Sizing estimation approaches

Minimum requirements

Minimum server parameters for the platform to operate in single-server mode:

  • 4 cores,

  • 8 GB RAM,

  • 100 GB HDD.

Recommended server parameters for the platform to operate in single-server mode with the DB and file storage on the server:

  • 8+ cores,

  • 32+ GB RAM,

  • 250+ GB SSD TBW 3500+ TB,

  • 1+ TB HDD/SSD for the DB,

  • 1+ TB HDD/SSD for file storage,

  • RAID.

Performance table

Load testing was performed on various processors in single-server and two-server configurations. Three tests were performed with different emphasis: either on CPS, or on the number of simultaneous conversations, or on the number of simultaneous IVRs. In each of them the other two indicators were obviously non-zero but significantly lower than the peak possible values.

The testing result for various processors was generalized and reduced to 1 core.

Average performance table per 1 core at 2.5 GHz:

Emphasis on quantity Per core Per server with 12 cores at 2.5 GHz

CPS

cps - 6, calls - 25, ivr - 25

cps - 72, calls - 300, ivr - 300

Calls

cps - 0.5, calls - 180, ivr - 0

cps - 6, calls - 2000, ivr - 0

IVR

cps - 0.5, calls - 125, ivr - 125

cps - 6, calls - 1500, ivr - 1500

Approaches to calculating used resources

Three approaches are distinguished:

  1. Expansion. We gradually load, evaluate actual resource usage. Not enough — we add.

  2. Load extrapolation. We measure the costs obtained on a small volume of executing processes in a fully functioning system. For example, on 10 operators. We extrapolate the result to the expected number of simultaneously executing processes. For example, to 1000 operators.

  3. Expert assessment. Based on calculated indicators for telephony, we apply a coefficient of 3-6 depending on the redundancy type and system size.

Approach 1. Expansion

Install on 1 server and configure functionality for full production operation in a single-server configuration. As load increases, add servers and distribute the configuration across more servers.

We always start with 1-2-4 servers depending on the redundancy type and roughly estimated load. If utilization approaches 60-70%, we identify costly services and allocate them to separate machines. Or reconfigure the system based on the increased number of servers.

Mainly the following are moved out:

  • databases;

  • microservices serving media traffic;

  • microservices serving SIP signaling;

  • microservices serving IVR and other scripts;

  • microservices serving the data model

  • web server microservices

  • product layer microservices

  • change subscription servicing microservices

Then, observing the resource usage of the remaining microservices, identify the leaders.

Approach 2. Load extrapolation

  1. The system is configured for full functionality on a known necessary number of servers (a number that obviously will not have to be abandoned, possibly on one server). All scripts are created, all processes, rules, connections, integrations are configured.

  2. A part of the planned processes or request flows is transferred to the system, the share of which can be quantified relative to the planned peak load. For example, 10%. Not 1000 operators are connected but 100, the balancer sends not the whole call flow but only 10%, etc.

  3. Actual system resource usage while servicing the allocated part is measured.

  4. The obtained value is extrapolated to the planned one based on the allocated share.

  5. An increasing redundancy coefficient is applied. From 1.5 to 2.

  6. An increasing peak coefficient is applied to ensure partial load. Coefficient 1.5 (the base CPU utilization level on servers should not exceed 70%, excluding possible point peaks).

  7. The required number of corresponding servers is calculated.

  8. The DB and archival data and file storages are considered separately.

It is important here, when allocating part of the processes, to adequately estimate its share of the total. Each process of a fully loaded system must execute and be represented in the measured part.

Quadratic (exponential) reduction of load in the measured part relative to the full one must not be allowed. For example, reducing a collection by 10 times reduces cross-intersection by 100 times, and a JOIN query completes that much faster.

Approach 3. Expert assessment

The telephony estimate can be built from the indicators given in the table.

  1. Roughly estimate the expected peak system load:

    • How many simultaneous conversations and of what kind.

    • How many subscribers waiting in queues (probably not more than the number of external trunks).

    • How many phone devices will be registered simultaneously and how often they will re-register.

    • Which data processing processes will run, how often, and with what load. This can be converted to the number of components per second.

    • How many users will be connected to the system simultaneously.

    • Which reports and dashboards will be built simultaneously with the main work.

    • What volume of conversations will be recorded and how long they will be stored.

    • How many domains will be in the system.

  2. Determine external services.

    • Whether external file storage will be connected.

    • Whether the DB will be moved to external servers.

  3. Determine the redundancy type based on goals.

    • Whether the installation will be ready for the failure of one server. Calculated coefficient (N+1)/N, but practically from 1.5 to 2 depending on configuration.

    • Whether data center redundancy will be used (installation across two data centers). Coefficient 2.

    • Whether zones ready for decommissioning will be formed. The coefficient depends on the number and size of zones. Rough coefficient 2.

  4. Estimate the required number of cores.

    • Based on the table above and the estimated peak load indicators, the required number of cores is estimated.

    • Based on the number of cores, the number of corresponding servers is estimated.

  5. Estimate the required disk system.

    • The disk subsystem is covered in the article Disk load

  6. Estimate the required RAM.

    • The base 32 GB of memory per server is used by standard telephony processes of any scale and a call center of up to 100 people with the DB on the server.

    • The following increase RAM requirements:

      • Large DBs retaining history for several years — on the servers where the DB runs;

      • The presence of report-building processes over long date periods — on the servers where the DB runs;

      • The presence of capacious collections with caching in the product layer data model.

  7. Add a reserve that reduces peak utilization.

    • average CPU load should not exceed 70%. Coefficient 1.5. (the base CPU utilization level on servers should not exceed 70%, excluding possible point peaks).

    • SSD/HDD disks should not be filled more than 50% (bytes, inodes).

    • The read/write queue load on an HDD should not be observed for periods longer than 1 second and in total should not exceed 1% (irrelevant with SSD).

  8. Add a reserve for the redundancy scheme. Calculated coefficient (N+1/N), where N is the calculated number of servers, in practice from 1.5 to 2.

  9. When installing the DB on cluster servers, add resources for DB operation. (from 2 cores and 100 GB to 24+ cores and 1+ TB + RAID)

    • Approximate size for calculation — 150 MB per month of storage for a domain containing 100 users and 30 thousand connections per month.

  10. When using local storage, add resources (SSD/HDD) to provide it.

    • Approximate size for calculation — 1 minute of stereo MP3 — 180 KB.

    • Additionally, other collections with attachments must be estimated and considered:

      • whether scripts store anything.

      • whether mailboxes are configured and how long messages are stored.

      • whether there are other historical data with attachments.

      • whether DB backup is performed by system tools (even to external storage).

  11. Apply a coefficient for the product layer working with the project data model (from 1.1 to 2).

    • coefficient 1.1 — if the product layer is installed in a single domain and is used only for viewing the call journal.

    • coefficient 1.5 — if the product layer is installed in all domains with a total of more than 20.

    • coefficient 2.0 — if users massively work in applications and non-voice servicing processes run.

  12. Video conferencing. Examples:

    • 1 server with 8 cores and a 1 Gbit/s channel handles:

      • 1 room for 200 people where everyone has cameras on and up to 10 speakers.

      • 1 room for 500 people where 1 speaker broadcasts video/screen and audio, the rest without cameras and microphones.

      • 10 rooms of 300 people where only audio is broadcast.

      • 100 rooms of 3-5 participants where everyone has video on and everyone is a speaker.

    • Mixing is performed deferred using free server resources.