URL: http://vblaze_name:17171/stats

Example Response:

{
    "stats": {
        "idletotal": [60.0,300.0,900.0,916225.595],
        "load.am": [0.0,0.0,0.0],
        "load.lm": [0.0,0.0,0.0],
        "load.stream": [0.0,0.0,0.003],
        "nsubs": [0,0,0,0],
        "stream.audiosecs": [0,0,0,3204034.2],
        "stream.count": [0,0,0,10819],
        "stream.errors": [0,0,0,0],
        "stream.latency": [0,0,0,0.169],
        "stream.rate": [0.0,0.0,0.0,3.457],
        "stream.speed": [0,0,0,299.706],
        "stream.tat": [0,0,0,47.428],
        "stream.tat.max": [0,0,0,133.787],
        "stream.warnings": [0,0,0,0],
        "utilization.am.all": [0.0,0.0,0.0],
        "utilization.am": [0.0,0.0,0.0],
        "utilization.gpu": [0.0,0.0,0.0],
        "utilization.gpu.memory": [0.0,0.0,0.0],
        "utilization.lm": [0.0,0.0,0.0],
        "utilization.stream": [0.0,0.0,0.005],
        "utt.audiosecs": [0,0,0,3035470.61],
        "utt.count": [0,0,0,209908],
        "utt.latency": [0,0,0,0.927],
        "utt.rate": [0.0,0.0,0.0,3.275],
        "utt.speed": [0,0,0,284.02],
        "utt.tat": [0,0,0,40.652],
        "utt.tat.max": [0,0,0,133.552]
    }
}

Explanation:

The example response above is a JSON object that shows status information about the V‑Blaze engine. The stats endpoint consists of a number of fields each of which has an array of 3 or 4 elements as a value. Each entry in the array corresponds to the measurement of the statistic during the last 1 minute, 5 minute, 15 minute, and all time windows respectively. Note that statistics with a 3 element array do not have an all time window element.

The following list explains each key and value returned by a /stats query:

{ 
 "advancedstats": true 
}

idletotal

Indicates the total idle time, or time when no streams were active, measured in seconds.

load.am

Indicates the average number of utterances being either actively decoded by an acoustic model process or waiting for an acoustic model process to become available. This is synonymous with UNIX load but for for acoustic model processes.

load.lm

Indicates the average number of utterances being either actively decoded by a language model process or waiting for a language model process to become available. This is synonymous with UNIX load but for language model processes.

load.stream

Indicates the average number of streams being either actively decoded by a decoder process or waiting for a decoder process to become available. This is synonymous with UNIX load but for decoder processes.

nsubs

Indicates the number of substitutions performed.

stream.audiosecs

Indicates the total audio processed measured in seconds of audio.

stream.count

Indicates the total number of streams processed.

stream.errors

Indicates the total number of stream errors.

stream.latency

Indicates the average stream latency. The latency of a stream is equal to the number of seconds taken to process the stream divided by the number of audio seconds within the stream.

stream.rate

Indicates the average rate of audio submission measured in audio seconds per second.

stream.speed

Indicates the average rate of audio decoding measured in audio seconds per second.

Note the difference between stream.speed and stream.rate : if only one stream consisting of 60 seconds of audio is submitted within the last minute and decoded in 1 second, then the 1 minute window value of stream.rate will be 1.0, corresponding to 1 second of audio submitted per second on average. The 1 minute window value of stream.speed will be 60.0, corresponding to 60 seconds of audio decoded per second on average.

stream.tat

Indicates the average stream turnaround time since V‑Blaze started, measured in seconds.

stream.tat.max

Indicates the maximum stream turnaround time since V‑Blaze started, measured in seconds.

stream.warnings

Indicates the total number of stream warnings.

utilization.am.all

Indicates the acoustic model process utilization percentage, which is between 0 and 100. This is calculated as the average percentage of AM processes which are actively decoding an utterance.

utilization.am

Indicates the acoustic model process utilization percentage, which is between 0 and 100, only for processes within the set of active acoustic model.

AM processes, unlike LM processes, are statically assigned a single acoustic model; therefore, on machines with a large number of acoustic models the number of idle AM processes is expected to be large as typical workloads rarely utilize more than one or two acoustic models simultaneously. As such, values seen in utilization.am.all may not be useful since inactive acoustic models are included in the calculation.

utilization.gpu

Indicates the average value of the query: nvidia-smi --query-gpu=utilization.gpu --format=csv

utilization.gpu.memory

The average value of the query: nvidia-smi --query-gpu=utilization.memory --format=csv

utilization.lm

Indicates the language model process utilization percentage, which is a value between 0 and 100. This is calculated as the average percentage of LM processes which are actively decoding an utterance.

utilization.stream

Indicates the decoder process utilization percentage, which is a value between 0 and 100. This is calculated as the average percentage of decoder process which are actively decoding a stream.

utt.audiosecs

Indicates the total utterance audio processed measured in seconds.

utt.count

Indicates the total number of utterances processed.

utt.latency

Indicates the average utterance latency. The latency of an utterance is equal to the number of seconds taken to process the utterance divided by the number of audio seconds within the stream

utt.rate

Indicates the average rate of utterance audio submission measured in audio seconds per second.

utt.speed

Indicates the average rate of utterance audio decoding measured in audio seconds per second. The same distinction made between stream.rate and stream.speed is valid for utt.rate and utt.speed .

utt.tat

Indicates the average turnaround time for utterances in seconds.

utt.tat.max

Indicates the maximum turnaround time for utterances in audio seconds.