Version 1.8.1
Released November 2nd, 2006 (corrections: Aug 31st, 2017)
Printed August 31, 2017
Editor: Dan Bonachea
http://gasnet.lbl.gov
Copyright © 2002-2017, Dan Bonachea.
Selected portions adapted from:
Permission is granted to freely distribute this specification and use it in creating GASNet clients or implementations. The authoritative version of the GASNet specification is maintained by Dan Bonachea and any proposed changes should be submitted for review.
Published by LBNL and U.C. Berkeley
This GASNet specification describes a network-independent and language-independent high-performance communication interface intended for use in implementing the runtime system for global address space languages (such as UPC or Titanium). GASNet stands for "Global-Address Space Networking".
The interface is divided into 2 layers - the GASNet core API and the GASNet extended API:
The core API is the minimum interface that must be implemented on each network when porting to a new system, and we provide a network-independent reference implementation of the extended API which is written purely in terms of the core API to ease porting and quick prototyping. Implementors for NIC’s that provide some hardware support for higher-level messaging operations (e.g. support for servicing remote reads/writes on the NIC without involving the main CPU) are encouraged to also implement an appropriate subset of the extended API directly on the network of interest (bypassing the core API) to achieve maximal performance for those operations (but this is an optimization and is not required to have a working system). Most clients will use calls to the extended API functions to implement the bulk of their communication work (thereby ensuring optimal performance across platforms). However the client is also permitted to use the core active message interface to implement non-trivial language-specific or compiler-specific communication operations which would not be appropriate in a language-independent API (e.g. implementing distributed language-level locks, distributed garbage collection, collective memory allocation, etc.).
Note the extended API interface is meant primarily as a low-level compilation target, not a library for hand-written code - as such, the goals of expressiveness and performance generally take precedence over readability and minimality.
gasnet_
GASNET_
gasnet_init(), and its associated
local memory space and system resources. The basic unit of control when
interfacing with GASNet.
Client code must #define exactly one of GASNET_PAR, GASNET_PARSYNC or
GASNET_SEQ when compiling the GASNet library and the client code (before
including gasnet.h) to indicate the threading environment.
GASNET_PAR
The most general configuration. Indicates a fully multi-threaded and thread-safe environment - the client may call GASNet concurrently from more than one thread. The exact threading system in use is system-specific, although for obvious reasons both GASNet and the client code must agree on the threading system - unless otherwise noted, the default mechanism is POSIX threads.
GASNET_PARSYNC
Indicates a multi-threaded but non-concurrent (non- threadsafe) GASNet environment, where multiple client threads may call GASNet, but their accesses to GASNet are fully serialized (e.g. by some level of synchronization above the GASNet interface). GASNet may safely assume that it will never be called from more than one client thread concurrently (and the client must ensure this property holds). Client code must still use GASNet No-Interrupt Sections and Handler-Safe Locks to ensure correct operation.
GASNET_SEQ
Indicates a single-threaded, non-threadsafe environment. GASNet may safely assume that it will only ever be called from one unique client thread. Client code must still use GASNet No-Interrupt Sections and Handler-Safe Locks to ensure correct operation.
|
Many GASNet core functions return 0 on success (GASNET_OK), or else they
return errors from the following list, as specified by each function:
GASNET_OK = 0 (no error)
GASNET_ERR_RESOURCE
GASNET_ERR_BAD_ARG
GASNET_ERR_NOT_INIT
GASNET_ERR_BARRIER_MISMATCH
GASNET_ERR_NOT_READY
Except where otherwise noted, errors that occur during a call to the extended API are fatal.
Many of the core API functions will return GASNET_ERR_RESOURCE to
indicate a generic failure in the hardware or communications system,
GASNET_ERR_BAD_ARG to indicate an illegal client argument, or
GASNET_ERR_NOT_INIT to indicate that gasnet_attach() has not been called.
If any node of a GASNet job crashes, aborts, or suffers a fatal hardware error, GASNet should make every attempt to ensure that the remaining nodes of the job are terminated in a timely manner to prevent creation of orphaned processes.
gasnet_ErrorName() and gasnet_ErrorDesc() convert the GASNet
error number errval into a string containing the name or description
(respectively) of the given error number. The client must not modify the
string returned.
gasnet_node_t
unsigned integer type representing a unique 0-based node index
gasnet_handle_t
an opaque type representing a non-blocking operation in-progress initiated using the extended API
gasnet_handler_t
an unsigned integer type representing an index into the core API AM handler table
gasnet_handlerarg_t
a 32-bit signed integer type which is used to express the user-provided arguments to all AM handlers. Platforms lacking a native 32-bit type may define this to a 64-bit type, but only the lower 32-bits are transmitted during an AM message send (and sign-extended on the receiver).
gasnet_token_t
an opaque type passed to core API handlers which may be used to query message information
gasnet_register_value_t
the largest unsigned integer type that can fit
entirely in a single CPU register for the current architecture and ABI.
SIZEOF_GASNET_REGISTER_VALUE_T is a preprocess-time literal integer constant
(i.e. not sizeof()) indicating the size of this type in bytes
gasnet_handlerentry_t
struct type used to negotiate handler registration
in gasnet_attach()
GASNET_SPEC_VERSION_MAJOR
GASNET_SPEC_VERSION_MINOR
Integral values corresponding to the major and minor version numbers of the GASNet specification version adhered to by a particular implementation. The minor version is incremented whenever new functionality is added to the specification without breaking backward compatibility. The major version is incremented whenever specification changes require breaking backward compatibility. The title page of this document provides the specification version corresponding to this version of the specification.
GASNET_RELEASE_VERSION_MAJOR
GASNET_RELEASE_VERSION_MINOR
GASNET_RELEASE_VERSION_PATCH
Integral values corresponding to the major, minor and patch version numbers of the release identifiers corresponding to the packaging on an implementation of GASNet. The significance of these values is implementation-defined.
GASNET_VERSION (deprecated)
equivalent to GASNET_SPEC_VERSION_MAJOR
GASNET_CONFIG_STRING
a string representing any of the relevant GASNet compile-time configuration settings that can be compared using string compare to verify version compatibility. The string is also embedded into the library itself such that it can be scanned for within a binary executable which is statically linked with GASNet.
GASNET_MAXNODES
an integer representing the maximum number of nodes supported in a single
GASNet job. This value must be representable as a gasnet_node_t.
GASNET_ALIGNED_SEGMENTS
defined by the GASNet implementation to the value 1 if gasnet_attach()
guarantees that the remote-access memory segment will be aligned at the same virtual
address on all nodes. Defined to 0 otherwise.
GASNET_PAGESIZE
a preprocessor constant integer which provides the memory granularity size used for various GASNet parameters which are required to be page-aligned. On many systems this will be the system page size.
gasnet_exit
The core API consists of:
Job startup in GASNet is a two-step process. GASNet programs should start by calling gasnet_init() as the first statement in their main() function, which bootstraps the nodes and establishes command-line arguments and the job environment. All nodes then call the gasnet_attach() function to initialize the network and register shared memory segments.
GASNet initialization may register some UNIX signal handlers (e.g. to support
interrupt-based implementations or aggressive segment registration policies). Client
code which registers signal handlers must be careful not to preempt any
GASNet-registered signal handlers (even for seemingly fatal signals such as
SIGABRT) - the only signal which the client may always safely catch is
SIGQUIT.
Any GASNet library implementation can be built in one of the following three
configurations, which affects the behavior of remote-access memory segment registration
during gasnet_attach().
The gasnet.h header file will define the appropriate preprocessor symbol
to indicate which configuration is active.
GASNET_SEGMENT_FAST
The remote-access memory segment is limited to an implementation-defined "reasonable" size,
and optimized in an implementation-specific way to provide the fastest possible
remote accesses. The maximum segment size may be queried using
gasnet_getMaxLocalSegmentSize().
GASNET_SEGMENT_LARGE
This configuration allows clients with larger shared data requirements to
register a larger remote-access memory segment, possibly at some cost in the
efficiency of remote accesses. The maximum segment size may be queried using
gasnet_getMaxLocalSegmentSize(), and should be comparable to the maximum
total data size allowed for processes on the given system.
GASNET_SEGMENT_EVERYTHING
The entire virtual memory space of each process is made available for remote access, in a way such that any memory access that would succeed when executed locally by this node would also succeed if executed by other nodes remotely. This can be used by clients which need to make the entire memory heap, stack and static data areas available for remote access.
|
Bootstraps a GASNet job and performs any system-specific setup required.
Called by all GASNet-based applications upon startup to bootstrap the nodes,
before any other processing takes place. Must be called before
any calls to any other functions in this specification, and before any
investigation of the command-line parameters passed to the program in
argc/argv, which may be modified or augmented by this call.
The semantics of any code executing before the call to gasnet_init()
is implementation-specific (for example, it is undefined whether
stdin/stdout/stderr are functional, or even how many nodes will run that code).
Upon return from gasnet_init(), all the nodes of the job will be running,
stdout/stderr will be functional, and the basic job environment will be established,
however the primary network resources may not yet have been initialized.
The following GASNet functions are the only ones that may be called between
gasnet_init() and gasnet_attach():
gasnet_mynode()gasnet_nodes()gasnet_getMaxLocalSegmentSize()gasnet_getMaxGlobalSegmentSize()gasnet_getenv()gasnet_exit()
All other GASNet calls are prohibited until after a successful gasnet_attach().
gasnet_init() may fail with a fatal error and implementation-defined message if
the nodes of the job cannot be successfully bootstrapped. It also may return
an error code such as GASNET_ERR_RESOURCE to indicate there was a problem
acquiring network or system resources. Otherwise, it returns GASNET_OK to indicate success.
May only be called once during a process lifetime, subsequent calls will
return an error.
typedef struct {
gasnet_handler_t index; // == 0 for don't care
void (*fnptr)();
} gasnet_handlerentry_t;
Initializes the GASNet network system and performs any system-specific setup required.
table is an array of numentries gasnet_handlerentry_t elements used for registering
active-message handlers provided by the client code. Clients that never explicitly
call the active-message request functions in the core API need not register
any handlers, and may pass a NULL pointer for table. Clients wishing to
register some handlers should fill in table with function pointers and
the desired handler index (or index 0 for "don’t-care") - note that handlers
0..127 are reserved for GASNet internal use, and handlers 128..255 are
available for client-provided handlers. Once gasnet_attach() returns, any
"don’t care" handler indexes in the table will be modified in place to
reflect the handler index assigned for each handler - the assignment
algorithm is deterministic: passing the same handler table on each node will
guarantee an identical resulting assignment on each node. Handler function
prototypes should match the prototypes described in the Active Message
Interface section.
segsize and minheapoffset are used to communicate the desired size and
location of the remote-access memory data segment for the local node that
will be used for all remote accesses (i.e. using the data transfer functions
of the extended API) or as the target of any Long active-messages in
the core API. The client passes the desired size of this area in bytes as
segsize, which must be a multiple of GASNET_PAGESIZE, and
should be less than or equal to the value returned by gasnet_getMaxLocalSegmentSize().
minheapoffset specifies the minimum amount of virtual memory space (in bytes)
to leave between the end of the current memory heap and the beginning of the
remote-access memory segment (on some systems the size of this offset
may limit the total future growth of the local memory heap, on other systems it may be irrelevant).
All nodes are required to pass the same value for minheapoffset.
Note that specifying a large
minheapoffset may limit the possible size of the remote-access segment on some systems.
Passing a segsize of zero disables the remote-access segment for this node, meaning
other nodes cannot access it with remote-memory operations and this node
cannot be the target of any Long AM messages.
GASNet will attempt to place the data segment in an area of the virtual memory space
whose pages are currently unused (e.g. by calling mmap).
The actual remote-access segment size achieved may be less than segsize if insufficient
system resources are available - the exact size and location of the segment
for all nodes should be queried after attach using gasnet_getSegmentInfo().
The segment assignment is guaranteed to have a GASNET_PAGESIZE-aligned base address and size,
but may differ in size across nodes, according to the requested segment sizes and
system resource availability.
GASNet will not initialize data within the memory segment in any way, nor
will it attempt to access the memory locations within the segment until
directed to do so by a data transfer function or Long active message.
If the GASNet implementation defines the macro GASNET_ALIGNED_SEGMENTS to 1, then
gasnet_attach() guarantees that the base of the remote-access memory segment will be
aligned at the same virtual address across all nodes (and will fail if it cannot
provide this). Otherwise, this guarantee is not provided. Note the segment sizes may
still differ across nodes, based on segsize and system resource availability.
In the GASNET_SEGMENT_FAST and GASNET_SEGMENT_LARGE
configurations, GASNet guarantees that data transfer functions,
Long active messages and local accesses referencing memory locations
in the remote-access memory segment will succeed, even before any local activity
takes place on those pages
(i.e. in an implementation performing lazy registration, first touch = allocate).
segsize and minheapoffset are ignored in the GASNET_SEGMENT_EVERYTHING
configuration, as the entire virtual memory space is implicitly shared for remote access.
Under this configuration, it is the client’s responsibility to ensure that any
remote-memory references fall within the legal areas of the current heap and data segment
for the target node -
remote accesses or Long active messages to locations outside these areas will have undefined effects
(for example, they may cause a segmentation fault on the target node).
gasnet_attach() may fail with a fatal error and implementation-defined message if
the network cannot be successfully initialized. It also may return
an error code such as GASNET_ERR_RESOURCE to indicate there was a problem
acquiring network or system resources. Otherwise, it returns GASNET_OK to indicate success.
A successful call acts as a global
barrier and blocks until all other nodes which are part of this parallel job
have successfully called gasnet_attach().
May only be called once during a process lifetime, subsequent calls will
return an error.
|
Retrieve an approximate, optimistic maximum size in bytes for the remote-access memory segment
that may be provided to gasnet_attach() under the current configuration.
The return value of this function may depend on
current system resource usage, and may return different values on different nodes
of a job, according to current system utilization. The value returned will
always be a multiple of GASNET_PAGESIZE.
The value returned is an optimistic approximation of the segment size
which can be acquired by gasnet_attach() - the actual size achieved can be queried
after attach using gasnet_getSegmentInfo().
On many implementations, this function will return different values in the
GASNET_SEGMENT_FAST and GASNET_SEGMENT_LARGE configurations.
Under the GASNET_SEGMENT_EVERYTHING configuration, this function returns -1.
This function has undefined behavior after gasnet_attach().
Returns a global minimum value
that would be returned by a call to gasnet_getMaxLocalSegmentSize
on any node of the current job (i.e. the smallest max segment size
estimated for any node in the job).
This function has undefined behavior after gasnet_attach().
Terminate the current GASNet job and return the given exitcode to the
console which invoked the job (in a system-specific way). This call is not
a collective operation, meaning any node may call it at any time after
initialization. It causes the system to flush all I/O, release all resources
and terminate the job for all active nodes. If several nodes and/or threads
call it simultaneously with different exit codes within a given
synchronization phase, the result provided to the console will be one of the
provided exit codes (chosen arbitrarily). This function should be called at
the end of main() after a barrier to ensure proper system exit, and should
also be called in the event of any fatal errors. GASNet clients are
encouraged to call gasnet_exit() before explicitly exiting (by calling
exit(), abort()) to reduce the possibility and lifetime of orphaned nodes,
but this is not required.
GASNet will send a SIGQUIT signal to the node if it detects that a remote
node has called gasnet_exit or crashed (in which case the node should catch
the signal, perform any system-specific shutdown, then call gasnet_exit() to
end the local node process). GASNet will also send a SIGQUIT signal if it
detects that the job has received a different catchable
terminate-the-program signal (e.g. SIGTERM, SIGINT) since some of these other signals
may be meaningful (and non-fatal) to certain GASNet implementations.
returns the unique, 0-based node index representing this node in the current GASNet job
returns the number of nodes in the current GASNet job
typedef struct {
void *addr;
uintptr_t size;
} gasnet_seginfo_t;
Query the segment base addresses and sizes for all the nodes in the job.
seginfo_table is an array of gasnet_seginfo_t
(and numentries is the number of entries in the table).
GASNet fills in the table with the remote-access segment base address and size
in bytes for each node whose index is less than numentries.
The value of numentries is usually equal to gasnet_nodes(), but is permitted
to be greater (in which case higher array entries are left untouched) or less (in which case
the higher-numbered nodes are not reported).
This is a non-collective operation.
Returns GASNET_OK on success.
Note that when GASNET_ALIGNED_SEGMENTS=1, the base addresses are guaranteed to be
equal (i.e. all remote-access segments start at the same virtual addresses).
However, in any case
the segment sizes may differ across nodes, and specifically they may differ from the
size requested by the client in the gasnet_attach() size hint.
Has the same semantics as the POSIX getenv() call, except it queries the
system-specific environment which was used to spawn the job (e.g. the
environment of the spawning console). Calling POSIX getenv() directly on
some implementations may not correctly return values reflecting the
environment that initiated the job spawn, consequently GASNet clients wishing to
query a consistent snapshot of the spawning environment across nodes
should never call getenv() directly. The semantics of POSIX setenv()
are undefined in GASNet jobs (specifically, it will probably fail to
propagate changes across nodes).
Active message communication is formulated as logically matching request and reply operations. Upon receipt of a request message, a request handler is invoked; likewise, when a reply message is received, the reply handler is invoked. Request handlers can reply at most once to the requesting node. If no explicit reply is made, the layer may generate one (to an implicit do-nothing reply handler). Thus a request handler can call reply at most once, and may only reply to the requesting node. Reply handlers cannot request or reply.
Here is a high-level description of a typical active message exchange between two nodes, A and B:
gasnet_AMRequest*() to send a request to B.
The call includes arguments, data payload, the node index of B and the index of
the request handler to run on B when the request arrives
gasnet_AMRequest*() call.
The request handler does some work on the arguments, and usually finishes
by calling gasnet_AMReply*() to issue a reply message before it exits (replying is
optional in GASNet, but required in AM2 - if the request handler does not
reply then no further actions are taken).
gasnet_AMReply*() takes the token passed to the request handler, arguments and data
payload, and the index of the reply handler to run when the reply message
arrives. It does not take a node index because a request handler is only
permitted to send a reply to the requesting node
gasnet_AMReply*() call.
The reply handler does some work on the arguments and then exits. It is
not permitted to send further messages.
The message layer will deliver requests and replies to destination nodes
barring any catastrophic errors (e.g. node crashes). From a sender’s point
of view, the request and reply functions block until the message is sent. A
message is defined to be sent once it is safe for the caller to reuse the storage
(registers or memory) containing the message (one notable exception to this
policy is gasnet_RequestLongAsyncM()). In implementations which copy or
buffer messages for transmission, the definition still holds: message sent
means the layer has copied the message and promises to deliver the copy with
its "best effort", and the original message storage may be reused. By best
effort, the message layer promises it will take care of all the details
necessary to transmit the message. These details include any retransmission
attempts and buffering issues on unreliable networks.
However, in either case, sent does not imply received. Once control returns from a request or reply function, clients cannot assume that the message has been received and handled at the destination. The message layer only guarantees that if a request or reply is sent, and, if the receiver occasionally polls for arriving messages, then the message will eventually be received and handled. From a receiver’s point of view, a message is defined to be received only once its handler function is invoked. The contents of partially received messages and messages whose handlers have not executed are undefined.
If the client sends an AM request or AM reply to a handler index which has not been registered on the destination node, GASNet will print an implementation-defined error message and terminate the job. It is implementation-defined whether this checking happens on the sending or receiving node.
There are three categories of active messages:
These messages carry only a few integer arguments (up to gasnet_AMMaxArgs())
handler prototype:
void handler(gasnet_token_t token,
gasnet_handlerarg_t arg0, ... gasnet_handlerarg_t argM-1);
In addition to integer arguments, these messages can can carry an opaque
data payload (up to gasnet_AMMaxMedium() bytes in length), that will
be made available to the handler when it is run on the remote node.
handler prototype:
void handler(gasnet_token_t token,
void *buf, size_t nbytes,
gasnet_handlerarg_t arg0, ... gasnet_handlerarg_t argM-1);
In addition to integer arguments, these messages can carry an opaque
data payload (up to gasnet_AMMaxLong{Request,Reply}() bytes in length) which is destined
for a particular predetermined address in the segment of the remote node
(often implemented using RDMA hardware assistance)
handler prototype:
void handler(gasnet_token_t token,
void *buf, size_t nbytes,
gasnet_handlerarg_t arg0, ... gasnet_handlerarg_t argM-1);
For more discussion on these three categories, see the Appendix.
The number of handler arguments (M) is specified upon issuing a request or reply by choosing the request/reply function of the appropriate name. The category of message and value of M used in the request/reply message sends determines the appropriate handler prototype, as detailed above. If a request or reply is sent to a handler whose prototype does not match the requirements as detailed above, the result is undefined.
|
These functions are used to query the maximum size messages of each category supported by a given implementation. These are likely to be implemented as macros for efficiency of client code which uses them (within packing loops, etc.)
Returns the maximum number of handler arguments (i.e. M) that may be passed
with any AM request or reply function. This value is guaranteed to be at
least (2 * MAX(sizeof(int),sizeof(void*))) (i.e. 8 for 32-bit systems, 16
for 64-bit systems), which ensures that 8 ints and/or pointers can be sent
with any active message. All implementations must support all values of M
from 0...gasnet_AMMaxArgs().
Returns the maximum number of bytes that can be sent in the payload of a single medium AM request or reply. This value is guaranteed to be at least 512 bytes on any implementation.
Returns the maximum number of bytes that can be sent in the payload of a single long AM request. This value is guaranteed to be at least 512 bytes on any implementation. Implementations which use RDMA to implement long messages are likely to support a much larger value.
Returns the maximum number of bytes that can be sent in the payload of a single long AM reply. This value is guaranteed to be at least 512 bytes on any implementation. Implementations which use RDMA to implement long messages are likely to support a much larger value.
In the function descriptions below, M is to be replaced with a number in [0 ... gasnet_AMMaxArgs()]
Send a short AM request to node dest,
to run the handler registered on the destination node at handler table index handler,
with the given M arguments.
gasnet_AMRequestShortM returns control to the calling thread of computation
after sending the request message. Upon receipt, the receiver invokes the
appropriate active message request handler function with the M integer
arguments.
Returns GASNET_OK on success.
Send a medium AM request to node dest, to run the handler registered on the destination node at handler table index handler, with the given M arguments.
The message also carries a data payload copied from the local node’s memory
space as indicated by source_addr and nbytes
(which need not fall within the registered data segment on the local node).
The value of nbytes must be no larger than the value returned by
gasnet_AMMaxMedium(), and is permitted to be zero (in which case
source_addr is ignored and the buf value passed to the handler
is undefined).
gasnet_AMRequestMediumM returns control to the calling thread of computation
after sending the associated request, and the source memory may be freely
modified once the function returns. The active message is logically
delivered after the data transfer finishes.
Upon receipt, the receiver invokes the appropriate request handler function
with a pointer to temporary storage containing the data payload
(in a buffer which is suitably aligned to hold any datatype), the number
of data bytes transferred, and the M integer arguments. The dynamic scope of
the storage is the same as the dynamic scope of the handler. The data should
be copied if it is needed beyond this scope.
Returns GASNET_OK on success.
Send a long AM request to node dest, to run the handler registered on the destination node at handler table index handler, with the given M arguments.
The message also carries a data payload copied from the local node’s memory
space as indicated by source_addr and nbytes
(which need not fall within the registered data segment on the local node).
The value of nbytes must be no larger than the value returned by
gasnet_AMMaxLongRequest(), and is permitted to be zero (in which case
source_addr is ignored and the buf value passed to the handler
is undefined).
The memory specified by [dest_addr...(dest_addr+nbytes-1)] must fall
entirely within the memory segment registered for remote access by the
destination node. This area will receive the data transfer before the handler runs.
If the source and destination memory overlap (e.g. in a loopback message),
the result is undefined.
gasnet_AMRequestLongM returns control to the calling thread of computation
after sending the associated request, and the source memory may be freely
modified once the function returns. The active message is logically
delivered after the bulk transfer finishes. Upon receipt, the receiver
invokes the appropriate request handler function with a pointer into the
memory segment where the data was placed, the number of data bytes
transferred, and the M integer arguments.
Returns GASNET_OK on success.
gasnet_AMRequestLongAsyncM() has identical semantics to
gasnet_AMRequestLongM(), except that the handler is required to send an AM reply
and the data payload source memory must NOT be modified until this matching
reply handler has begun execution.
Some implementations may leverage this additional constraint to provide
higher performance (e.g. by reducing extra data copying).
|
The following active message reply functions may only be called from the context of a running active message request handler, and a reply function may be called at most once from any given request handler (it is an error to do otherwise). The request and reply categories need not match (e.g. a short AM request handler may send a long AM reply).
Send a short AM reply to the indicated handler on the requesting node (i.e. the
node responsible for this particular invocation of the request handler), and
include the given M arguments.
gasnet_AMReplyShortM returns control to the calling thread of computation
after sending the reply message.
Upon receipt, the receiver invokes the appropriate active message reply
handler function with the M integer arguments.
Returns GASNET_OK on success.
Send a medium AM reply to the indicated handler on the requesting node (i.e. the
node responsible for this particular invocation of the request handler), with
the given M arguments and given data payload copied from the local node’s
memory space (source_addr need not fall within the registered data segment
on the local node).
The value of nbytes must be no larger than the value returned by
gasnet_AMMaxMedium(), and is permitted to be zero (in which case
source_addr is ignored and the buf value passed to the handler
is undefined).
gasnet_AMReplyMediumM returns control to the calling thread of computation
after sending the associated reply, and the source memory may be freely
modified once the function returns. The active message is logically
delivered after the data transfer finishes.
Upon receipt, the receiver invokes the appropriate reply handler function
with a pointer to temporary storage containing the data payload, the number
of data bytes transferred, and the M integer arguments. The dynamic scope of
the storage is the same as the dynamic scope of the handler. The data should
be copied if it is needed beyond this scope.
Returns GASNET_OK on success.
Send a long AM reply to the indicated handler on the requesting node (i.e. the
node responsible for this particular invocation of the request handler), with
the given M arguments and given data payload copied from the local node’s
memory space (source_addr need not fall within the registered data segment
on the local node).
The value of nbytes must be no larger than the value returned by
gasnet_AMMaxLongReply(), and is permitted to be zero (in which case
source_addr is ignored and the buf value passed to the handler
is undefined).
The memory specified by [dest_addr...(dest_addr+nbytes-1)] must fall
entirely within the memory segment registered for remote access by the
destination node.
If the source and destination memory overlap (e.g. in a loopback message),
the result is undefined.
gasnet_AMReplyLongM returns control to the calling thread of computation
after sending the associated reply, and the source memory may be freely
modified once the function returns. The active message is logically
delivered after the bulk transfer finishes.
Upon receipt, the receiver
invokes the appropriate reply handler function with a pointer into the
memory segment where the data was placed, the number of data bytes
transferred, and the M integer arguments.
Returns GASNET_OK on success.
An explicit call to service the network, process pending messages and run
handlers as appropriate.
Most of the message-sending primitives in GASNet poll the network
implicitly.
Purely polling-based implementations of GASNet may require occasional calls
to this function to ensure progress of remote nodes during compute-only
loops. Any client code which spin-waits for the arrival of a message should
call this function within the spin loop to optimize response time.
This call may be a no-op on some implementations (e.g. purely interrupt-based
implementations).
Returns GASNET_OK unless an error condition was detected.
#define GASNET_BLOCKUNTIL(cond) ???
This is a macro which implements a busy-wait/blocking polling loop in the way most efficient for the current GASNet core implementation. The macro blocks execution of the current thread and services the network until the provided condition becomes true. cond is an arbitrary C expression which will be evaluated by the macro one or more times as active messages arrive until the condition evaluates to a non-zero value. cond is an expression whose value is altered by the execution of an AM handler which the client thread is waiting for - GASNet may safely assume that the value of cond will only change while an AM handler is executing.
Example usage:
int doneflag = 0; gasnet_AMRequestShort1(..., &doneflag); // reply handler sets doneflag to 1 GASNET_BLOCKUNTIL(doneflag == 1);
Note that code like this would be illegal and could cause node 0 to sleep forever:
static int doneflag = 0; node 0: node 1: GASNET_BLOCKUNTIL(doneflag == 1); gasnet_put_val(0, &doneflag, 1, sizeof(int));
because gasnet_put_val (and other extended API functions) might not be
implemented using AM handlers. Also note that cond may be evaluated
concurrently with handler execution, so the client is responsible for negotiating
any atomicity concerns between the cond expression and handlers (for example, protecting
both with a handler-safe lock if the cond expression reads two or more values which are all
updated by handlers). Finally, note that unsynchronized handler code which modifies
one or more locations and then performs a flag write to signal a different thread
may need to execute a local memory barrier before the flag write to ensure correct
ordering on non-sequentially-consistent SMP hardware.
|
Can be called by handlers to query the source of the message being handled.
The token argument must be the token passed into the handler on entry.
Returns GASNET_OK on success.
Handlers may run asynchronously with respect to the main computation (in an
implementation which uses interrupts to run some or all handlers), and they
may run concurrently with each other on separate threads (e.g. in an
implementation where several threads may be polling the network at once). An
implementation using interrupts may result in handler code running within a
signal handler context. Some implementations may even choose to run handlers
on a separate private thread created by GASNet (making handlers asynchronous
with respect to all client threads). Note that polling-based GASNet
implementations are likely to poll (and possibly run handlers) from within
any GASNet call (i.e. not just gasnet_AMPoll()). Because of all this,
handler code should run quickly and to completion without making blocking
calls, and should not make assumptions about the context in which it is
being run (special care must be taken to ensure safety in a signal handler
context, see below).
Regardless, handlers themselves are not interruptible - any given thread
will only be running a single AM handler at a time and will never be
interrupted to run another AM handler (there is one exception to this rule -
the gasnet_AMReply*() call in a request handler may cause reply handlers to
run synchronously, which may be necessary to avoid deadlock in some
implementations. This should not be a problem since gasnet_AMReply*() is
often the last action taken by a request handler). Handlers are
specifically prohibited from initiating random network communication to
prevent deadlock - request handlers must generate at most one reply (to the
requestor) and make no other communication calls (including polling), and
reply handlers may not communicate or poll at all.
The asynchronous nature of handlers requires two mechanisms to make them safe: a mechanism to ensure signal safety for GASNet implementations using interrupt-based mechanisms, and a locking mechanism to allow atomic updates from handlers to data structures shared with the client threads and other handlers.
Traditionally, code running in signal handler context is extremely
circumscribed in what it can do: e.g. none of the standard pthreads/System V
synchronization calls are on the list of signal-safe functions (for such a
list see POSIX System Interfaces 2.4, IEEE Std 1003.1-2001).
Note that even most "thread-safe" libraries will break or deadlock if
called from a signal handler by the same thread currently executing a
different call to that library in an earlier stack frame. One specific case
where this is likely to arise in practice is calls to malloc()/free(). To
overcome these limitations, and allow our handlers to be more useful, the
normal limitations on signal handlers will be avoided by allowing the client
thread to temporarily disable the network interrupts that run handlers. All
function calls that are not signal-safe and could possibly access state
shared by functions also called from handlers MUST be called within a GASNet
"No-Interrupt Section":
gasnet_hold_interrupts() and gasnet_resume_interrupts() are used to define a
GASNet No-Interrupt Section (any code which dynamically executes between the
hold and resume calls is said to be "inside" the No-Interrupt Section).
These are likely to be implemented as macros and highly tuned for
efficiency.
The hold and resume calls must be paired, and may not be nested
recursively or the results are undefined (this means that clients should be
especially careful when calling other functions in the client from within a
No-Interrupt Section).
Both calls will return immediately in the common case, although one or both
may cause messages to be serviced on some implementations.
GASNet guarantees that no handlers will run asynchronously
on the current thread within the No-Interrupt Section.
The no-interrupt state is a
per-thread setting, and GASNet may continue running handlers synchronously or
asynchronously on other client threads or GASNet-private threads (even in a
GASNET_SEQ configuration) - specifically, a No-Interrupt Section does not
guarantee atomicity with respect to handler code, it merely provides a way
to ensure that handlers won’t run on a given thread while it’s inside a call
to a non-signal-safe library.
There is a strict set of conventions governing the use of No-Interrupt Sections which must be followed in order to ensure correct operation on all GASNet implementations. Clients which violate any of these rules may be subject to intermittent crashes, fatal errors or network deadlocks.
gasnet_hold_interrupts() and gasnet_resume_interrupts() need not be
called from within a handler context - handlers are run within an implicit
No-Interrupt Section, and gasnet_hold_interrupts() and gasnet_resume_interrupts()
calls are ignored within a handler context.
gasnet_mynode(),gasnet_nodes(),gasnet_hsl_*(),gasnet_exit(),gasnet_AMReply*()
Note that due to the previous rule, these are also the only GASNet functions
that may legally be called within a handler context (and gasnet_AMReply*()
is only legal in a request handler).
gasnet_hsl_lock within a No-Interrupt Section (subject to the rules
in section see Restrictions on Handler-Safe Locks).
|
In order to support handlers atomically updating data structures accessed by the main-line client code and other handlers, GASNet provides the Handler-Safe Lock (HSL) mechanism. As the name implies, these are a special kind of lock which are distinguished as being the only type of lock which may be safely acquired from a handler context. There is also a set of restrictions on their usage which allows this to be safe (see below). All lock-protected data structures in the client that need to be accessed by handlers should be protected using a Handler-Safe Lock (i.e. instead of a standard POSIX mutex).
gasnet_hsl_t is an opaque type representing a Handler-Safe Lock.
HSL’s operate analogously to POSIX mutexes, in that they are always
manipulated using a pointer.
gasnet_hsl_t hsl = GASNET_HSL_INITIALIZER;
Similarly to POSIX mutexes, HSL’s can be created in two ways. They can be
statically declared and initialized using the GASNET_HSL_INITIALIZER
constant. Alternately, HSL’s allocated using other means (such as dynamic
allocation) may be initialized by calling gasnet_hsl_init().
gasnet_hsl_destroy() may be called on either type of HSL once it’s no longer
needed to release any system resources associated with it.
It is erroneous to call gasnet_hsl_init() on a given HSL more than once. It
is erroneous to destroy an HSL which is currently locked. Any errors
detected in HSL initialization/destruction are fatal.
Lock and unlock HSL’s.
gasnet_hsl_lock(hsl) will block until the hsl lock can be
acquired by the current thread. gasnet_hsl_lock() may be called from within
main-line client code or from within handlers - this is the only blocking
call which is permitted to execute within a GASNet handler context (e.g. it
is erroneous to call POSIX mutex locking functions).
gasnet_hsl_trylock(hsl) attempts to acquire hsl for the current
thread, returning immediately (without blocking). If the lock was successfully
acquired, this function returns GASNET_OK. If the lock could not be
acquired (e.g it was found to be held by another thread) then this function
returns GASNET_ERR_NOT_READY and the lock is not acquired. It is not
legal for an AM handler to spin-poll a lock without bound using
gasnet_hsl_trylock() waiting for success - AM handlers must always use
gasnet_hsl_lock() when they wish to block to acquire an HSL.
gasnet_hsl_unlock(hsl) releases the hsl lock previously acquired
using gasnet_hsl_lock(hsl) or a successful
gasnet_hsl_trylock(hsl), and not yet released.
It is erroneous to call any of these functions on HSL’s which have not been
properly initialized.
Note that under the GASNET_SEQ configuration, HSL locking functions may only
be called from handlers and the designated GASNet client thread (not from
other client threads that may happen to exist - those threads are not
permitted to make any GASNet calls, which includes HSL locking calls).
All HSL locking/unlocking calls must follow the usage rules documented in the next section.
There is a strict set of conventions governing the use of HSL’s which must be followed in order to ensure correct operation on all GASNet implementations. Amongst other things, the restrictions are designed to ensure that HSL’s are always held for a strictly bounded amount of time, to ensure that acquiring them from within a handler can’t lead to deadlock. Clients which violate any of these rules may be subject to intermittent crashes, fatal errors or network deadlocks.
gasnet_hold_interrupts() and gasnet_resume_interrupts()
are ignored while holding an HSL.
gasnet_AMReply*()
gasnet_hsl_lock()
or gasnet_hsl_trylock(hsl) on a
lock already held by the current thread) and attempting to do so will lead
to undefined behavior. It is permitted for a thread to acquire more than
one HSL, although the traditional cautions about the possibility of deadlock
in the presence of multiple locks apply (e.g. the common solution is to
define a partial order on locks and always acquire them in a monotonically
ascending sequence).
|
Errors in calls to the extended API are considered fatal and abort the job
(by sending a SIGABRT signal) after printing an appropriate error message.
These comments apply to all put/get functions:
gasnet_attach()), or the results are undefined
gasnet_node_t) and a void * virtual memory address,
which logically represent a global pointer to the given address on the given
node. These global pointers need not be remote - the node rank passed to
these functions may in fact be the rank of the current node -
implementations must support this form of loopback, and should probably
attempt to optimize it by avoiding network traffic for such purely local
operations.
Blocking get/put operations for aligned data. The get operation fetches nbytes bytes from the address src on node node and places them at dest in the local memory space. The put operation sends nbytes bytes from the address src in the local address space, and places them at the address dest in the memory space of node node. A call to these functions blocks until the transfer is complete, and the contents of the destination memory are undefined until it completes. If the contents of the source memory change while the operation is in progress the result will be implementation-specific. The src and dest addresses (whether local or remote) must be properly aligned for accessing objects of size nbytes. nbytes must be >= 0 and has no maximum size, but implementations will likely optimize for small powers of 2.
Blocking get/put operations for bulk (unaligned) data. These function similarly to the aligned get/put operations above, except the data is permitted to be unaligned, and implementations are likely to optimize for larger sizes of nbytes.
Blocking operation that has the same effect as if the dest node had executed the POSIX call
memset(dest, val, nbytes).
As with puts, the destination memory must fall entirely within the memory area
registered for remote access by the dest node (see gasnet_attach).
The following functions provide non-blocking, split-phase memory access to shared data.
All such non-blocking operations require an initiation (generally a put or get) and a subsequent synchronization on the completion of that operation before the result is guaranteed.
There are two basic categories of non-blocking operations, defined by the synchronization mechanism used:
These operations return a specific handle from the initiation that is used for synchronization. The handle can be used to synchronize a specific subset of the nb operations in-flight
These operations don’t return a handle from the initiation - synchronization is accomplished by calling a synchronization routine that synchronizes all outstanding nbi operations.
Successful synchronization of a non-blocking get operation means the local result is ready to be examined, and will contain a value held by the source location at some time in the interval between the call to the initiation function and the successful completion of the synchronization (note this specifically allows implementations to delay the underlying read until the synchronization operation is called, provided they preserve the blocking semantics of the synchronization function).
Successful synchronization of a put operation means the source data has been written to the destination location and get operations issued subsequently by any thread (or load instructions issued by the destination node) will receive the new value or a subsequently written value (assuming no other threads are writing the location)
Note that the order in which non-blocking operations complete is intentionally unspecified - the system is free to coalesce and/or reorder non-blocking operations with respect to other blocking or non-blocking operations, or operations initiated from a separate thread - the only ordering constraints that must be satisfied are those explicitly enforced using the synchronization functions (i.e. the non-blocking operation is only guaranteed to occur somewhere in the interval between initiation and successful synchronization on that operation).
Implementors should attempt to make the non-blocking initiation operations return as quickly as possible - however in some cases (e.g. when a large number of non-blocking operations have been issued or the network is otherwise busy) it may be necessary to block temporarily while waiting for the network to become available. In any case, all implementations must support at least 2^16-1 non-blocking operations in-progress - that is, the client is free to issue up to 2^16-1 non-blocking operations before issuing a sync operation, and the implementation must handle this correctly without deadlock or livelock.
The explicit-handle non-blocking data transfer functions return a
gasnet_handle_t value to represent the non-blocking operation in flight.
gasnet_handle_t is an opaque scalar type whose contents are implementation-defined,
with one exception - every implementation must provide a scalar value corresponding
to an "invalid" handle (GASNET_INVALID_HANDLE) and furthermore this value
must be the result of setting all the bytes in the gasnet_handle_t datatype
to zero. Implementators are free to define the gasnet_handle_t type to be
any reasonable and appropriate size, although they are recommended to use a
type which fits within a single standard register on the target
architecture. In any case, the datatype should be wide enough to express at least
2^16-1
different handle values, to prevent limiting the number of
non-blocking operations in progress due to the number of handles available.
gasnet_handle_t has value semantics, so for example it
is permitted for clients to pass them across function call boundaries.
In the case of multithreaded clients (GASNET_PAR or GASNET_PARSYNC),
gasnet_handle_t values are thread-specific. In other words, it is an error
to obtain a handle value by initiating a non-blocking operation on one
thread, and later pass that handle into a synchronization function from a
different thread.
Any explicit-handle, non-blocking operation may return GASNET_INVALID_HANDLE
to indicate it was possible to complete the operation immediately without
blocking (e.g. operations where the "remote" node is actually the local
node)
It is always an error to discard the gasnet_handle_t value for an
explicit-handle operation in-flight - i.e. to initiate an operation and never
synchronize on its completion.
Non-blocking get/put functions for aligned data. These functions operate
similarly to their blocking counterparts, except they initiate a
non-blocking operation and return immediately with a handle (gasnet_handle_t)
which must later be used (by calling an explicit gasnet_*_syncnb*() function), to
synchronize on completion of the non-blocking operation. The contents of the
destination memory address are undefined until a synchronization completes
successfully for the non-blocking operation. For the put version, the source
memory may be safely overwritten once the initiation function returns.
Non-blocking get/put functions for bulk (unaligned) data. For the put version, the source memory may not be safely overwritten until a successful synchronization for the operation. If the contents of the source memory change while the operation is in progress the result will be implementation-specific. These otherwise behave identically to the non-bulk variants (but are likely to be optimized for large transfers).
Non-blocking operation that has the same effect as if the dest node had executed the POSIX call
memset(dest, val, nbytes).
As with puts, the destination memory must fall entirely within the memory area
registered for remote access by the dest node (see gasnet_attach).
The synchronization behavior is identical to a non-blocking, explicit-handle put operation (the
gasnet_handle_t return value must be synchronized using an explicit-handle
synchronization operation).
GASNet supports two basic types of synchronization for non-blocking
operations - trying (polling) and waiting (blocking). All explicit-handle
synchronization functions take one or more gasnet_handle_t values as input
and either return an indication of whether the operation has completed or
block until it completes.
Synchronize on the completion of a single specified explicit-handle
non-blocking operation that was initiated by the calling thread.
gasnet_wait_syncnb() blocks until the specified operation has completed (or
returns immediately if it has already completed). In any case, the handle
value is "dead" after gasnet_wait_syncnb() returns and may not be used in
future synchronization operations.
gasnet_try_syncnb() always returns immediately, with the value GASNET_OK if
the operation is complete (at which point the handle value is "dead", and
may not be used in future synchronization operations), or
GASNET_ERR_NOT_READY if the operation is not yet complete and future
synchronization is necessary to complete this operation.
It is legal to pass GASNET_INVALID_HANDLE as input to these functions -
gasnet_wait_sync(GASNET_INVALID_HANDLE) returns immediately and
gasnet_try_sync(GASNET_INVALID_HANDLE) returns GASNET_OK.
It is an error to pass a gasnet_handle_t value for an operation which has
already been successfully synchronized using one of the explicit-handle
synchronization functions.
Synchronize on the completion of an array of non-blocking explicit-handle
operations (all of which were initiated by this thread). numhandles
specifies the number of handles in the provided array of handles.
gasnet_wait_syncnb_all() blocks until all the specified operations have
completed (or returns immediately if they have all already completed).
gasnet_try_syncnb_all always returns immediately, with the value GASNET_OK
if all the specified operations have completed, or GASNET_ERR_NOT_READY if
one or more of the operations is not yet complete and future synchronization
is necessary to complete some of the operations.
Both functions will modify the provided array to reflect completions -
handles whose operations have completed are overwritten with the value
GASNET_INVALID_HANDLE, and the client may test against this value when
gasnet_try_syncnb_all() returns GASNET_ERR_NOT_READY to determine which
operations are complete and which are still pending.
It is legal to pass the value GASNET_INVALID_HANDLE in some of the array
entries, and both functions will ignore it so that it has no effect on
behavior. For example, if all entries in the array are GASNET_INVALID_HANDLE
(or numhandles==0), then gasnet_try_syncnb_all() will return GASNET_OK.
These operate analogously to the gasnet_*_syncnb_all variants, except they only
wait/test for at least one operation corresponding to a valid handle in
the provided list to be complete (the valid handles values are all those
which are not GASNET_INVALID_HANDLE). Specifically,
gasnet_wait_syncnb_some() will block until at least one of the valid handles
in the list has completed, and indicate the operations that have completed
by setting the corresponding handles to the value GASNET_INVALID_HANDLE.
Similarly, gasnet_try_syncnb_some will check if at least one valid handle in
the list has completed (setting those completed handles to
GASNET_INVALID_HANDLE) and return GASNET_OK if it detected at least one
completion or GASNET_ERR_NOT_READY otherwise.
Both functions ignore GASNET_INVALID_HANDLE values so those values have no
effect on behavior. If the input array is empty or consists only of
GASNET_INVALID_HANDLE values, gasnet_wait_syncnb_some will return
immediately and gasnet_try_syncnb_some will return GASNET_OK.
Non-blocking get/put functions for aligned and unaligned (bulk) data. These functions operate similarly to their explicit-handle counterparts, except they do not return a handle and must be synchronized using the implicit-handle synchronization operations. The contents of the destination memory address are undefined until a synchronization completes successfully for the non-blocking operation. As with the explicit-handle variants, the source memory for the non-bulk put operation may be safely overwritten once the initiation function returns, but the bulk put version requires the source memory to remain unchanged until the operation has been successfully completed using a synchronization.
gasnet_memset_nbi behaves identically to gasnet_memset_nb,
except that it is synchronized as if it were a non-blocking, implicit-handle put operation.
The following functions are used to synchronize implicit-handle non-blocking operations.
In the case of multithreaded clients, implicit-handle synchronization functions only synchronize the implicit-handle non-blocking operations initiated from the calling thread. Operations initiated by other threads sharing the GASNet interface proceed independently and are not synchronized. Implicit-handle synchronization functions will synchronize operations initiated within other function frames by the calling thread (but this cannot affect the correctness of correctly synchronized code).
These functions implicitly specify a set of non-blocking operations on which
to synchronize. They synchronize on a set of outstanding non-blocking
implicit-handle operations initiated by this thread - either all such gets,
all such puts, or all such puts and gets (where outstanding is defined as
all those implicit-handle operations which have been initiated (outside an
access region) but not yet completed through a successful implicit
synchronization). The wait variants block until all operations in this
implicit set have completed (indicating these operations have been
successfully synchronized). The try variants test whether all operations in
the implicit set have completed, and return GASNET_OK if so (which indicates
these operations have been successfully synchronized) or
GASNET_ERR_NOT_READY otherwise (in which case none of these operations may
be considered successfully synchronized).
If there are no outstanding implicit-handle operations, these
synchronization functions all return immediately (with GASNET_OK for the try
variants).
|
In some cases, it may be useful or desirable to initiate a number of
non-blocking shared-memory operations (possibly without knowing how many at
compile-time) and synchronize them at a later time using a single, fast
synchronization.
Simple implicit handle synchronization may not be appropriate for this
situation if there are intervening implicit accesses which are not to be
synchronized.
This situation could be handled using explicit-handle non-blocking
operations and a list synchronization (e.g. gasnet_wait_syncnb_all()), but
this may not be desirable because it requires managing an array of handles
(which could have negative cache effects on performance, or could be
expensive to allocate when the size is not known until runtime).
To handle these cases, we provide "implicit access region" synchronization,
described below.
gasnet_begin_nbi_accessregion() and gasnet_end_nbi_accessregion() are used
to define an implicit access region (any code which dynamically executes
between the begin and end calls is said to be "inside" the region)
The begin and end calls must be paired, and may not be nested recursively or
the results are undefined.
It is erroneous to call any implicit-handle synchronization function within
the access region.
All implicit-handle non-blocking operations initiated inside the region
become "associated" with the abstract access region handle being
constructed. gasnet_end_nbi_accessregion() returns an explicit handle which
jointly represents all the associated implicit-handle operations (those
initiated within the access region).
This handle can then be passed to the regular explicit-handle
synchronization functions, and will be successfully synchronized when all of
the associated non-blocking operations (both puts and gets) initiated in the
access region have completed.
The associated operations cease to be implicit-handle operations, and are
not synchronized by subsequent calls to the implicit-handle
synchronization functions occurring after the access region (e.g.
gasnet_wait_syncnbi_all()).
Explicit-handle operations initiated within the access region operate as
usual and do not become associated with the access region.
Sample code:
gasnet_begin_nbi_accessregion(); // begin the access region
gasnet_put_nbi(...); // becomes assoc. with access region
while (...) {
gasnet_put_nbi(...); // becomes assoc. with access region
}
// unrelated explicit-handle operation not assoc. with access region
h2 = gasnet_get_nb(...);
gasnet_wait_syncnb(h2);
// end the access region and get the handle
handle = gasnet_end_nbi_accessregion();
.... // other code, which may include unrelated implicit-handle
// operations+syncs, or other regions, etc
// wait for all the operations assoc. with access region to complete
gasnet_wait_syncnb(handle);
Register-memory operations allow client code to avoid forcing communicated data to pass through the local memory system. Some interconnects may be able to take advantage of this capability and launch remote puts directly from registers or recieve remote gets directly into registers.
Register-to-remote-memory put - these functions take the value to be put as
input parameter to avoid forcing outgoing values to local memory in client
code.
Otherwise, the behavior is identical to the memory-to-memory versions of put above.
Requires: nbytes > 0 && nbytes <= SIZEOF_GASNET_REGISTER_VALUE_T.
The value written to the target address is a direct byte copy of the
8*nbytes low-order bits of value, written with the endianness appropriate
for an nbytes integral value on the current architecture.
The non-blocking forms of value put must be synchronized using the explicit
or implicit synchronization functions defined above, as appropriate
This function returns the fetched value to avoid
forcing incoming values through local memory (on architectures
which pass the return value in a register).
Otherwise, the behavior is identical to the memory-to-memory blocking get.
Requires: nbytes > 0 && nbytes <= SIZEOF_GASNET_REGISTER_VALUE_T.
The value returned is the one obtained by reading the nbytes bytes starting
at the source address with the endianness appropriate for an nbytes integral
value on the current architecture and setting the high-order bits (if any)
to zero (i.e. no sign-extension)
This operates similarly to the blocking form of value get, but is split-phase. Non-blocking value gets are synchronized independently of all other operations in GASNet.
typedef ??? gasnet_valget_handle_t;
gasnet_get_nb_val initiates a non-blocking value get and returns an explicit
handle which must be synchronized using gasnet_wait_syncnb_valget.
gasnet_wait_syncnb_valget synchronizes one such outstanding operation
and returns the retrieved value as described for the blocking version.
Note that gasnet_valget_handle_t and gasnet_handle_t are completely
different datatypes and may not be intermixed (i.e. gasnet_valget_handle_t
cannot be used with other synchronization functions, and
gasnet_handle_t cannot be passed to gasnet_wait_syncnb_valget).
The gasnet_valget_handle_t type is completely opaque (with no special
"invalid" value), although implementors are recommended to make
sizeof(gasnet_valget_handle_t) <= sizeof(gasnet_register_value_t) to
facilitate register reuse.
There is no try variant of value get synchronization, and no implicit-handle variant.
The following functions can be used to execute a parallel split-phase
barrier with the given barrier identifier across all nodes in the job.
Note that the barrier wait/notify functions should only be called once (i.e.
by one representative thread) on each node per barrier phase.
The client must synchronize its own accesses to the barrier functions and
ensure that only one thread is ever inside a GASNet barrier function at a
time (esp. gasnet_barrier_try()).
#define GASNET_BARRIERFLAG_ANONYMOUS ??? #define GASNET_BARRIERFLAG_MISMATCH ???
Execute the notification for a split-phase barrier, with a barrier value id. This is a non-blocking operation that completes immediately after noting the barrier value. No synchronization is performed on outstanding non-blocking memory operations.
Generates a fatal error if this is the second call to
gasnet_barrier_notify() on this node since the last call to
gasnet_barrier_wait() or the beginning of the program.
If flags == 0 then this is a "named" barrier notify that carries the given
id value.
If flags == GASNET_BARRIERFLAG_ANONYMOUS, then id is ignored and the
barrier is anonymous - it has no specific value.
If flags == GASNET_BARRIERFLAG_MISMATCH, then the subsequent
gasnet_barrier_wait() call on every node will return
GASNET_ERR_BARRIER_MISMATCH (i.e. allows the client to force a global
mismatch error when a mismatch was detected locally).
Execute the wait for a split-phase barrier, with a barrier value.
This is a blocking operation that returns only after all remote nodes have
called gasnet_barrier_notify().
No synchronization is performed on outstanding non-blocking memory
operations .
Generates a fatal error if there were no preceding calls to
gasnet_barrier_notify() on this node, or if this is the second call to
gasnet_barrier_wait() (or successful call to gasnet_barrier_try()) since the
last call to gasnet_barrier_notify() on this node.
On a GASNET_PAR or GASNET_PARSYNC configuration, the thread calling
gasnet_barrier_notify() is permitted to differ from the thread which calls
the paired gasnet_barrier_wait(), but the ordering between the calls must
still be maintained.
Returns GASNET_ERR_BARRIER_MISMATCH if flags is not equal to the flags value
passed to the preceding gasnet_barrier_notify() call made by this node.
Returns GASNET_ERR_BARRIER_MISMATCH if the flags value passed to
gasnet_barrier_notify() on this or any other node was
GASNET_BARRIERFLAG_MISMATCH.
Returns GASNET_ERR_BARRIER_MISMATCH if flags==0 and the supplied id value
doesn’t match the id value provided in the preceding gasnet_barrier_notify()
call made by this node.
Returns GASNET_ERR_BARRIER_MISMATCH if any two nodes passed non-anonymous
barrier values which didn’t match during the gasnet_barrier_notify() calls
which began this barrier phase.
Otherwise, returns GASNET_OK to indicate that all nodes have called a
matching gasnet_barrier_notify() and the barrier phase is complete.
gasnet_barrier_try() functions similarly to gasnet_wait(), except that it
always returns immediately.
If the barrier has been notified by all nodes, the call behaves as a call to
gasnet_barrier_wait() with the same barrier id and flags, and returns
GASNET_OK (or GASNET_ERR_BARRIER_MISMATCH in the case a mismatch is
detected).
If the barrier has not yet been notified by some node, the call is a no-op
and returns the value GASNET_ERR_NOT_READY.
Generates a fatal error if there were no preceding calls to
gasnet_barrier_notify() on this node, or if this is the second call to
gasnet_barrier_wait() (or successful call to gasnet_barrier_try()) since the
last call to gasnet_barrier_notify() on this node.
When compiled in the GASNET_PAR or GASNET_PARSYNC configurations, GASNet is
capable of handling multiple client threads. It is likely that GASNet
implementations will need to distinguish these threads, specifically they
may need to store some metadata associated with each client thread.
Unfortunately, the overhead of discovering the identity of a particular
client thread making a GASNet call (hereafter termed "thread discovery") can
have a non-trivial overhead on some threading systems (e.g. the cost of
calling pthread_self() or pthread_getspecific()). Many of the simpler GASNet
functions could have their performance dominated by this cost if they need
to perform thread discovery on every call.
The following macros provide a way for the client to amortize the cost of
thread discovery over many GASNet calls made by the same thread. This is an
optimization which is totally optional - clients need not make any of the
calls below to have a working system, although GASNet performance may
suffer without it in a GASNET_PAR or GASNET_PARSYNC configuration
on some platforms.
typedef void *gasnet_threadinfo_t;
gasnet_threadinfo_t is an opaque pointer representing the internal GASNet
metadata associated with a particular client thread.
#define GASNET_GET_THREADINFO() ???
Returns a value of type gasnet_threadinfo_t which represents the GASNet
internal metadata associated with the current client thread. This
gasnet_threadinfo_t value can be passed into or out of functions and may be
posted for GASNet’s use with GASNET_POST_THREADINFO(). May be called from
anywhere in the client program, at any time after GASNet initialization. It
is erroneous to hand-off this gasnet_threadinfo_t value to a different
client thread.
#define GASNET_POST_THREADINFO(info) ???
This macro may optionally be placed (followed by a semi-colon)
at the top of functions which make
calls to GASNet. It has no runtime semantics, but it may provide a
performance boost on some implementations (especially in functions which
make multiple calls to the extended API - e.g. it provides the
implementation with a place for minimal per-function initialization or
temporary storage that may be helpful in amortizing implementation-specific
overheads).
When used, it must appear only at the very beginning of a function or block
(before any declarations or calls to the API in that function). It may not
appear as a global declaration. The info argument must be a
gasnet_threadinfo_t value acquired from a previous call to
GASNET_GET_THREADINFO() on this thread.
#define GASNET_BEGIN_FUNCTION() ???
A convenience macro that may optionally be placed (followed by a semi-colon) at the top of functions which repeatedly make GASNet calls, to amortize the overhead of thread discovery on some implementations.
It has behavior equivalent to GASNET_POST_THREADINFO(GASNET_GET_THREADINFO()),
however some implementations may choose to lazily postpone performing thread discovery
until the first place where it is actually needed.
Optional call which gives the GASNet implementation a hint about how aggressively threads within blocking GASNet calls should contend for CPU resources. wait_mode must be one of the following recognized values:
GASNET_WAIT_SPIN
contend aggressively for CPU resources while waiting (spin)
GASNET_WAIT_BLOCK
yield CPU resources immediately while waiting (block)
GASNET_WAIT_SPINBLOCK
spin for an implementation-dependent period, then block
Wait mode is a per-node hint which is permitted to differ across GASNet nodes.
Returns GASNET_OK on success.
The GASNet core API was originally based on Active Messages 2.0 (as described in A. Mainwaring and D. Culler in "Active Message Applications Programming Interface and Communication Subsystem Organization"), however we’ve removed some of the generality which is not required (and can lead to performance degradation and more implementation effort), and stripped it down to the bare essentials required for active messages in a purely SPMD environment. The final spec more closely resembles the "Generic Active Message Interface Specification v.1.1", by D.Culler et al., however we describe the differences from AM2.0 for readers familiar with that specification (and because we envision a number of the GASNet core implementations being simply a thin wrapper over the existing AM2.0 implementations on a number of platforms).
Here are a summary of the changes (informal style.. this is not really part of the spec):
AM_Init, AM_Terminate,
AM_AllocateBundle, AM_AllocateEndpoint, AM_FreeEndpoint, AM_FreeBundle,
AM_MoveEndpoint, AM_GetXferM, AM_GetDestEndpoint
gasnet_attach(), and the
maximum number of handlers is fixed at 256 (including handler 0, the error
handler)
AM_SetHandler and
AM_SetHandlerAny, AM_GetNumHandlers, AM_SetNumberHandlers, AM_MaxNumHandlers
gasnet_attach() (using a uintptr_t to
allow entire VA space)
AM_SetSeg and AM_MaxSegLength
(still have AM_GetSeg)
gasnet_attach requests a size larger than what underlying
AM_SetSeg can provide, then we turn off large AM Xfers and emulate
gasnet_Xfer using medium messages)
dest_offset argument to the Xfer functions is changed to a void *
address
AM_Map, AM_MapAny,
AM_Unmap, AM_SetTag, AM_GetTag, AM_GetTranslationName, AM_GetTranslationTag,
AM_GetTranslationInUse, AM_MaxNumTranslations, AM_GetNumTranslations,
AM_SetNumTranslations, AM_GetMsgTag
en_t * argument to AM_GetSourceEndpoint is now an gasnet_node_t *
and returns the node rank of the sender (the now-opaque token could be
implemented as the integer node index itself, although we allow
implementations to still use it as a ptr to metadata if required)
AM_RequestXferAsyncM has more useful semantics (may block)
AM_SetExpectedResources no longer exists
AM_PAR (multi-threaded) access mode
(GASNET_PAR configuration)
(void*)’s can be sent)
cons: handler code needs to be rewritten for 64-bit platforms to perform packing/unpacking
AM_GetEventMask and AM_SetEventMask no longer exist
AM_WaitSema is replaced with GASNET_BLOCKUNTIL()
ReplyXfer in favor of GetXfer
ReplyXfer’s (with software
flow control & reliability)
AM_MaxLong into AM_MaxLongRequest, and
AM_MaxLongReply
GetXfer doesn’t add any expressiveness - really want a way to get
from remote segment into arbitrary local memory address
Newcomers to Active Messages and GASNet occasionally express confusion over the concepts of Short, Medium and Long AM’s. Despite the somewhat misleading naming convention, the three categories of messages may actually bear only a loose correllation to the actual message/data sizes. The important distinctions are semantic, and sufficiently minor that one might imagine replacing the three categories with a single, more general type of AM that provides the functionality of each GASNet AM category as a special case.
Specifically, a GASNet Short AM can be seen as a special case of a Medium or Long AM where the payload has length zero. Furthermore, the only important semantic distinction between Medium and Long AM’s are that Medium AM’s provide the payload to the handler in a temporary network buffer, whereas Long AM’s write the payload (often using RDMA) to a sender-specified location in the user memory segment of the target node before running the handler (each semantic is useful for different usage scenarios).
Hence, some users may find it helpful to consider building "unified" AM request/reply functions such as suggested below:
/* unified request function */
int unified_AMRequestM(
gasnet_node_t dest, gasnet_handler_t handler,
void *buf, size_t buf_len,
void *dest_addr,
int32 arg0, int32 arg1, ...) {
if (buf == NULL)
return gasnet_AMRequestShortM(dest,handler,arg0,arg1,...);
else if (dest_addr == NULL)
return gasnet_AMRequestMediumM(dest,handler,buf,buf_len,arg0,arg1,...);
else
return gasnet_AMRequestLongM(dest,handler,buf,buf_len,dest_addr,arg0,arg1,...);
}
gasnet_AMMaxArgs()
dest_addr == NULL requires buf_len <= gasnet_AMMaxMedium(),
and the payload is delivered in a temporary buffer
dest_addr != NULL requires buf_len <= gasnet_AMMaxLongRequest()
(or gasnet_AMMaxLongReply() for replies),
and the payload is written into the target node segment at dest_addr
buf == NULL:
void handler_nopayload(gasnet_token_t token,
gasnet_handlerarg_t arg0, gasnet_handlerarg_t arg1...);
buf != NULL:
void handler_withpayload(gasnet_token_t token,
void *buf, size_t buf_len,
gasnet_handlerarg_t arg0, gasnet_handlerarg_t arg1...);
| Jump to: | A B C D E G I J L M N P R S T U |
|---|
| Jump to: | A B C D E G I J L M N P R S T U |
|---|