On 31 Oct 2026, at 2:02, Ilya Maximets wrote:
> On 10/15/26 2:17 PM, Timothy Redaelli via dev wrote:
>> When --format json is passed to ovs-appctl, pmd-perf-show returns a
>> JSON object keyed by thread identifier ("pmd-cNN" for PMD threads,
>> "main" for the main thread). Each entry contains:
>>
>> - "core" (only for PMD threads) and "numa"
>> - "measurement-duration-s" (seconds as real)
>> - "cycles" (used, idle and busy cycles, only for PMD threads)
>> - "iterations" (total, idle, busy and sleep iterations, only for PMD
>> threads)
>> - "sleep" (sleep time, only for PMD threads)
>> - "packets" (received, recirculated, transmitted and tx batches)
>> - "flow-cache-hits" and "upcalls" (only if packets were received)
>> - "averages" (per packet, per batch and per upcall ratios)
>>
>> Percentages name what they are calculated from, the same way the text
>> output does: "used-percentage-of-total" is relative to all cycles in
>> the measurement interval, "idle-percentage-of-used" and
>> "busy-percentage-of-used" to the used cycles, and
>> "sleep-percentage-of-total" to all iterations.
>>
>> Keys are omitted when the value is not available for a thread, rather
>> than being reported as null. Derived values that cannot be calculated
>> because their denominator is zero are omitted as well, instead of being
>> reported as 0. The rates in kpps and the cycles per packet are integers,
>> rounded the same way as in the text output.
>>
>> When the detailed metrics are enabled, the command options select the
>> same additional sections as in text mode: "histograms" unless -nh is
>> given, "iteration-history" for -it and "millisecond-history" for -ms.
>> Each histogram is an object keyed by the upper bound of each bin, with
>> the sample count as value. The last bin, which collects everything
>> above the last bound, is keyed "more". The two histories are arrays
>> with the most recent sample first.
>>
>> The counters are read and the arithmetic is done in a single place:
>> pmd_perf_format_overall_stats_json() builds the JSON object and
>> pmd_perf_format_overall_stats() renders the text output from it, so
>> the two formats cannot drift apart.
>>
>> Example output:
>> {"pmd-c03": {"core": 3, "numa": 0, "measurement-duration-s": 1.0,
>> "cycles": {"used": 9600, "used-percentage-of-total": 0.4,
>> "idle": 9000, "idle-percentage-of-used": 93.75,
>> "busy": 600, "busy-percentage-of-used": 6.25},
>> "iterations": {"total": 100, "idle": 80, "busy": 20,
>> "sleep": 0, "sleep-percentage-of-total": 0,
>> "us-per-iteration": 0.04},
>> "sleep": {"time-us": 0},
>> "packets": {"received": 20, "recirculated": 0,
>> "transmitted": 20, "tx-batches": 20},
>> "flow-cache-hits": {"partial-hardware-offload": 0,
>> "simple": 0, "exact": 19,
>> "signature": 0, "megaflow": 0},
>> "upcalls": {"success": 1, "failure": 0},
>> "averages": {"datapath-passes-per-packet": 1,
>> "rx-kpps": 0, "cycles-per-packet": 30,
>> "us-per-upcall": 2.5, "tx-kpps": 0,
>> "packets-per-tx-batch": 1}}}
>>
>> Signed-off-by: Timothy Redaelli <[email protected]>
>> ---
>> NEWS | 3 +-
>> lib/dpif-netdev-perf.c | 505 ++++++++++++++++++++++++++++++++++++-----
>> lib/dpif-netdev-perf.h | 11 +
>> lib/dpif-netdev.c | 107 +++++++--
>> tests/pmd.at | 172 ++++++++++++++
>> 5 files changed, 724 insertions(+), 74 deletions(-)
>>
>
> [...]
>
> I didn't go over the C code very carefully, as I believe it will need some
> changes anyway, see below.
>
>> diff --git a/tests/pmd.at b/tests/pmd.at
>> index 955be50d1..dfb04ae79 100644
>> --- a/tests/pmd.at
>> +++ b/tests/pmd.at
>> @@ -484,6 +484,178 @@ pmd thread numa_id <cleared> core_id <cleared>:
>> Tx batches: 20 (1.00 pkts/batch)
>> ])
>>
>> +dnl Check JSON output.
>> +dnl Mask dynamic values: core id, measurement duration, all cycle,
>> +dnl iteration and sleep fields, kpps, cycles-per-packet, and us-per-upcall.
>> +AT_CHECK([ovs-appctl --format json --pretty dpif-netdev/pmd-perf-show | dnl
>> + sed -e 's/"pmd-c[[0-9]][[0-9]]"/"pmd-c<cleared>"/' dnl
>> + -e '/"cycles": {/,/}/s/: [[0-9.e+-]][[0-9.e+-]]*/:
>> <cleared>/' dnl
>> + -e '/"iterations": {/,/}/s/: [[0-9.e+-]][[0-9.e+-]]*/:
>> <cleared>/' dnl
>> + -e '/"sleep": {/,/}/s/: [[0-9.e+-]][[0-9.e+-]]*/: <cleared>/'
>> dnl
>> + -e 's/"core": [[0-9]][[0-9]]*/"core": <cleared>/' dnl
>> + -e 's/"measurement-duration-s":
>> [[0-9.e+-]]*/"measurement-duration-s": <cleared>/' dnl
>> + -e 's/"rx-kpps": [[0-9]][[0-9]]*/"rx-kpps": <cleared>/' dnl
>> + -e 's/"tx-kpps": [[0-9]][[0-9]]*/"tx-kpps": <cleared>/' dnl
>> + -e 's/"cycles-per-packet":
>> [[0-9]][[0-9]]*/"cycles-per-packet": <cleared>/' dnl
>> + -e 's/"us-per-upcall": [[0-9.e+-]]*/"us-per-upcall":
>> <cleared>/'], [0], [dnl
>
> As with the previous patches, please, use standard line continuation '\'
> for commands here and in other places in the tests.
>
>> +{
>> + "main": {
>> + "measurement-duration-s": <cleared>,
>> + "packets": {
>> + "received": 0,
>> + "recirculated": 0,
>> + "transmitted": 0,
>> + "tx-batches": 0}},
>> + "pmd-c<cleared>": {
>> + "averages": {
>> + "cycles-per-packet": <cleared>,
>> + "datapath-passes-per-packet": 1,
>> + "packets-per-tx-batch": 1,
>> + "rx-kpps": <cleared>,
>> + "tx-kpps": <cleared>,
>> + "us-per-upcall": <cleared>},
>> + "core": <cleared>,
>> + "cycles": {
>> + "busy": <cleared>,
>> + "busy-percentage-of-used": <cleared>,
>> + "idle": <cleared>,
>> + "idle-percentage-of-used": <cleared>,
>> + "used": <cleared>,
>> + "used-percentage-of-total": <cleared>},
>
> This still doesn't look like a good JSON object. We need to try to encode
> more things structurally. We may also report more things in order to keep
> the output more consistent. In human-readable output we have to balance
> readability with the amount of information we're providing. But for JSON
> we should lean into easier parsing, i.e. more common structure instead of
> custom names. Maybe we could do something like this instead:
>
> "cycles": {
> "count": 123,
> "used": {
> "count": 100,
> "percentage": 80,
> "busy": {
> "count": 80,
> "percentage": 80
> },
> "idle": {
> "count": 20,
> "percentage": 20
> },
> "sleep": {
> "count": 5,
> "percentage": 5
> }
> }
> }
>
> "iterations": {
> "count": 100,
> "busy": {
> "count": 80,
> "percentage": 80
> },
> "idle": {
> "count": 15,
> "percentage": 15
> },
> "sleep": {
> "count": 5,
> "percentage": 5
> }
> }
>
> Here "percentage" is always the percentage of the "count" from the one
> level higher.
>
>> + "flow-cache-hits": {
>> + "exact": 19,
>> + "megaflow": 0,
>> + "partial-hardware-offload": 0,
>> + "signature": 0,
>> + "simple": 0},
>> + "iterations": {
>> + "busy": <cleared>,
>> + "idle": <cleared>,
>> + "sleep": <cleared>,
>> + "sleep-percentage-of-total": <cleared>,
>> + "total": <cleared>,
>> + "us-per-iteration": <cleared>},
>
> "us-per-iteration" should probably go to "averages".
>
>> + "measurement-duration-s": <cleared>,
>> + "numa": 0,
>> + "packets": {
>> + "received": 20,
>> + "recirculated": 0,
>> + "transmitted": 20,
>> + "tx-batches": 20},
>
> "tx-batches" may need to be a standalone field.
>
>> + "sleep": {
>> + "time-us": <cleared>},
>
> This one is also a bit weird. It's not clear why ii is an object
> if it only has one field. It might be good to pir this with the
> measurement duration maybe? They are both time measurements of
> the same thing. So, something like:
>
> "time": {
> "usec": 4232134343,
> "sleep": {
> "usec": 1232434,
> "percentage": 42
> }
> }
>
> Here it maybe makes sense to unify the units to usec, so it's easier
> to relate them to each other. Once again, we don't need to prioritize
> readability here by reporting measurement duration in seconds, an
> application reading the data can convert.
>
> WDYT? Eelco? Kevin?
>
>
> Or maybe an alternative layout with a fully uniform structure:
>
> "pmd-cXX": {
> "cycles": xxx, // total cycles of the measurement duration
> "iterations": xxx, // total iterations
> "usec": xxx, // measurement duration
>
> "used": {
> "cycles": xxx, // used cycles
> "iterations": xxx, // this will always be equal to "iterations" above
> "usec": xxx, // time spent inside iterations
>
> "percentage": {
> "cycles": xxx, // % of total cycles
> "iterations": xxx, // 100
> "usec": xxx // % of total time
> },
>
> "busy": {
> "cycles": xxx, // cycles in busy iterations
> "iterations": xxx, // number of busy iterations
> "usec": xxx, // time spent in busy iterations
>
> "percentage": {
> "cycles": xxx, // % of used.cycles
> "iterations": xxx, // % of used.iterations
> "usec": xxx // % of used.usec
> },
> },
>
> "idle": { // Same for this block
> "cycles": xxx,
> "iterations": xxx,
> "usec": xxx,
>
> "percentage": {
> "cycles": xxx,
> "iterations": xxx,
> "usec": xxx
> },
> }
> },
>
> "sleep": { // Sleep is separate because it is part of both busy and idle
> // and we only want absolute numbers and percentages of total
> // not of 'used'.
> "cycles": xxx, // cycles slept
> "iterations": xxx, // iterations that involved sleep
> "usec": xxx, // time slept
>
> "percentage": {
> "cycles": xxx, // % of total cycles
> "iterations": xxx, // % of total iterations
> "usec": xxx // % of measurement duration
> }
> }
> }
>
> Thoughts?
I like this layout as it's clear who used what.
>> + "upcalls": {
>> + "failure": 0,
>> + "success": 1}}}
>> +])
>> +
>> +dnl Check that the JSON output honours -nh, -it and -ms.
>> +AT_CHECK([ovs-vsctl set Open_vSwitch . other_config:pmd-perf-metrics=true])
>> +AT_CHECK([ovs-appctl time/warp 100], [0], [ignore])
>> +
>> +AT_CHECK([ovs-appctl --format json --pretty dpif-netdev/pmd-perf-show dnl
>> + -it 2 -ms 3 > perf.json])
> Instead of redirect, please capture the output with [stdout] and then
> do AT_CHECK([mv stdout perf.json]), so we have the data in the testsuite
> log as well instead of only in the file.
>
> Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev