NAME¶

cmtime - RDMA CM connection steps timing test.

SYNOPSIS¶

cmtime -C controller_address
		(-B bind_interface | -b bind_address)
		[-L]
		[-c connections_per_pair] [-P num_peers]
		[-p controller_port]
		[-q base_qpn] [-n num_threads]
		[-m mimic_qp_delay_us] [-r retries]
		[-t timeout_ms]
		[-S]

DESCRIPTION¶

Determines min, max, and average times for various "steps" in RDMA CM connection setup and teardown between clients and servers.

To use, start one or more servers (-L), plus one or more clients. Each client will establish -c number of connections to each server. By default, the test runs with 1 client and 1 server.

One process will act as a controller process. The controller coordinates the clients and servers -- uses out-of-band sockets to collect and distribute addresses and keeps clients and servers in sync as they advance through each "step".

"Steps" that are timed are: create ID, bind address, resolve address, resolve route, create QP, modify QP to INIT, modify QP to RTR, modify QP to RTS, CM connect, client establish, disconnect, destroy QP, and destroy ID.

Many operations are asynchronous, allowing progress on multiple connections simultanesously. The 'sum' output adds the time that all connections took for a given step. The average 'us/conn' is the sum divided by the number of connections. This is useful to identify steps which take a significant amount of time. The min and max values are the smallest and largest times that any single connection took to complete a given step.

The 'total' and 'avg/iter' times measure the time to complete a given step for all connections. These two values take into account asynchronous operations. For steps which are serial, the total and sum values will be roughly the same. For asynchronous steps, the total may be significantly lower than the sum, as multiple connections will be in progress simultanesously. The avg/iter is the total time divided by the number of connections.

In many cases, times may not be available or only available on the clients. Is such situations, the output will show 0. Stats printed by each instance pertain to connections handled by it -- on runs with multiple clients and/or servers (-P > 2) further external aggregation might be required.

OPTIONS¶

-C controller_address

The network name or IP address of the instance which will provide step synchronization. This option must be specified by all instances. The first instance to start which discovers this address as local will act as controller and will synchronize all other instances via an out-of-band channel as benchmark steps progress. Any type of instance (client or server) can act as controller.

-B bind_interface

The local RDMA interface to bind to. Only one of -B or -b is accepted.

-b bind_address

The local network address to bind to. Only one of -B or -b is accepted.

-L

Whether this instance is an RDMA CM server (set) or client (unset). (default client/unset) -c connections_per_pair The number of connections to establish between each client-server pair. (default 100)

-P num_peers

Total number of peers (clients, servers) which are going to be started. (default 2)

-p controller_port

The controller's port number.

-q base_qpn

The first QP number to use when creating connections without allocating hardware QPs. The test will use the values between base_qpn and base_qpn plus connections when connecting. (default 1000)

-n num_threads

Sets the number of threads to spawn used to process connection events and hardware operations. (default 1)

-m mimic_qp_delay_us

"Simulates" QP creation and modify calls by replacing them with a simple sleep function instead. This allows testing the CM at larger scale than would be practical, or even possible given system configuration settings, if HW resources needed to be allocated.

-r retries

Number of retries when resolving address or route. (default 2)

-S

Run connection rate test using sockets. This provides a baseline comparison of RDMA connections versus TCP connections. Sockets are set to blocking mode. Only supported in 1 client 1 server mode (-P 2).

-t timeout_ms

Timeout in millseconds (ms) when resolving address or route. (default 2000 - 2 seconds)

EXAMPLES¶

One client connecting to one server:

server$ cmtime -B enp1s0f0np0 -C 192.0.2.1 -L
client$ cmtime -B enp1s0f0np0 -C 192.0.2.1

Two clients connecting to three servers, 6k connections in total:

client1$ cmtime -B enp1s0f0np0 -P 5 -C 192.0.2.1 -c 1000
client2$ cmtime -B enp1s0f0np0 -P 5 -C 192.0.2.1 -c 1000
server1$ cmtime -B enp1s0f0np0 -P 5 -C 192.0.2.1 -c 1000 -L
server2$ cmtime -B enp1s0f0np0 -P 5 -C 192.0.2.1 -c 1000 -L
server3$ cmtime -B enp1s0f0np0 -P 5 -C 192.0.2.1 -c 1000 -L

NOTES¶

Because this test maps RDMA resources to userspace, users must ensure that they have available system resources and permissions. See the libibverbs README file for additional details.

SEE ALSO¶

rdma_cm(7)