sFlow has two components:

• sFlow Agent - should be enabled within the switch to send sFlow datagrams to the collector
• sFlow collector - collects sflow datagrams

switch (config)# protocol sflow
switch (config)# sflow enable
switch (config)# sflow
switch (config sflow)# collector-ip aaa.bbb.ccc.ddd
switch (config sflow)# sampling-rate 4000
switch (config sflow)# counter-poll-interval 5

1. Download sFlow-RT
2. Run command: sflow-rt/get-app.sh sflow-rt flow-trend
3. Restart sFlow-RT

Alternatively, use the Docker image: https://hub.docker.com/r/sflow/flow-trend/

For more information, visit: https://sFlow-RT.com

A window on the local host should be visible after starting the sflow-rt service with flow-trend app downloaded.

On defining the flow specification with the appropriate keys, the flow-trend graph shows only the RoCEv2 traffic.

Further, a filter can be added:

• udpdestinationport = 4791

For more information, visit: https://sflow-rt.com/define_flow.php

PCP is a framework and services to support system-level performance monitoring and performance management. The code can be found in the folder pcp. The metrics can be viewed both on command line and in a graphical way.

For more information regarding implementing a PMDA, visit: https://ryandoyle.net/posts/writing-a-pmda-for-pcp/

A set of bandwidth benchmarks was used to test the monitoring tools. The benchmarks generate a synthetic stream of operations, which is very useful for hardware and software benchmarking and analysis. The benchmarks are not designed to emulate any real application traffic. Real application traffic may be affected by many parameters, and hence might not be predictable based only on the results of those benchmarks.

To test the monitoring tools on the HPC message-passing protocol, the Message Passing Interface (MPI) was used. The Open MPI is an open-source implementation of the Message Passing Interface (MPI). It RDMA through InfiniBand and RoCE protocols. The script used for testing can be found in the folder mpi.