition for more information.
Currently, the compiler does not attempt to inline or propagate constants across module boundaries, except for exports of the built-in modules (such as the one that originally provides +).
The later section letrec Performance provides some additional caveats concerning inlining of module bindings.
18.4 Function-Call Optimizations
When the compiler detects a function call to an immediately visible function, it generates more efficient code than for a generic call, especially for tail calls. For example, given the program
(letrec ([odd (lambda (x) (if (zero? x) #f (even (sub1 x))))] [even (lambda (x) (if (zero? x) #t (odd (sub1 x))))]) (odd 40000000))
the compiler can detect the odd–even loop and produce code that runs much faster via loop unrolling and related optimizations.
Within a module form, defined variables are lexically scoped like letrec bindings, and definitions within a module therefore permit call optimizations, so
(define (odd x) ....) (define (even x) ....)
within a module would perform the same as the letrec version.
Primitive operations like pair?, car, and cdr are inlined at the machine-code level by the JIT compiler. See also the later section Fixnum and Flonum Optimizations for information about inlined arithmetic operations.
18.5 Mutation and Performance
Using set! to mutate a variable can lead to bad performance. For example, the microbenchmark
#lang racket/base (define (subtract-one x) (set! x (sub1 x)) x) (time (let loop ([n 4000000]) (if (zero? n) 'done (loop (subtract-one n)))))
runs much more slowly than the equivalent
#lang racket/base (define (subtract-one x) (sub1 x)) (time (let loop ([n 4000000]) (if (zero? n) 'done (loop (subtract-one n)))))
In the first variant, a new location is allocated for x on every iteration, leading to poor performance. A more clever compiler could unravel the use of set! in the first example, but since mutation is discouraged (see Guidelines for Using Assignment), the compiler’s effort is spent elsewhere.
More significantly, mutation can obscure bindings where inlining and constant-propagation might otherwise apply. For example, in
(let ([minus1 #f]) (set! minus1 sub1) (let loop ([n 4000000]) (if (zero? n) 'done (loop (minus1 n)))))
the set! obscures the fact that minus1 is just another name for the built-in sub1.
18.6 letrec Performance
When letrec is used to bind only procedures and literals, then the compiler can treat the bindings in an optimal manner, compiling uses of the bindings efficiently. When other kinds of bindings are mixed with procedures, the compiler may be less able to determine the control flow.
For example,
(letrec ([loop (lambda (x) (if (zero? x) 'done (loop (next x))))] [junk (display loop)] [next (lambda (x) (sub1 x))]) (loop 40000000))
likely compiles to less efficient code than
(letrec ([loop (lambda (x) (if (zero? x) 'done (loop (next x))))] [next (lambda (x) (sub1 x))]) (loop 40000000))
In the first case, the compiler likely does not know that display does not call loop. If it did, then loop might refer to next before the binding is available.
This caveat about letrec also applies to definitions of functions and constants as internal definitions or in modules. A definition sequence in a module body is analogous to a sequence of letrec bindings, and non-constant expressions in a module body can interfere with the optimization of references to later bindings.
18.7 Fixnum and Flonum Optimizations
A fixnum is a small exact integer. In this case, “small” depends on the platform. For a 32-bit machine, numbers that can be expressed in 30 bits plus a sign bit are represented as fixnums. On a 64-bit machine, 62 bits plus a sign bit are available.
A flonum is used to represent any inexact real number. They correspond to 64-bit IEEE floating-point numbers on all platforms.
Inlined fixnum and flonum arithmetic operations are among the most important advantages of the JIT compiler. For example, when + is applied to two arguments, the generated machine code tests whether the two arguments are fixnums, and if so, it uses the machine’s instruction to add the numbers (and check for overflow). If the two numbers are not fixnums, then it checks whether whether both are flonums; in that case, the machine’s floating-point operations are used directly. For functions that take any number of arguments, such as +, inlining works for two or more arguments (except for -, whose one-argument case is also inlined) when the arguments are either all fixnums or all flonums.
Flonums are typically boxed, which means that memory is allocated to hold every result of a flonum computation. Fortunately, the generational garbage collector (described later in Memory Management) makes allocation for short-lived results reasonably cheap. Fixnums, in contrast are never boxed, so they are typically cheap to use.
See Parallelism with Futures for an example use of flonum-specific operations.
The racket/flonum library provides flonum-specific operations, and combinations of flonum operations allow the JIT compiler to generate code that avoids boxing and unboxing intermediate results. Besides results within immediate combinations, flonum-specific results that are bound with let and consumed by a later flonum-specific operation are unboxed within temporary storage. Finally, the compiler can detect some flonum-valued loop accumulators and avoid boxing of the accumulator. The bytecode decompiler (see raco decompile: Decompiling Bytecode) annotates combinations where the JIT can avoid boxes with #%flonum, #%as-flonum, and #%from-flonum.
Unboxing of local bindings and accumualtors is not supported by the JIT for PowerPC.
The racket/unsafe/ops library provides unchecked fixnum- and flonum-specific operations. Unchecked flonum-specific operations allow unboxing, and sometimes they allow the compiler to reorder expressions to improve performance. See also Unchecked, Unsafe Operations, especially the warnings about unsafety.
18.8 Unchecked, Unsafe Operations
The racket/unsafe/ops library provides functions that are like other functions in racket/base, but they assume (instead of checking) that provided arguments are of the right type. For example, unsafe-vector-ref accesses an element from a vector without checking that its first argument is actually a vector and without checking that the given index is in bounds. For tight loops that use these functions, avoiding checks can sometimes speed the computation, though the benefits vary for different unchecked functions and different contexts.
Beware that, as “unsafe” in the library and function names suggest, misusing the exports of racket/unsafe/ops can lead to crashes or memory corruption.
18.9 Memory Management
The Racket implementation is available in two variants: 3m and CGC. The 3m variant uses a modern, generational garbage collector that makes allocation relatively cheap for short-lived objects. The CGC variant uses a conservative garbage collector which facilitates interaction with C code at the expense of both precision and speed for Racket memory management. The 3m variant is the standard one.
Although memory allocation is reasonably cheap, avoiding allocation altogether is normally faster. One particular place where allocation can be avoided sometimes is in closures, which are the run-time representation of functions that contain free variables. For example,
(let loop ([n 40000000] [prev-thunk (lambda () #f)]) (if (zero? n) (prev-thunk) (loop (sub1 n) (lambda () n))))
allocates a closure on every iteration, since (lambda () n) effectively saves n.
The compiler can eliminate many closures automatically. For example, in
(let loop ([n 40000000] [prev-val #f]) (let ([prev-thunk (lambda () n)]) (if (zero? n) prev-val (loop (sub1 n) (prev-thunk)))))
no closure is ever allocated for prev-thunk, because its only application is visible, and so it is inlined. Similarly, in
(let n-loop ([n 400000]) (if (zero? n) 'done (let m-loop ([m 100]) (if (zero? m) (n-loop (sub1 n)) (m-loop (sub1 m))))))
then the expansion of the let form to implement m-loop involves a closure over n, but the compiler automatically converts the closure to pass itself n as an argument instead.
18.10 Parallelism with Futures
The racket/future library provides support for performance improvement through parallelism with the future and touch functions. The level of parallelism available from those constructs, however, is limited by several factors, and the current implementation is best suited to numerical tasks.
Other functions, such as thread, support the creation of reliably concurrent tasks. However, thread never run truly in parallel, even if the hardware and operating system support parallelism.
As a starting example, the any-double? function below takes a list of numbers and determines whether any number in the list has a double that is also in the list:
(define (any-double? l) (for/or ([i (in-list l)]) (for/or ([i2 (in-list l)]) (= i2 (* 2 i)))))
This function runs in quadratic time, so it can take a long time (on the order of a second) on large lists like l1 and l2:
(define l1 (for/list ([i (in-range 5000)]) (+ (* 2 i) 1))) (define l2 (for/list ([i (in-range 5000)]) (- (* 2 i) 1))) (or (any-double? l1) (any-double? l2))
The best way to speed up any-double? is to use a different algorithm. However, on a machine that offers at least two processing units, the example above can run in about half the time using future and touch:
(let ([f (future (lambda () (any-double? l2)))]) (or (any-double? l1) (touch f)))
The future f runs (any-double? l2) in parallel to (any-double? l1), and the result for (any-double? l2) becomes available about the same time that it is demanded by (touch f).
Futures run in parallel as long as they can do so safely, but the notion of “safe” for parallelism is inherently tied to the system implementation. The distinction between “safe” and “unsafe” operations may be far from apparent at the level of a Racket program.
Consider the following core of a Mandelbrot-set computation:
(define (mandelbrot iterations x y n) (let ((ci (- (/ (* 2.0 y) n) 1.0)) (cr (- (/ (* 2.0 x) n) 1.5))) (let loop ((i 0) (zr 0.0) (zi 0.0)) (if (> i iterations) i (let ((zrq (* zr zr)) (ziq (* zi zi))) (cond ((> (+ zrq ziq) 4.0) i) (else (loop (add1 i) (+ (- zrq ziq) cr) (+ (* 2.0 zr zi) ci)))))))))
The expressions (mandelbrot 10000000 62 500 1000) and (mandelbrot 10000000 62 501 1000) each take a while to produce an answer. Computing them both, of course, takes twice as long:
(list (mandelbrot 10000000 62 500 1000) (mandelbrot 10000000 62 501 1000))
Unfortunately, attempting to run the two computations in parallel with future does not improve performance:
(let ([f (future (lambda () (mandelbrot 10000000 62 501 1000)))]) (list (mandelbrot 10000000 62 500 1000) (touch f)))
One problem is that the * and / operations in the first two lines of mandelbrot involve a mixture of exact and inexact real numbers. Such mixtures typically trigger a slow path in execution, and the general slow path is not safe for parallelism. Consequently, the future created in this example is almost immediately suspended, and it cannot resume until touch is called.
Changing the first two lines of mandelbrot addresses that first the problem:
(define (mandelbrot iterations x y n) (let ((ci (- (/ (* 2.0 (->fl y)) (->fl n)) 1.0)) (cr (- (/ (* 2.0 (->fl x)) (->fl n)) 1.5))) ....))
With that change, mandelbrot computations can run in parallel. Nevertheless, performance still does not improve. The problem is that most every arithmetic operation in this example produces an inexact number whose storage must be allocated. Especially frequent allocation triggers communication between parallel tasks that defeats any performance improvement.
By using flonum-specific operations (see Fixnum and Flonum Optimizations), we can re-write mandelbot to use much less allocation:
(define (mandelbrot iterations x y n) (let ((ci (fl- (fl/ (* 2.0 (->fl y)) (->fl n)) 1.0)) (cr (fl- (fl/ (* 2.0 (->fl x)) (->fl n)) 1.5))) (let loop ((i 0) (zr 0.0) (zi 0.0)) (if (> i iterations) i (let ((zrq (fl* zr zr)) (ziq (fl* zi zi))) (cond ((fl> (fl+ zrq ziq) 4.0) i) (else (loop (add1 i) (fl+ (fl- zrq ziq) cr) (fl+ (fl* 2.0 (fl* zr zi)) ci)))))))))
This conversion can speed mandelbrot by a factor of 8, even in sequential mode, but avoiding allocation also allows mandelbrot to run usefully faster in parallel.
As a general guideline, any operation that is inlined by the JIT compiler runs safely in parallel, while other operations that are not inlined (including all operations if the JIT compiler is disabled) are considered unsafe. The mzc decompiler tool annotates operations that can be inlined by the compiler (see raco decompile: Decompiling Bytecode), so the decompiler can be used to help predict parallel performance.
To more directly report what is happening in a program that uses future and touch, operations are logged when they suspend a computation or synchronize with the main computation. For example, running the original mandelbrot in a future produces the following output in the 'debug log level:
To see 'debug logging output on stderr, set the PLTSTDERR environment variable to debug or start racket with -W debug.
future 1, process 1: BLOCKING on process 0; time: .... |
.... |
future 1, process 0: HANDLING: *; time: .... |
The messages indicate which internal future-running task became blocked on an unsafe operation, the time it blocked (in terms of current-inexact-miliseconds), and the operation that caused the computation it to block.
The first revision to mandelbrot avoids suspending at *, but produces many log entries of the form
future 1, process 0: synchronizing: [allocate memory]; time: .... |
The [allocate memory] part of the message indicates that synchronization was needed for memory allocation.