vllmbench2pcp - Man Page

vllmbench2pcp [-hv] [-a archive] [-H hostname] results [...]

vllmbench2pcp is a performance metrics importer tool that creates a PCP archive from the JSON results of a vLLM benchmark.

An archive spanning the time for which the benchmark and its run iterations were active. Discrete PCP metrics are created for each recorded aspect of a run, and logged once at the start and once at the end of each iteration, followed by a mark record.

This allows PCP tools to report on these metrics in a meaningful way alongside other system (kernel, hardware, vLLM, and so on) metrics gathered during the benchmark.

The only mandatory command line argument is the JSON results file from a vllm bench invocation. This JSON can optionally be wrapped with additional Model Furnace metadata, and if this is detected additional PCP labels will be extracted and stored (such as model and inference server names).

The available command line options are:

-a archive, --archive=archive

Specified the name of the output PCP archive to be produced. By default, an archive name using the timestamp of the start of the benchmark run will be used.

-H hostname, --hostname=hostname

Sets the hostname that will be stamped into the PCP archive label header.

-h,  --help

Display usage message and exit.

-v,  --verbose

Report progress as each run is identified and its values saved.

PCPIntro(1), pminfo(1), pmrep(1), LOGARCHIVE(5), and PMNS(5).