| fi_trigger(3) | #VERSION# | fi_trigger(3) |
fi_trigger - Triggered operations
#include <rdma/fi_trigger.h>
Triggered operations allow an application to queue a data transfer request that is deferred until a specified condition is met. A typical use is to send a message only after receiving all input data. Triggered operations can help reduce the latency needed to initiate a transfer by removing the need to return control back to an application prior to the data transfer starting.
An endpoint must be created with the FI_TRIGGER capability in order for triggered operations to be specified. A triggered operation is requested by specifying the FI_TRIGGER flag as part of the operation. Such an endpoint is referred to as a trigger-able endpoint.
Any data transfer operation is potentially trigger-able, subject to provider constraints. Trigger-able endpoints are initialized such that only those interfaces supported by the provider which are trigger-able are available.
Triggered operations require that applications use struct fi_triggered_context as their per operation context parameter, or if the provider requires the FI_CONTEXT2 mode, struct fi_trigger_context2. The use of struct fi_triggered_context[2] replaces struct fi_context[2], if required by the provider. Although struct fi_triggered_context[2] is not opaque to the application, the contents of the structure may be modified by the provider once it has been submitted as an operation. This structure has similar requirements as struct fi_context[2]. It must be allocated by the application and remain valid until the corresponding operation completes or is successfully canceled.
Struct fi_triggered_context[2] is used to specify the condition that must be met before the triggered data transfer is initiated. If the condition is met when the request is made, then the data transfer may be initiated immediately. The format of struct fi_triggered_context[2] is described below.
struct fi_triggered_context {
enum fi_trigger_event event_type; /* trigger type */
union {
struct fi_trigger_threshold threshold;
struct fi_trigger_xpu xpu;
void *internal[3]; /* reserved */
} trigger;
};
struct fi_triggered_context2 {
enum fi_trigger_event event_type; /* trigger type */
union {
struct fi_trigger_threshold threshold;
struct fi_trigger_xpu xpu;
void *internal[7]; /* reserved */
} trigger;
};
The triggered context indicates the type of event assigned to the trigger, along with a union of trigger details that is based on the event type.
Completion based triggers defer a data transfer until one or more related data transfers complete. For example, a send operation may be deferred until a receive operation completes, indicating that the data to be transferred is now available.
The following trigger event related to completion based transfers is defined.
struct fi_trigger_threshold {
struct fid_cntr *cntr; /* event counter to check */
size_t threshold; /* threshold value */
};
Threshold operations are triggered in the order of the threshold values. This is true even if the counter increments by a value greater than 1. If two triggered operations have the same threshold, they will be triggered in the order in which they were submitted to the endpoint.
XPU based triggers work in conjunction with heterogenous memory (FI_HMEM capability). XPU triggers define a split execution model for specifying a data transfer separately from initiating the transfer. Unlike completion triggers, the user controls the timing of when the transfer starts by writing data into a trigger variable location.
XPU transfers allow the requesting and triggering to occur on separate computational domains. For example, a process running on the host CPU can setup a data transfer, with a compute kernel running on a GPU signaling the start of the transfer. XPU refers to a CPU, GPU, FPGA, or other acceleration device with some level of computational ability.
Endpoints must be created with both the FI_TRIGGER and FI_XPU capabilities to use XPU triggers. XPU triggered enabled endpoints only support XPU triggered operations. The behavior of mixing XPU triggered operations with normal data transfers or non-XPU triggered operations is not defined by the API and subject to provider support and implementation.
The use of XPU triggers requires coordination between the fabric provider, application, and submitting XPU. The result is that hardware implementation details need to be conveyed across the computational domains. The XPU trigger API abstracts those details. When submitting a XPU trigger operation, the user identifies the XPU where the triggering will occur. The triggering XPU must match with the location of the local memory regions. For example, if triggering will be done by a GPU kernel, the type of GPU and its local identifier are given. As output, the fabric provider will return a list of variables and corresponding values. The XPU signals that the data transfer is safe to initiate by writing the given values to the specified variable locations. The number of variables and their sizes are provider specific.
XPU trigger operations are submitted using the FI_TRIGGER flag with struct fi_triggered_context or struct fi_triggered_context2, as required by the provider. The trigger event_type is:
struct fi_trigger_var {
enum fi_datatype datatype;
int count;
void *addr;
union {
uint8_t val8;
uint16_t val16;
uint32_t val32;
uint64_t val64;
uint8_t *data;
} value;
};
struct fi_trigger_xpu {
int count;
enum fi_hmem_iface iface;
union {
uint64_t reserved;
int cuda;
int ze;
} device;
struct fi_trigger_var *var;
};
On input to a triggered operation, the iface field indicates the software interface that will be used to write the variables. The device union specifies the device identifier. For valid iface and device values, see fi_mr(3). The iface and device must match with the iface and device of any local HMEM memory regions. Count should be set to the number of fi_trigger_var structures available, with the var field pointing to an array of struct fi_trigger_var. The user is responsible for ensuring that there are sufficient fi_trigger_var structures available and of an appropriate size. The count and size of fi_trigger_var structures can be obtained by calling fi_getopt() on the endpoint with the FI_OPT_XPU_TRIGGER option. See fi_endpoint(3) for details.
Each fi_trigger_var structure referenced should have the datatype and count fields initialized to the number of values referenced by the struct fi_trigger_val. If the count is 1, one of the val fields will be used to return the necessary data (val8, val16, etc.). If count > 1, the data field will return all necessary data used to signal the trigger. The data field must reference a buffer large enough to hold the returned bytes.
On output, the provider will set the fi_trigger_xpu count to the number of fi_trigger_var variables that must be signaled. Count will be less than or equal to the input value. The provider will initialize each valid fi_trigger_var entry with information needed to signal the trigger. The datatype indicates the size of the data that must be written. Valid datatype values are FI_UINT8, FI_UINT16, FI_UINT32, and FI_UINT64. For signal variables <= 64 bits, the count field will be 1. If a trigger requires writing more than 64-bits, the datatype field will be set to FI_UINT8, with count set to the number of bytes that must be written. The data that must be written to signal the start of an operation is returned through either the value union val fields or data array.
Users signal the start of a transfer by writing the returned data to the given memory address. The write must occur from the specified input XPU location (based on the iface and device fields). If a transfer cannot be initiated for some reason, such as an error occurring before the transfer can start, the triggered operation should be canceled to release any allocated resources. If multiple variables are specified, they must be updated in order.
Note that the provider will not modify the fi_trigger_xpu or fi_trigger_var structures after returning from the data transfer call.
In order to support multiple provider implementations, users should trigger data transfer operations in the same order that they are queued and should serialize the writing of triggers that reference the same endpoint. Providers may return the same trigger variable for multiple data transfer requests.
The following feature and description are enhancements to triggered operation support.
The deferred work queue interface is designed as primitive constructs that can be used to implement application-level collective operations. They are a more advanced form of triggered operation. They allow an application to queue operations to a deferred work queue that is associated with the domain. Note that the deferred work queue is a conceptual construct, rather than an implementation requirement. Deferred work requests consist of three main components: an event or condition that must first be met, an operation to perform, and a completion notification.
Because deferred work requests are posted directly to the domain, they can support a broader set of conditions and operations. Deferred work requests are submitted using struct fi_deferred_work. That structure, along with the corresponding operation structures (referenced through the op union) used to describe the work must remain valid until the operation completes or is canceled. The format of the deferred work request is as follows:
struct fi_deferred_work {
struct fi_context2 context;
uint64_t threshold;
struct fid_cntr *triggering_cntr;
struct fid_cntr *completion_cntr;
enum fi_trigger_op op_type;
union {
struct fi_op_msg *msg;
struct fi_op_tagged *tagged;
struct fi_op_rma *rma;
struct fi_op_atomic *atomic;
struct fi_op_fetch_atomic *fetch_atomic;
struct fi_op_compare_atomic *compare_atomic;
struct fi_op_cntr *cntr;
} op;
};
Once a work request has been posted to the deferred work queue, it will remain on the queue until the triggering counter (success plus error counter values) has reached the indicated threshold. If the triggering condition has already been met at the time the work request is queued, the operation will be initiated immediately.
On the completion of a deferred data transfer, the specified completion counter will be incremented by one. Note that deferred counter operations do not update the completion counter; only the counter specified through the fi_op_cntr is modified. The completion_cntr field must be NULL for counter operations.
Because deferred work targets support of collective communication operations, posted work requests do not generate any completions at the endpoint by default. For example, completed operations are not written to the EP’s completion queue or update the EP counter (unless the EP counter is explicitly referenced as the completion_cntr). An application may request EP completions by specifying the FI_COMPLETION flag as part of the operation.
It is the responsibility of the application to detect and handle situations that occur which could result in a deferred work request’s condition not being met. For example, if a work request is dependent upon the successful completion of a data transfer operation, which fails, then the application must cancel the work request.
To submit a deferred work request, applications should use the domain’s fi_control function with command FI_QUEUE_WORK and struct fi_deferred_work as the fi_control arg parameter. To cancel a deferred work request, use fi_control with command FI_CANCEL_WORK and the corresponding struct fi_deferred_work to cancel. The fi_control command FI_FLUSH_WORK will cancel all queued work requests. FI_FLUSH_WORK may be used to flush all work queued to the domain, or may be used to cancel all requests waiting on a specific triggering_cntr.
Deferred work requests are not acted upon by the provider until the associated event has occurred; although, certain validation checks may still occur when a request is submitted. Referenced data buffers are not read or otherwise accessed. But the provider may validate fabric objects, such as endpoints and counters, and that input parameters fall within supported ranges. If a specific request is not supported by the provider, it will fail the operation with -FI_ENOSYS.
fi_getinfo(3), fi_endpoint(3), fi_mr(3), fi_alias(3), fi_cntr(3)
OpenFabrics.
| 2022-12-11 | Libfabric Programmer’s Manual |