DOKK / manpages / debian 12 / libfabric-dev / fabric.7.en
fabric(7) #VERSION# fabric(7)

fabric - Fabric Interface Library

#include <rdma/fabric.h>

Libfabric is a high-performance fabric software library designed to provide low-latency interfaces to fabric hardware.

Libfabric provides `process direct I/O' to application software communicating across fabric software and hardware. Process direct I/O, historically referred to as RDMA, allows an application to directly access network resources without operating system interventions. Data transfers can occur directly to and from application memory.

There are two components to the libfabric software:

Conceptually, a fabric provider may be viewed as a local hardware NIC driver, though a provider is not limited by this definition. The first component of libfabric is a general purpose framework that is capable of handling different types of fabric hardware. All fabric hardware devices and their software drivers are required to support this framework. Devices and the drivers that plug into the libfabric framework are referred to as fabric providers, or simply providers. Provider details may be found in fi_provider(7).
The second component is a set of communication operations. Libfabric defines several sets of communication functions that providers can support. It is not required that providers implement all the interfaces that are defined; however, providers clearly indicate which interfaces they do support.

The fabric interfaces are designed such that they are cohesive and not simply a union of disjoint interfaces. The interfaces are logically divided into two groups: control interfaces and communication operations. The control interfaces are a common set of operations that provide access to local communication resources, such as address vectors and event queues. The communication operations expose particular models of communication and fabric functionality, such as message queues, remote memory access, and atomic operations. Communication operations are associated with fabric endpoints.

Applications will typically use the control interfaces to discover local capabilities and allocate necessary resources. They will then allocate and configure a communication endpoint to send and receive data, or perform other types of data transfers, with remote endpoints.

The control interfaces APIs provide applications access to network resources. This involves listing all the interfaces available, obtaining the capabilities of the interfaces and opening a provider.

The fi_getinfo call is the base call used to discover and request fabric services offered by the system. Applications can use this call to indicate the type of communication that they desire. The results from fi_getinfo, fi_info, are used to reserve and configure fabric resources.

fi_getinfo returns a list of fi_info structures. Each structure references a single fabric provider, indicating the interfaces that the provider supports, along with a named set of resources. A fabric provider may include multiple fi_info structures in the returned list.

A fabric domain represents a collection of hardware and software resources that access a single physical or virtual network. All network ports on a system that can communicate with each other through the fabric belong to the same fabric domain. A fabric domain shares network addresses and can span multiple providers. libfabric supports systems connected to multiple fabrics.
An access domain represents a single logical connection into a fabric. It may map to a single physical or virtual NIC or a port. An access domain defines the boundary across which fabric resources may be associated. Each access domain belongs to a single fabric domain.
A fabric endpoint is a communication portal. An endpoint may be either active or passive. Passive endpoints are used to listen for connection requests. Active endpoints can perform data transfers. Endpoints are configured with specific communication capabilities and data transfer interfaces.
Event queues, are used to collect and report the completion of asynchronous operations and events. Event queues report events that are not directly associated with data transfer operations.
Completion queues are high-performance event queues used to report the completion of data transfer operations.
Event counters are used to report the number of completed asynchronous operations. Event counters are considered light-weight, in that a completion simply increments a counter, rather than placing an entry into an event queue.
Memory regions describe application local memory buffers. In order for fabric resources to access application memory, the application must first grant permission to the fabric provider by constructing a memory region. Memory regions are required for specific types of data transfer operations, such as RMA transfers (see below).
Address vectors are used to map higher level addresses, such as IP addresses, which may be more natural for an application to use, into fabric specific addresses. The use of address vectors allows providers to reduce the amount of memory required to maintain large address look-up tables, and eliminate expensive address resolution and look-up methods during data transfer operations.

Fabric endpoints are associated with multiple data transfer interfaces. Each interface set is designed to support a specific style of communication, with an endpoint allowing the different interfaces to be used in conjunction. The following data transfer interfaces are defined by libfabric.

Message queues expose a simple, message-based FIFO queue interface to the application. Message data transfers allow applications to send and receive data with message boundaries being maintained.
Tagged message lists expose send/receive data transfer operations built on the concept of tagged messaging. The tagged message queue is conceptually similar to standard message queues, but with the addition of 64-bit tags for each message. Sent messages are matched with receive buffers that are tagged with a similar value.
RMA transfers are one-sided operations that read or write data directly to a remote memory region. Other than defining the appropriate memory region, RMA operations do not require interaction at the target side for the data transfer to complete.
Atomic operations can perform one of several operations on a remote memory region. Atomic operations include well-known functionality, such as atomic-add and compare-and-swap, plus several other pre-defined calls. Unlike other data transfer interfaces, atomic operations are aware of the data formatting at the target memory region.

Logging can be controlled using the FI_LOG_LEVEL, FI_LOG_PROV, and FI_LOG_SUBSYS environment variables.

FI_LOG_LEVEL controls the amount of logging data that is output. The following log levels are defined.
Warn is the least verbose setting and is intended for reporting errors or warnings.
Trace is more verbose and is meant to include non-detailed output helpful to tracing program execution.
Info is high traffic and meant for detailed output.
Debug is high traffic and is likely to impact application performance. Debug output is only available if the library has been compiled with debugging enabled.
The FI_LOG_PROV environment variable enables or disables logging from specific providers. Providers can be enabled by listing them in a comma separated fashion. If the list begins with the `^' symbol, then the list will be negated. By default all providers are enabled.

Example: To enable logging from the psm and sockets provider: FI_LOG_PROV=“psm,sockets”

Example: To enable logging from providers other than psm: FI_LOG_PROV=“^psm”

The FI_LOG_SUBSYS environment variable enables or disables logging at the subsystem level. The syntax for enabling or disabling subsystems is similar to that used for FI_LOG_PROV. The following subsystems are defined.
Provides output related to the core framework and its management of providers.
Provides output specific to interactions associated with the fabric object.
Provides output specific to interactions associated with the domain object.
Provides output specific to endpoint non-data transfer operations, such as CM operations.
Provides output specific to endpoint data transfer operations.
Provides output specific to address vector operations.
Provides output specific to completion queue operations.
Provides output specific to event queue operations.
Provides output specific to memory registration.

The libfabric build scripts will install all providers that are supported by the installation system. Providers that are missing build prerequisites will be disabled. Installed providers will dynamically check for necessary hardware on library initialization and respond appropriately to application queries.

Users can enable or disable available providers through build configuration options. See `configure –help' for details. In general, a specific provider can be controlled using the configure option `–enable-'. For example, `–enable-udp' (or `–enable-udp=yes') will add the udp provider to the build. To disable the provider, `–enable-udp=no' can be used.

Providers can also be enable or disabled at run time using the FI_PROVIDER environment variable. The FI_PROVIDER variable is set to a comma separated list of providers to include. If the list begins with the `^' symbol, then the list will be negated.

Example: To enable the udp and tcp providers only, set: FI_PROVIDER=“udp,tcp”

The fi_info utility, which is included as part of the libfabric package, can be used to retrieve information about which providers are available in the system. Additionally, it can retrieve a list of all environment variables that may be used to configure libfabric and each provider. See fi_info(1) for more details.

Core features of libfabric and its providers may be configured by an administrator through the use of environment variables. Man pages will usually describe the most commonly accessed variables, such as those mentioned above. However, libfabric defines interfaces for publishing and obtaining environment variables. These are targeted for providers, but allow applications and users to obtain the full list of variables that may be set, along with a brief description of their use.

A full list of variables available may be obtained by running the fi_info application, with the -e or –env command line option.

Because libfabric is designed to provide applications direct access to fabric hardware, there are limits on how libfabric resources may be used in conjunction with system calls. These limitations are notable for developers who may be familiar programming to the sockets interface. Although limits are provider specific, the following restrictions apply to many providers and should be adhered to by applications desiring portability across providers.

Fabric resources are not guaranteed to be available by child processes. This includes objects, such as endpoints and completion queues, as well as application controlled data buffers which have been assigned to the network. For example, data buffers that have been registered with a fabric domain may not be available in a child process because of copy on write restrictions.

In some cases, calls to cudaMemcpy within libfabric may result in a deadlock. This typically occurs when a CUDA kernel blocks until a cudaMemcpy on the host completes. To avoid this deadlock, cudaMemcpy may be disabled by setting FI_HMEM_CUDA_ENABLE_XFER=0. If this environment variable is set and there is a call to cudaMemcpy with libfabric, a warning will be emitted and no copy will occur. Note that not all providers support this option.

Another mechanism which can be used to avoid deadlock is Nvidia’s gdrcopy. Using gdrcopy requires an external library and kernel module available at https://github.com/NVIDIA/gdrcopy. Libfabric must be configured with gdrcopy support using the --with-gdrcopy option, and be run with FI_HMEM_CUDA_USE_GDRCOPY=1. This may be used in conjunction with the above option to provide a method for copying to/from CUDA device memory when cudaMemcpy cannot be used. Again, this may not be supported by all providers.

libfabric releases maintain compatibility with older releases, so that compiled applications can continue to work as-is, and previously written applications will compile against newer versions of the library without needing source code changes. The changes below describe ABI updates that have occurred and which libfabric release corresponds to the changes.

Note that because most functions called by applications actually call static inline functions, which in turn reference function pointers in order to call directly into providers, libfabric only exports a handful of functions directly. ABI changes are limited to those functions, most notably the fi_getinfo call and its returned attribute structures.

The ABI version is independent from the libfabric release version.

The initial libfabric release (1.0.0) also corresponds to ABI version 1.0. The 1.0 ABI was unchanged for libfabric major.minor versions 1.0, 1.1, 1.2, 1.3, and 1.4.

A number of external data structures were appended starting with libfabric version 1.5. These changes included adding the fields to the following data structures. The 1.1 ABI was exported by libfabric versions 1.5 and 1.6.

Added api_version
Added cntr_cnt, mr_iov_limit, caps, mode, auth_key, auth_key_size, max_err_data, and mr_cnt fields. The mr_mode field was also changed from an enum to an integer flag field.
Added auth_key_size and auth_key fields.

The 1.2 ABI version was exported by libfabric versions 1.7 and 1.8, and expanded the following structure.

The fi_info structure was expanded to reference a new fabric object, fid_nic. When available, the fid_nic references a new set of attributes related to network hardware details.

The 1.3 ABI version was exported by libfabric versions 1.9, 1.10, and 1.11. Added new fields to the following attributes:

Added tclass
Added tclass

The 1.4 ABI version was exported by libfabric 1.12. Added fi_tostr_r, a thread-safe (re-entrant) version of fi_tostr.

ABI version starting with libfabric 1.13. Added new fi_open API call.

ABI version starting with libfabric 1.14. Added fi_log_ready for providers.

fi_info(1), fi_provider(7), fi_getinfo(3), fi_endpoint(3), fi_domain(3), fi_av(3), fi_eq(3), fi_cq(3), fi_cntr(3), fi_mr(3)

OpenFabrics.

2022-12-11 Libfabric Programmer’s Manual