11.5 Places

Parallelism with Places in The Racket Guide introduces places.

Places enable the development of parallel programs that take advantage of machines with multiple processors, cores, or hardware threads.

Currently, parallel support for places is enabled only for the 3m (main) and CS variants of Racket, and only by default for Windows, Linux x86/x86_64, and Mac OS x86/x86_64. To enable support for other platforms, use --enable-places with configure when building Racket. The place-enabled? function reports whether places run in parallel.

Implementation and operating-system constraints may limit the scalability of places. For example, although places can perform garbage collections independently in the 3m variant, a garbage collection may need to manipulate a page table that is shared across all places, and that shared page table can be a bottleneck with enough places—perhaps around 8 or 16.

A place is a parallel task that is effectively a separate instance of the Racket virtual machine, although all places run within a single operating-system process. Places communicate through place channels, which are endpoints for a two-way buffered communication.

To a first approximation, place channels support only immutable, transparent values as messages. In addition, place channels themselves can be sent across channels to establish new (possibly more direct) lines of communication in addition to any existing lines. Finally, mutable values produced by shared-flvector, make-shared-flvector, shared-fxvector, make-shared-fxvector, shared-bytes, and make-shared-bytes can be sent across place channels; mutation of such values is visible to all places that share the value, because they are allowed in a shared memory space. See place-message-allowed?.

A place channel can be used as a synchronizable event (see Events) to receive a value through the channel. A place channel is ready for synchronization when a message is available on the channel, and the place channel’s synchronization result is the message (which is removed on synchronization). A place can also receive messages with place-channel-get, and messages can be sent with place-channel-put.

Two place channels are equal? if they are endpoints for the same underlying channels while both or neither is a place descriptor. Place channels can be equal? without being eq? after being sent messages through a place channel.

Constraints on messages across a place channel—and therefore on the kinds of data that places share—enable greater parallelism than future, even including separate garbage collection of separate places. At the same time, the setup and communication costs for places can be higher than for futures.

For example, the following expression launches two places, echoes a message to each, and then waits for the places to terminate:

(let ([pls (for/list ([i (in-range 2)])

(dynamic-place "place-worker.rkt" 'place-main))])

(for ([i (in-range 2)]

[p pls])

(place-channel-put p i)

(printf "~a\n" (place-channel-get p)))

(map place-wait pls))

The "place-worker.rkt" module must export the place-main function that each place executes, where place-main must accept a single place channel argument:

#lang racket

(provide place-main)

(define (place-main pch)

(place-channel-put pch (format "Hello from place ~a"

(place-channel-get pch))))

Place channels are subject to garbage collection, like other Racket values, and a thread that is blocked reading from a place channel can be garbage collected if place channel’s writing end becomes unreachable. However, unlike normal channel blocking, if otherwise unreachable threads are mutually blocked on place channels that are reachable only from the same threads, the threads and place channels are all considered reachable, instead of unreachable.

When a place is created, its parameter values are generally set to the initial values of the parameters in the creating place, except that the current values of the following parameters are used: current-library-collection-paths, current-library-collection-links, and current-compiled-file-roots.

A newly created place is registered with the current custodian, so that the place is terminated when the custodian is shut down.

11.5.1 Using Places

The module indicated by module-path must export a function with the name start-name. The function must accept a single argument, which is a place channel that corresponds to the other end of communication for the place descriptor returned by place.

If location is provided, it must be a place location, such as a distributed places node produced by create-place-node.

When the place is created, the initial exit handler terminates the place, using the argument to the exit handler as the place’s completion value. Use (exit v) to immediately terminate a place with the completion value v. Since a completion value is limited to an exact integer between 0 and 255, any other value for v is converted to 0.

If the function indicated by module-path and start-name returns, then the place terminates with the completion value 0.

In the created place, the current-input-port parameter is set to an empty input port, while the values of the current-output-port and current-error-port parameters are connected to the current ports in the creating place. If the output ports in the creating place are file-stream ports, then the connected ports in the created place share the underlying streams, otherwise a thread in the creating place pumps bytes from the created place’s ports to the current ports in the creating place.

The module-path argument must not be a module path of the form (quote sym) unless the module is predefined (see module-predefined?).

The dynamic-place binding is protected in the sense of protect-out, so access to this operation can be prevented by adjusting the code inspector (see Code Inspectors).

The caller of dynamic-place* is responsible for closing all returned ports; none are closed automatically.

The dynamic-place* procedure returns four values:

• a place descriptor value representing the created place;

• an output port piped to the place’s standard input, or #f if in was a port;

• an input port piped from the place’s standard output, or #f if out was a port;

• an input port piped from the place’s standard error, or #f if err was a port or 'stdout.

The dynamic-place* binding is protected in the same way as dynamic-place.

The generated submodule has the name place-body-n for an integer n, and the submodule exports a main function that takes a place channel for the new place. The submodule is not intended for use, however, except by the expansion of the place form.

The place binding is protected in the same way as dynamic-place.

The place* binding is protected in the same way as dynamic-place.

If any pumping threads were created to connect a non-file-stream port to the ports in the place for p (see dynamic-place), place-wait returns only when the pumping threads have completed.

If any pumping threads were created to connect a non-file-stream port to the ports in the place for p (see dynamic-place), the event returned by place-dead-evt may become ready even if a pumping thread is still running.

Typically, one place channel is used by the current place to send messages to a destination place; the other place channel is sent to the destination place (via an existing place channel).

See place-message-allowed? form information on automatic coercions in v, such as converting a mutable string to an immutable string.

If (place-enabled?) returns #f, then the result is always #t and no conversions are performed on v as a message. Otherwise, the following kinds of data are allowed as messages:

• numbers, characters, booleans, keywords, and #<void>;

• symbols, where the eq?ness of uninterned symbols is preserved within a single message, but not across messages;

• strings and byte strings, where mutable strings and byte strings are automatically replaced by immutable variants;

• paths (for any platform);

• pairs, lists, vectors, and immutable prefab structures containing message-allowed values, where a mutable vector is automatically replaced by an immutable vector and where impersonators of vectors and prefab structures are copied;

• hash tables where mutable hash tables are automatically replaced by immutable variants, and where a hash table impersonator is copied;

• place channels, where a place descriptor is automatically replaced by a plain place channel;

• file-stream ports and TCP ports, where the underlying representation (such as a file descriptor, socket, or handle) is duplicated and attached to a fresh port in the receiving place;

• C pointers as created or accessed via ffi/unsafe; and

• values produced by shared-flvector, make-shared-flvector, shared-fxvector, make-shared-fxvector, shared-bytes, and make-shared-bytes.

A place location can be passed as the #:at argument to dynamic-place, which in turn simply calls the prop:place-location value of the place location.

A distributed places note created with create-place-node is an example of a place location.

11.5.2 Places Logging

Place events are reported to a logger named 'place. In addition to its string message, each event logged for a place has a data value that is an instance of a place-event prefab structure:

(struct place-event (place-id action value time)

#:prefab)

The place-id field is an exact integer that identifies a place.

The time field is an inexact number that represents time in the same way as current-inexact-milliseconds.

The action field is a symbol:

• 'create: a place was created. This event is logged in the creating place, and the event’s value field has the ID for the created place.

• 'reap: a place that was previously created in the current place has exited (and that fact has been detected, possibly via place-wait). The event’s value field has the ID for the exited place.

• 'enter: a place has started, logged within the started place. The event’s value field has #f.

• 'exit: a place is exiting, logged within the exiting place. The event’s value field has #f.

• 'put: a place-channel message has been sent. The event’s value field is a positive exact integer that approximates the message’s size.

• 'get: a place-channel message has been received. The event’s value field is a positive exact integer that approximates the message’s size.

Changed in version 6.0.0.2 of package base: Added logging via 'place and place-event.