VICIfast
Guides & tutorials

VICIdial clusters: how they work, when you need one, and how to size one

How a VICIdial cluster splits the work across machines, the signs one server is no longer enough, and how to count dialers and size the database from measured numbers.

VICIfast Support
··12 min read
VICIdial clusters: how they work, when you need one, and how to size one

Your floor is growing, and the one server that runs it is starting to show the strain. Dispositions save late, agent screens hang for a second, and someone has told you the answer is "a cluster".

Before you buy more machines, it helps to know what a cluster actually is, whether you need one yet, and how many machines it takes. The numbers below come from live floors we measured, not from rules of thumb.

How a VICIdial cluster works

A VICIdial cluster is one VICIdial system spread over several servers. Your agents log in the same way, and each Campaign (a set of leads and dialing rules) runs the same way it did on one server. What changes is which machine does which job.

In our layout there are three kinds of machine, plus the network that joins them.

The database and web server

This machine runs MariaDB, the database that holds every lead, campaign and call record. In our layout the same machine also serves the VICIdial admin pages, the agent screens and the reports. When an agent opens their screen, their browser loads it from here.

Every other machine depends on it. Dialers query it many times a second, so its health sets the pace for the whole floor. For more on the split, see running the VICIdial database on its own server.

The dialer servers

Dialers run Asterisk, the phone engine underneath VICIdial. They place and answer the calls, host the agents' phones and hold the conferences, including every Three-way call where an agent brings a third person onto the line. Your carriers' trunks connect here, and so do your agents' phones.

We run one dialer per 300 concurrent calls or 25 agents, whichever needs more, and never fewer than two.

The standby database

The standby copies the database as it changes, so it is always a live copy. If the database machine fails, the standby is promoted and takes over. Without it, the database is the one machine whose loss stops everything.

One datacenter, one private network

Every machine in the cluster sits in the same datacenter on a private network. That is not a preference. Dialers query the database many times a second, and a round trip between Germany and Finland is over 20 ms, which is enough to stall VICIdial.

Carriers and agents connect to the dialers. Database queries cross the private network every second. Here is how the pieces fit:

flowchart LR
  T[Your SIP trunks] --> D1[Dialer 1]
  T --> D2[Dialer 2]
  P[Agent phones] --> D1
  P --> D2
  B[Agent browsers] --> W[(Database + web)]
  D1 -- private network --> W
  D2 -- private network --> W
  W -. replicates .-> S[(Standby database)]

When one server stops being enough

Most floors start on one server and grow into a cluster. The hard part is knowing when.

By our estimates one machine carries about 600 Concurrent calls with room to spare. Beyond that, calls belong on dialers of their own. But call count is only one reason to split, and the other two matter just as much.

First, on one box Asterisk and MariaDB share the same cores. A busy report or a big list load slows the calls, because the database and the phone engine compete for the same processor. A cluster gives each its own machine.

Second, one box is one failure. If it stops, the whole floor stops. With two or more dialers and a standby database, a failed machine stops part of the floor, not all of it.

If you are still reading your current server's limits, start with how many agents one VICIdial server can handle and the signs your server needs more capacity.

What one measured floor actually used

On 14 October 2026 we measured a full shift on an outbound floor running on one of our 16-vCPU servers. Across 92 five-minute samples with calls up, it ran 166 concurrent calls with 17 agents and 365 channels, the individual call legs Asterisk keeps open.

That floor used about 7.5 cores in total. Asterisk took 4.3 of them and MariaDB 1.3. The database answered 2,230 queries a second.

Asterisk does most of the work, but the database grows with every call too. From that shift we built our per-machine estimate: about 2 cores of fixed load plus 0.03 cores per concurrent call, adjusted for each processor.

How to size a VICIdial cluster

Sizing comes down to three questions: how many dialers, how big a database, and what the machines cost.

Count the dialers

Use one dialer per 300 concurrent calls or 25 agents, whichever needs more, and never fewer than two. Work out both numbers and take the larger one.

A floor with 50 agents and 600 calls needs two dialers by either rule. At 75 agents and 900 calls it needs three, and at 125 agents and 1,500 calls, five.

The floor of two is about failure, not load. If a dialer fails, the calls on it drop and the other dialers keep the floor working. With one dialer, that same failure stops every call.

Size the database for locks, not cores

On 15 October 2026 we replayed five minutes of that floor's real database traffic on a 16-core AMD Ryzen 9 7950X3D with 128 GB and mirrored NVMe disks. We played it faster and faster, up to seven times the live load.

SpeedAbout this many callsQueries a secondDB cores avg / peak99% of queries99.9% of queries
1x1802,1550.4 / 1.30.45 ms18 ms
3x5406,4641.1 / 2.10.6 ms24 ms
5x90010,7741.6 / 2.50.69 ms208 ms
7x1,26015,0832.1 / 3.10.77 ms290 ms

The database never needed more than 3.1 of its cores. Even at about 1,260 calls, 99% of queries finished in 0.77 ms.

What grew was the slow tail: the 99.9% mark went from 18 ms at live speed to 208 ms at 5x and 290 ms at 7x. That came from table locks on VICIdial's call log, not the processor. At this size the fix is keeping that table small with VICIdial's log archiving, not a bigger database machine.

The machines we use

Our database + web server and the standby are the same build: AMD Ryzen 9 7950X3D, 16 cores / 32 threads, 128 GB, and 2 x 1.92 TB NVMe, mirrored. Each dialer is an Intel Core i5-13500 with 14 cores (6P + 8E), 64 GB and 2 x 512 GB NVMe.

What the hardware costs

At November 2026 prices, hardware only, with the standby database included:

FloorDialersMachinesHardware a month
600 calls / 50 agents24$1,376
900 calls / 75 agents35$1,515
1,500 calls / 125 agents57$1,793

There is also a one-time setup fee per machine; on these examples it comes to $298 in total. Building and running the cluster is quoted on a sizing call, and prices change, so check the live sizer on the VICIdial cluster page for today's numbers. We also build clusters on cloud plans; for that trade-off, read cloud cost vs bare metal for VICIdial, and for the wider money picture, VICIdial pricing and true cost in 2026.

Problems that look like capacity, and what fixes them

Many floors that think they need more hardware have a different problem. The symptoms look like an overloaded server, but the cause is a lock, a missing index or a background job nobody looked at. These are the ones we measured on live floors.

The database is waiting, not busy

VICIdial keeps call records in MyISAM tables, which lock the whole table while a read runs. A report, stats tool or lead search that reads the call log for a few seconds makes every Disposition (the result an agent picks for a call) wait behind it. On a 60-agent floor, one day-long count held up to 334 statements for as long as 4 seconds.

The same locks cause drops. VICIdial hands each answered call to a free agent through the database. When the tables it needs are locked, the hand-off waits but the drop timer keeps counting. On one floor, taking a stats tool off the call log cut lock waits from 58 a second to 13.

If your Drop rate climbs while agents sit ready, read the common causes of dropped calls and how to lower your VICIdial drop rate. Other database costs we found:

  • A missing phone-number index. VICIdial's call log lacked one, so a lead search by phone read the whole log. On one floor that meant 4–12 second dialer freezes. Every cluster we build gets that index.
  • Idle campaigns left active. With large hoppers, they were 41% of all database time on one floor. Deactivate campaigns nobody dials and keep each Hopper, the queue of leads waiting to be dialed, sized to the floor.
  • A CRM polling the API once per agent. On one floor that was 25–39% of web requests. One call for all logged-in agents, cached and polled less often, does the same job.
  • The no-answer log scan. A stock VICIdial job that scans the call log for unanswered calls took 8–27% of database time on two floors. An index cut that cost by about 95% in our test.
  • Log tables never archived. On one floor they had grown to 7–12.5 million rows each. VICIdial ships a log-archiving script, ADMIN_archive_log_tables.pl, that moves old rows out. Run it on a schedule.
  • The midnight freeze. A nightly job that archives the agent latency log froze one floor for 17 seconds at midnight.

When the hopper runs dry, agents just sit idle and nothing on their screen says why. How to fix an empty hopper walks through the usual causes.

Agent screens that stop moving

The agent screen asks the server for news every second. On one floor, agent-screen polling was 87.5% of all web requests. When the database or Asterisk's manager interface (AMI (Asterisk Manager Interface)) backs up, those requests queue and the screen stops updating. What causes latency gaps on the agent screen shows how to read that symptom.

Asterisk 21 and 22 add one more cause. They leak a manager subscription on every dialplan reload while the old users.conf file is present. After days of reloads that can trip Asterisk's overload protection, which turns away new calls and phone registrations. We have fixed it on every server we run.

Problems on the dialers themselves

  • Choppy audio. Linux's default network buffers are sized for a desktop. On a busy dialer they overflow in bursts and voice packets are thrown away: one 60-agent server dropped 458,000 of them in four days. See diagnosing choppy or robotic audio.
  • Customers connected to silence. When an agent's Webphone, the phone that runs in their browser, loses its connection (a closed tab, a sleeping laptop, a new IP address), Asterisk keeps its side of the call up. VICIdial still sees a ready agent and keeps sending customers to a line nobody hears. Every agent phone and carrier trunk we set up hangs up after two minutes without audio.
  • Leave 3-way fails. VICIdial frees conference rooms with a background process that is often left switched off. Without it, every room used for a 3-way call stays taken until the dialer has none left.
  • Too few cores. One floor ran 18 agents at dial level 4–8 on 4 cores and saw 25% CPU pressure. On 8 cores it fell to 6%. Safe CPU and load thresholds explains how to read yours.
Check which drop rate you are looking at. Some dashboards divide dropped calls by every number dialed, while VICIdial's report divides by calls a person answered, which is what drop limits are measured against. One campaign showed 1.23% on a dashboard and 28.89% in VICIdial on the same day.

Moving to a cluster, and running one

We build the cluster beside your old server and move the data across on a date you choose. The migration is quoted with the cluster.

The move, step by step

  1. A sizing call: agents, dial level, inbound and recording. That gives the layout and a monthly price.
  2. The machines are ordered together from the datacenter on one private network, with delivery estimates shown first.
  3. The database goes in first, then the standby starts replicating, then each dialer is joined.
  4. Test calls run through your own trunks, and a pilot campaign runs before the floor moves.
  5. Leads, lists, campaigns and recordings move across on the chosen date.
  6. Monitoring, patching and a named contact take over from there.

If you have moved servers before, planning a zero-downtime cutover and moving VICIdial to new server hardware cover the same ground from the single-server side.

When a machine fails

If a dialer fails, the calls on it drop and the other dialers keep the floor working. If the database machine fails, the standby is promoted. Recovery times are written into your agreement.

Watching it day to day

In a cluster, a database problem shows up as a symptom on a dialer while the CPU and memory graphs look fine. Our cluster page checks every server every 10 seconds and says in plain words when something needs a person: an agent screen that stopped responding, a dialer running out of conference rooms, a campaign over its drop limit, a hopper about to run dry. When a shift goes wrong, we can record the next one second by second.

Patches tend to arrive late on clusters, because patching means touching servers mid-shift. When the PHP2Shell flaw reached VICIdial servers in October 2026, we patched every server we run within a day; here is what PHP2Shell is and how we patched it.

Upgrades have their own trap: they replace VICIdial's files, which wipes custom changes. A new revision runs on our demo server first and is tested end to end. Your changed files are merged with the new version, and servers are upgraded only when nobody is logged in.

Frequently asked

› How many agents before I need a VICIdial cluster?
There is no single agent number. By our estimates one machine carries about 600 concurrent calls with room to spare, but a cluster also separates the database from the calls and removes the single point of failure, and either can matter sooner.
› Can the machines in a cluster sit in different datacenters?
No. Dialers query the database many times a second, and a round trip between Germany and Finland is over 20 ms, which is enough to stall VICIdial. Every machine sits in one datacenter on a private network.
› Is one dialer enough for a small cluster?
We never build fewer than two. If a dialer fails, the calls on it drop and the other dialer keeps the floor working.

If one server is still enough for your floor, start on a plan and grow later. When you need more, size your cluster and see today's prices.

About VICIfast

We run VICIdial servers for call centres, so you don’t have to. You get the dialer set up, secured and kept running — we handle the server, the updates and the backups. Bring your own phone carrier at no markup, or use VICIfast Voice, our own US calling and numbers, from $0.0085 a minute. From $49 a month per server.

Questions? Call +1 636 556 0022 and talk to someone who runs dialers.

Citing this article

VICIfast Engineering. “VICIdial clusters: how they work, when you need one, and how to size one”. VICIfast LLC, November 18, 2026. Retrieved from https://vicifast.com/blog/vicidial-cluster-guide

Have questions?

Related posts

Comments

Comments are reviewed before they appear. We never publish your email.

No comments yet - be the first.