Skip to content

Mercury RPC Layer

The RPC layer is used for sending and receiving RPCs. RPC arguments are generally small. For handling larger arguments, this layer is complemented with a bulk interface that is covered in the next section.

HG Interface

The HG interface directly builds on top of the NA interface. Sending an RPC generally results in two NA operations being posted: an expected receive for the response and an unexpected send for the RPC request. The entire interface is therefore non-blocking with a goal of asynchronously executing RPCs on a given target. In order to achieve that, the HG interface provides the following primitives: target address lookup, RPC registration, RPC execution, progress and cancelation.


To initialize the HG RPC interface, two options are available, either by using the default HG_Init() function and specifying an init info string as described in the NA plugin section:

hg_class_t *
HG_Init(const char *info_string, hg_bool_t listen);

Or by using the HG_Init_opt() function, which allows for passing extra options.

Option Description
na_init_info See NA initialization options.
na_class Enables initialization from an existing NA class.
request_post_init Controls the initial number of requests that are posted on context creation.
Default value is: 256
request_post_incr Controls the number of requests that are incrementally posted when the initial number of requests is exhausted.
Default value is: 256
auto_sm Controls whether shared-memory should be automatically used when origin and target share the same node.
sm_info_string Overrides default init string for NA SM plugin.
checksum_level Control checksum level on RPC (none by default).
no_bulk_eager Prevents small bulk data from being automatically embedded along with the RPC request.
no_loopback Disables internal loopback interface that enables forwarding of RPC to self addresses.
stats Print RPC stats at exit.
no_multi_recv Disable use of multi_recv feature (ofi plugin only)
release_input_early Attempt to release input buffers as early as possible once HG_Get_input() has been called.
struct hg_init_info {
    struct na_init_info na_init_info;
    na_class_t *na_class;
    hg_uint32_t request_post_init;
    hg_uint32_t request_post_incr;
    hg_bool_t auto_sm;
    const char *sm_info_string;
    hg_checksum_level_t checksum_level;
    hg_bool_t no_bulk_eager;
    hg_bool_t no_loopback;
    hg_bool_t stats;
    hg_bool_t no_multi_recv;
    hg_bool_t release_input_early;

hg_class_t *
HG_Init_opt2(const char *na_info_string, hg_bool_t na_listen,
             unsigned int version, const struct hg_init_info *hg_init_info);

Similar to the NA layer, the HG_Init() call results in the creation of a new hg_class_t object. The hg_class_t object can later be released after a call to:

HG_Finalize(hg_class_t *hg_class);

Once the interface has been initialized, a context of execution must be created, which (similar to the NA layer) internally associates a specific queue to the operations that will complete:

hg_context_t *
HG_Context_create(hg_class_t *hg_class);

It can then be destroyed using:

HG_Context_destroy(hg_context_t *context);


Before an RPC can be sent, the HG class needs a way of identifying it so that a callback corresponding to that RPC can be executed on the target. Additionally, the functions to serialize and deserialize the function arguments associated to that RPC must be provided. This is done through the HG_Register_name() function. Note that this step can be slightly simplified by using the MERCURY_REGISTER macro. Registration must be done on both the origin and the target with the same func_name identifier. Alternatively HG_Register() can be used to pass a user-defined unique identifier and avoid internal hashing of the provided function name.

typedef hg_return_t (*hg_proc_cb_t)(hg_proc_t proc, void *data);
typedef hg_return_t (*hg_rpc_cb_t)(hg_handle_t handle);

HG_Register_name(hg_class_t *hg_class, const char *func_name,
                 hg_proc_cb_t in_proc_cb, hg_proc_cb_t out_proc_cb,
                 hg_rpc_cb_t rpc_cb);

It is also possible (but not necessary) to deregister an existing RPC ID, in the case where an RPC should no longer be received by using:

HG_Deregister(hg_class_t *hg_class, hg_id_t id);

In the case where an RPC does not require a response, one can indicate that no response is expected (and therefore avoid waiting for a message to be sent back) by using the following call (on an already registered RPC):

HG_Registered_disable_response(hg_class_t *hg_class, hg_id_t id, hg_bool_t disable);

Target Address Lookup

As mentioned in the overview section, there is no real distinction between client and server since it may be desirable for a client to also act as a server for other processes. Therefore, the interface only uses the distinction of origin and target.

Similar to NA's API, the first step consists of retrieving the target's address, this can be done by calling on the target:

HG_Addr_self(hg_class_t *hg_class, hg_addr_t *addr);

And then convert it to string using:

HG_Addr_to_string(hg_class_t *hg_class, char *buf, hg_size_t *buf_size, hg_addr_t addr);

The string can then be exchanged back with the origin through out-of-band mechanisms (e.g., using a file, etc), which can then look the target up using the function:

HG_Addr_lookup(hg_class_t *hg_class, const char *name, hg_addr_t *addr);

All addresses must then be freed using:

HG_Addr_free(hg_class_t *hg_class, hg_addr_t addr);

RPC Execution

Executing an RPC is generally composed of two parts, one on the origin, which will send the RPC request and receive a response, one on the target, which will receive the request, execute it and send a response back.


Using the RPC ID defined after a call to HG_Register(), one can use the HG_Create() call to define a new hg_handle_t object that can be used (and later re-used without reallocating resources) to set/get input/output arguments.

HG_Create(hg_context_t *context, hg_addr_t addr, hg_id_t id, hg_handle_t *handle);

This handle can be destroyed with HG_Destroy()—and a reference count prevents resources from being released while the handle is still in use.

HG_Destroy(hg_handle_t handle);

The second step is to pack the input arguments within a structure, for which a serialization function is provided with the HG_Register() call. The HG_Forward() function can then be used to send that structure (which describes the input arguments). This function is non-blocking. When it completes, the associated callback can be executed by calling HG_Trigger().

typedef hg_return_t (*hg_cb_t)(const struct hg_cb_info *callback_info);

HG_Forward(hg_handle_t handle, hg_cb_t callback, void *arg, void *in_struct);

When HG_Forward() completes (i.e., when the user callback can be triggered), the RPC has been remotely executed and a response with the output results has been sent back. This output can then be retrieved (usually within the callback) with the following function:

HG_Get_output(hg_handle_t handle, void *out_struct);

Retrieving the output may result in the creation of memory objects, which must then be released by calling:

HG_Free_output(hg_handle_t handle, void *out_struct);

To be safe and if necessary, one must make a copy of the results before calling HG_Free_output(). Note that in the case of an RPC with no response, completion occurs after the RPC has been successfully sent (i.e., there is no output to retrieve).


On the target, the RPC callback function passed to the HG_Register() call must be defined.

typedef hg_return_t (*hg_rpc_cb_t)(hg_handle_t handle);

Whenever a new RPC is received, that callback will be invoked. The input arguments can be retrieved with:

HG_Get_input(hg_handle_t handle, void *in_struct);

Retrieving the input may result in the creation of memory objects, which must then be released by calling:

HG_Free_input(hg_handle_t handle, void *in_struct);

When the input has been retrieved, the arguments contained in the input structure can be passed to the actual function call. When the execution is done, an output structure can be filled with the return value and/or the output arguments of the function. It can then be sent back using:

typedef hg_return_t (*hg_cb_t)(const struct hg_cb_info *callback_info);

HG_Respond(hg_handle_t handle, hg_cb_t callback, void *arg, void *out_struct);

This call is also non-blocking. When it completes, the associated callback is placed onto a completion queue. It can then be triggered after a call to HG_Trigger(). Note that in the case of an RPC with no response, calling HG_Respond() will return an error.

Progress and Cancelation

Mercury uses a callback model. Callbacks are passed to non-blocking functions and are pushed to the context's completion queue when the operation completes. Explicit progress is made by calling HG_Progress(). HG_Progress() returns when an operation completes, is in the completion queue or timeout is reached.

HG_Progress(hg_context_t *context, unsigned int timeout);

When an operation completes, calling HG_Trigger() allows the callback execution to be separately controlled from the main progress loop.


Operations may complete with an HG_SUCCESS return code or with an error return code if failure occured after the operation was submitted. The ret field from the hg_cb_info struct should therefore always be checked for potential errors.

HG_Trigger(hg_context_t *context, unsigned int timeout,
           unsigned int max_count, unsigned int *actual_count);

In some cases, one may want to call HG_Progress() then HG_Trigger() or have them execute in parallel by using separate threads.

When it is desirable to cancel an HG operation, one can call HG_Cancel() on a HG handle to cancel an on-going HG_Forward() or HG_Respond() operation. Please refer to this page for additional details regarding cancellation of operations and handling of timeouts.

HG_Cancel(hg_handle_t handle);

Cancelation is always asynchronous. When/if the operation is successfully canceled, it will be pushed to the completion queue and the callback ret value will set with an HG_CANCELED error return code.

Last update: June 15, 2023