Fabtests is a set of examples for fabric providers that demonstrates various features of libfabric- high-performance fabric software library.
Libfabric defines sets of interface that fabric providers can support. The purpose of Fabtests examples is to demonstrate some of the major features. The goal is to familiarize users with different functionalities libfabric offers and how to use them. Although most tests report performance numbers, they are designed to test functionality and not performance. The exception are the benchmarks and ubertest.
The tests are divided into the following categories. Except the unit tests all of them are client-server tests. Not all providers will support each test.
The test names try to indicate the type of functionality each test is verifying. Although some tests work with any endpoint type, many are restricted to verifying a single endpoint type. These tests typically include the endpoint type as part of the test name, such as dgram, msg, or rdm.
These tests are a mix of very basic functionality tests that show major features of libfabric.
Tests communication for connectionless endpoints, as addresses are inserted and removed from the local address vector.
Verifies exchanging CM data as part of connecting endpoints.
Tranfers messages with CQ data.
A basic datagram endpoint example.
Transfers datagrams using waitsets for completion notification.
Sends messages using the FI_INJECT_COMPLETE operation flag.
A simple multicast test.
A basic message endpoint example.
Transfers messages with completion queues configured to use file descriptors as wait objects. The file descriptors are retrieved by the program and used directly with the Linux epoll API.
Verifies that the address assigned to a passive endpoint can be transitioned to an active endpoint. This is required applications that need socket API semantics over RDMA implementations (e.g. rsockets).
Performs data transfers over multiple endpoints in parallel.
Issues RMA write operations to multiple memory regions, using completion counters of inbound writes as the notification mechanism.
Exchanges data over RDM endpoints using poll sets to drive completion notifications.
A basic RDM endpoint example.
Test and verifies atomic operations over an RDM endpoint.
Test triggered operations and deferred work queue support.
Performs data transfers over multiple endpoints, with each endpoint belonging to a different opened domain.
Transfers multiple messages over an RDM endpoint that are received into a single buffer, posted using the FI_MULTI_RECV flag.
An RMA write example over an RDM endpoint that uses RMA events to notify the peer that the RMA transfer has completed.
A basic example of queuing an RMA write operation that is initiated upon the firing of a triggering completion. Works with RDM endpoints.
Spawns child processes to verify basic functionality of using a shared address vector with RDM endpoints.
Basic test of using the FI_PEEK operation flag with tagged messages. Works with RDM endpoints.
Tests canceling posted receives for tagged messages.
Tests the resource management enabled feature. This verifies that the provider prevents applications from overrunning local and remote command queues and completion queues. This corresponds to setting the domain attribute resource_mgmt to FI_RM_ENABLED.
Performs data transfers over scalable endpoints, endpoints associated with multiple transmit and receive contexts.
Performs data transfers between multiple endpoints, where the endpoints share transmit and/or receive contexts.
Tests the send and receive handling of unexpected tagged messages.
Tests data transfers where the transmit buffer is mmapped and unmapped between each transfer, but the virtual address of the transmit buffer tries to remain the same. This test is used to validate the correct behavior of memory registration caches.
Performs a one-sided bandwidth test with an option for data verification. A sleep time on the receiving side can be enabled in order to allow the sender to get ahead of the receiver.
The client and the server exchange messages in either a ping-pong manner, for pingpong named tests, or transfer messages one-way, for bw named tests. These tests can transfer various messages sizes, with controls over which features are used by the test, and report performance numbers. The tests are structured based on the benchmarks provided by OSU MPI. They are not guaranteed to provide the best latency or bandwidth performance numbers a given provider or system may achieve.
Latency test for datagram endpoints
Message transfer bandwidth test for connected (MSG) endpoints.
Message transfer latency test for connected (MSG) endpoints.
Message transfer latency test for reliable-datagram (RDM) endpoints that uses counters as the completion mechanism.
Message transfer latency test for reliable-datagram (RDM) endpoints.
Tagged message bandwidth test for reliable-datagram (RDM) endpoints.
Tagged message latency test for reliable-datagram (RDM) endpoints.
An RMA read and write bandwidth test for reliable (MSG and RDM) endpoints.
These are simple one-sided unit tests that validate basic behavior of the API. Because these are single system tests that do not perform data transfers their testing scope is limited.
Verify address vector interfaces.
Tests counter creation and destruction.
Tests completion queue creation and destruction.
Tests domain creation and destruction.
Tests event queue creation, destruction, and capabilities.
Tests provider response to fi_getinfo calls with varying hints.
Tests memory registration.
Tests provider MR cache eviction capabilities.
This test runs a series of tests over multiple formats and patterns to help validate at scale. The patterns are an all to all, one to all, all to one and a ring. The tests also run across multiple capabilities, such as messages, rma, atomics, and tagged messages. Currently, there is no option to run these capabilities and patterns independently, however the test is short enough to be all run at once.
This is a comprehensive latency, bandwidth, and functionality test that can handle a variety of test configurations. The test is able to run a large number of tests by iterating over a large number of test variables. As a result, a full ubertest run can take a significant amount of time. Because ubertest iterates over input variables, it relies on a test configuration file for control, rather than extensive command line options that are used by other fabtests. A configuration file must be constructed for each provider. Example test configurations are at test_configs.
This test takes a configure file as input. The file contains a list of variables and their values to iterate over. The test will run a set of latency, bandwidth, and functionality tests over a given provider. It will perform one execution for every possible combination of all variables. For example, if there are 8 test variables, with 6 having 2 possible values and 2 having 3 possible values, ubertest will execute 576 total iterations of each test.
Config file options
The following keys and respective key values may be used in the config file.
Identify the provider(s) to test. E.g. udp, tcp, verbs, ofi_rxm;verbs, ofi_rxd;udp.
FT_TEST_LATENCY, FT_TEST_BANDWIDTH, FT_TEST_UNIT
FT_CAP_MSG, FT_CAP_TAGGED, FT_CAP_RMA, FT_CAP_ATOMIC
For FT_CAP_MSG and FT_CAP_TAGGED: FT_FUNC_SEND, FT_FUNC_SENDV, FT_FUNC_SENDMSG, FT_FUNC_INJECT, FT_FUNC_INJECTDATA, FT_FUNC_SENDDATA
For FT_CAP_RMA: FT_FUNC_WRITE, FT_FUNC_WRITEV, FT_FUNC_WRITEMSG, FT_FUNC_WRITEDATA, FT_FUNC_INJECT_WRITE, FT_FUNC_INJECT_WRITEDATA FT_FUNC_READ, FT_FUNC_READV, FT_FUNC_READMSG
For FT_CAP_ATOMIC: FT_FUNC_ATOMIC, FT_FUNC_ATOMICV, FT_FUNC_ATOMICMSG, FT_FUNC_INJECT_ATOMIC, FT_FUNC_FETCH_ATOMIC, FT_FUNC_FETCH_ATOMICV, FT_FUNC_FETCH_ATOMICMSG, FT_FUNC_COMPARE_ATOMIC, FT_FUNC_COMPARE_ATOMICV, FT_FUNC_COMPARE_ATOMICMSG
- constant_caps - values OR'ed together
FI_RMA, FI_MSG, FI_SEND, FI_RECV, FI_READ, FI_WRITE, FI_REMOTE_READ, FI_REMOTE_WRITE, FI_TAGGED, FI_DIRECTED_RECV
- mode - values OR'ed together
FI_EP_MSG, FI_EP_DGRAM, FI_EP_RDM
FT_COMP_QUEUE, FT_COMP_CNTR, FT_COMP_ALL
FI_WAIT_NONE, FI_WAIT_UNSPEC, FI_WAIT_FD, FI_WAIT_MUTEX_COND
FI_WAIT_NONE, FI_WAIT_UNSPEC, FI_WAIT_FD, FI_WAIT_MUTEX_COND
FI_WAIT_NONE, FI_WAIT_UNSPEC, FI_WAIT_FD, FI_WAIT_MUTEX_COND
FI_THREAD_UNSPEC, FI_THREAD_SAFE, FI_THREAD_FID, FI_THREAD_DOMAIN, FI_THREAD_COMPLETION, FI_THREAD_ENDPOINT
FI_PROGRESS_MANUAL, FI_PROGRESS_AUTO, FI_PROGRESS_UNSPEC
(Values OR'ed together) FI_MR_LOCAL, FI_MR_VIRT_ADDR, FI_MR_ALLOCATED, FI_MR_PROV_KEY
For FT_CAP_ATOMIC: FI_MIN, FI_MAX, FI_SUM, FI_PROD, FI_LOR, FI_LAND, FI_BOR, FI_BAND, FI_LXOR, FI_BXOR, FI_ATOMIC_READ, FI_ATOMIC_WRITE, FI_CSWAP, FI_CSWAP_NE, FI_CSWAP_LE, FI_CSWAP_LT, FI_CSWAP_GE, FI_CSWAP_GT, FI_MSWAP
For FT_CAP_ATOMIC: FI_INT8, FI_UINT8, FI_INT16, FI_UINT16, FI_INT32, FI_UINT32, FI_INT64, FI_UINT64, FI_FLOAT, FI_DOUBLE, FI_FLOAT_COMPLEX, FI_DOUBLE_COMPLEX, FI_LONG_DOUBLE, FI_LONG_DOUBLE_COMPLEX
- msg_flags - values OR'ed together
For FT_FUNC_[SEND,WRITE,READ,ATOMIC]MSG: FI_REMOTE_CQ_DATA, FI_COMPLETION
- rx_cq_bind_flags - values OR'ed together
- tx_cq_bind_flags - values OR'ed together
- rx_op_flags - values OR'ed together
- tx_op_flags - values OR'ed together
- test_flags - values OR'ed together
How to Run Tests
Fabtests requires that libfabric be installed on the system, and at least one provider be usable.
Install fabtests on the system. By default all the test executables are installed in /usr/bin directory unless specified otherwise.
All the client-server tests have the following usage model:
fi_ [OPTIONS] start server fi_ connect to server
Command Line Options
Tests share command line options where appropriate. The following command line options are available for one or more test. To see which options apply for a given test, you can use the '-h' help option to see the list available for that test.
Displays help output for the test.
Restrict test to the specified fabric name.
Restrict test to the specified domain name.
Restrict test to the specified provider name.
Use the specified endpoint type for the test. Valid options are msg, dgram, and rdm. The default endpoint type is rdm.
Allocate data buffers on the specified device, rather than in host memory. Valid options are ze and cuda.
: The name of a shared address vector. This option only applies to tests that support shared address vectors.
Specifies the port number of the local endpoint, overriding the default.
Specifies the port number of the peer endpoint, overriding the default.
: Specifies the address of the local endpoint.
Specifies the address format.
Enables out-of-band (via sockets) address exchange and test synchronization. A port for the out-of-band connection may be specified as part of this option to override the default.
Enables out-of-band (via sockets) address exchange only. A port for the out-of-band connection may be specified as part of this option to override the default. Cannot be used together with the '-b' option.
Run fabtests with FI_DELIVERY_COMPLETE.
Number of data transfer iterations.
Associated any EQ with the domain, rather than directly with the EP.
Number of warm-up data transfer iterations.
Data transfer size or 'all' for a full range of sizes. By default a select number of sizes will be tested.
If specified, the starting address of transmit and receive buffers will be aligned along a page boundary.
Use machine readable output. This is useful for post-processing the test output with scripts.
Specify the type of completion mechanism to use. Valid values are queue and counter. The default is to use completion queues.
Indicate the type of processing to use checking for completed operations. Valid values are spin, sread, and fd. The default is to busy wait (spin) until the desired operation has completed. The sread option indicates that the application will invoke a blocking read call in libfabric, such as fi_cq_sread. Fd indicates that the application will retrieve the native operating system wait object (file descriptor) and use either poll() or select() to block until the fd has been signaled, prior to checking for completions.
For RMA based tests, specify the type of RMA operation to perform. Valid values are read, write, and writedata. Write operations are the default.
For multicast tests, specifies the address of the multicast group to join.
Add data verification check to data transfers.
A simple example
run server: <test_name> -p <provider_name> -s <source_addr> e.g. fi_msg_rma -p sockets -s 192.168.0.123 run client: <test_name> <server_addr> -p <provider_name> e.g. fi_msg_rma 192.168.0.123 -p sockets
An example with various options
run server: fi_rdm_atomic -p psm -s 192.168.0.123 -I 1000 -S 1024 run client: fi_rdm_atomic 192.168.0.123 -p psm -I 1000 -S 1024
This will run "fi_rdm_atomic" for all atomic operations with
- PSM provider - 1000 iterations - 1024 bytes message size - server node as 126.96.36.199
Run multinode tests
Server and clients are invoked with the same command: fi_multinode -n <number of processes> -s <server_addr> -C <mode> A process on the server must be started before any of the clients can be started succesfully. -C lists the mode that the tests will run in. Currently the options are
for rma and msg. If not provided, the test will default to msg.
run server: fi_ubertest run client: fi_ubertest -u /usr/share/fabtests/test_configs/sockets/quick.test 192.168.0.123
This will run "fi_ubertest" with
- sockets provider - configurations defined in /usr/share/fabtests/test_configs/sockets/quick.test - server node as 192.168.0.123
The config files are provided in /test_configs for sockets, verbs, udp, and usnic providers and distributed with fabtests installation.
For more usage options: fi_ubertest -h
Run the whole fabtests suite
A runscript scripts/runfabtests.sh is provided that runs all the tests in fabtests and reports the number of pass/fail/notrun.
Usage: runfabtests.sh [OPTIONS] [provider] [host] [client]
By default if none of the options are provided, it runs all the tests using
- sockets provider - 127.0.0.1 as both server and client address - for small number of optiond and iterations
Various options can be used to choose provider, subset tests to run, level of verbosity etc.
runfabtests.sh -vvv -t all psm 192.168.0.123 192.168.0.124
This will run all fabtests using
- psm provider - for different options and larger iterations - server node as 192.168.0.123 and client node as 192.168.0.124 - print test output for all the tests
For detailed usage options: runfabtests.sh -h
The man pages fi_av_test(1), fi_av_xfer(1), fi_bw(1), fi_cm_data(1), fi_cntr_test(1), fi_cq_data(1), fi_cq_test(1), fi_dgram(1), fi_dgram_pingpong(1), fi_dgram_waitset(1), fi_dom_test(1), fi_efa_ep_rnr_retry(1), fi_eq_test(1), fi_getinfo_test(1), fi_inj_complete(1), fi_mcast(1), fi_mr_test(1), fi_msg(1), fi_msg_bw(1), fi_msg_epoll(1), fi_msg_pingpong(1), fi_msg_sockets(1), fi_multi_ep(1), fi_multi_mr(1), fi_multi_recv(1), fi_poll(1), fi_rdm(1), fi_rdm_atomic(1), fi_rdm_cntr_pingpong(1), fi_rdm_deferred_wq(1), fi_rdm_multi_client(1), fi_rdm_multi_domain(1), fi_rdm_pingpong(1), fi_rdm_rma_event(1), fi_rdm_rma_trigger(1), fi_rdm_shared_av(1), fi_rdm_stress(1), fi_rdm_tagged_bw(1), fi_rdm_tagged_peek(1), fi_rdm_tagged_pingpong(1), fi_recv_cancel(1), fi_resmgmt_test(1), fi_rma_bw(1), fi_scalable_ep(1), fi_shared_ctx(1), fi_ubertest(1), fi_unexpected_msg(1) and fi_unmap_mem(1) are aliases of fabtests(7).