Elastic SIP trunking for predictive dialers: find the limiting resource

A limit must be measured in its own unit

CPS counts new attempts per second; concurrent calls count overlapping calls; Asterisk channels and media allocations are different objects again. Compare each observation with a limit expressed in the same unit and time scope.

The useful answer is the first demonstrated constraint for this workload. It may be the account, a destination, an application worker, media handling, CPU or network. A low monthly average cannot exclude a one-second burst.

You should seeYour worksheet names the workload and the interval before it names a bottleneck.

Read the configured account limits, not an unlimited label

As checked November 6, 2026, Twilio documents an initial termination allowance of 1 CPS and Console increases up to 5 CPS. It describes origination CPS and concurrency as unlimited subject to carrier support, and flags concurrency/self-service restrictions for new accounts without an approved Business Profile. Read the actual account before planning traffic.

For any second provider, obtain the equivalent outbound account/profile limits and enabled destinations. Compare the same direction and account state. An inbound routing feature is not an outbound CPS allowance.

Keep attempts, calls, channels and media separate

Count attempts in the provider’s relevant time buckets and measure overlapping call intervals separately. Retrying the same destination creates another attempt; it is not another person contacted. A call can use multiple local legs and may change its media topology during hold, transfer or recording.

Use actual allocation/traffic telemetry for media. Agent count multiplied by dial level is a planning input, not a count of active RTP streams. Keep rejected attempts and abandoned calls visible when interpreting throughput.

Account termination limit: 2 new attempts/secondObserved attempted calls in the busiest second: 4Observed overlapping calls: 52 of an approved limit of 80Observed allocated media sockets: 104 of a reviewed pool of 2000CPU and packet-loss checks: inside the agreed test thresholdsConclusion: this example exceeds the CPS limit; it does not demonstrate media exhaustion.

Read the configured RTP range without calling it capacity

Inspect the selected RTP configuration and any includes on the actual media host. The following reads a file; it does not prove which values the running process loaded or how many sockets are occupied. Check the installed module/configuration and effective network policy separately.

If a hypothetical policy reserves two ports per stream, 101 available ports permit 50 complete reservations with one spare. That arithmetic assumes the policy; RTCP multiplexing, multiple legs and other media arrangements can change allocation. Do not divide a range and label the result simultaneous calls.

if test -r /etc/asterisk/rtp.conf; then  grep -nE '^[[:space:]]*(rtpstart|rtpend)[[:space:]]*=' /etc/asterisk/rtp.confelse  printf '%s\n' 'Cannot read RTP configuration; ask its owner.' >&2fi

You should seeConfigured range, effective allowed range and observed allocation occupy three separate fields in the worksheet.

Complete the worked worksheet, then replace its evidence

Download the fictional example and blank worksheet. The example isolates a CPS mismatch while leaving other observations within explicitly fictional limits. Its next action is to lower the approved test attempt rate or seek a documented account change, then repeat the same workload.

For your workload, include both media directions, codec/transcoding, transfer/hold/recording behavior, CPU saturation and network loss. A test that omitted one of those behaviors cannot establish its capacity.

1. Record the approved workload, sample interval, destination class and stop conditions.
2. Fill every limit row with units and evidence date. Mark unmeasured rows Unknown.
3. Compare synchronized observations with the limits; retain the first failed and unknown boundaries.
4. Agree one bounded change and repeat the same test before widening the workload.

You should seeThe conclusion names a particular constraint and a repeatable next test, not an agent-count-to-stream assumption.

Fictional capacity observation →

Blank capacity worksheet →

Use call outcomes to correlate, not to infer the resource

Retain response classes and per-interval attempt counts alongside overlapping calls, actual media allocations, CPU and network observations. Correlate their timestamps. A status distribution alone cannot tell whether a rejection came from a provider limit or a local resource.

The optional query below counts rows by VICIdial status in one campaign and a defined time interval. It is neither a live-human answer rate nor a statutory abandonment calculation.

SELECT status, COUNT(*) AS logged_rowsFROM vicidial_logWHERE campaign_id = '<CAMPAIGN_ID>'  AND call_date >= '<START_LOCAL_TIME>'  AND call_date < '<END_LOCAL_TIME>'GROUP BY statusORDER BY logged_rows DESC;

Decide what can change and what remains unproven

An account adjustment, a resource change and a workload change solve different problems. Choose only the one supported by the observation. Re-run answer, audio, hold/transfer, hangup and outcome reconciliation after a relevant change.

If the system did not reach an agreed workload or a needed metric was unavailable, record that limit of the exercise. A provider’s elasticity and a larger port range cannot substitute for this evidence.

You should seeThe team has a scoped capacity decision and the evidence needed to revisit it.

• The limiting row uses matching units and a measured interval.
• No unmeasured stream count or peak was inferred from agents, dial level or daily totals.
• The proposed change has an owner, stop condition and reversal plan.