The ingress gateway and Voice Browser are separated from the servers that provide them with media files, VXML documents, and call control signaling. These factors make the bandwidth requirement for the Unified CVP.

For example, assume that all calls have 1 minute of VRU treatment and a single transfer to an agent for 1 minute. Each branch has 20 agents. If each agent handles 30 calls per hour, you have 600 calls per hour per branch. The call average rate is 0.166 calls per second (CPS) per branch.

Even a small change in these variables can have a large impact on sizing. Remember that 0.166 calls per second is an average for the entire hour. Typically, calls do not come in uniformly across an entire hour, and there are usually peaks and valleys within the busy hour. External factors can also affect the call volume. For example, bad weather increases call volumes for business like airlines and promotions can increase call volumes for retailers. Find the busiest traffic period for your business, and calculate the call arrival rate based on the worst-case scenario.

VXML Documents

A VXML document is generated for every prompt that is played to the caller. This document is generated based on voice application scripts that you write using either Unified ICM scripts or Cisco Unified Call Studio, or both. A VXML document varies in size, depending on the type of prompt being used. For example, menu prompts with multiple selections are larger in size than prompts that play announcements only.

Note	The approximate size of a VXML document for a Call Server or a VXML Server and the gateway is 7 kilobytes.

You can calculate bandwidth in the following ways:

Bandwidth Estimated by Prompts

You can estimate the bandwidth for a branch office as follows:

CPS * Bits per Prompt * Prompts per call = Bandwidth in bps

For the previous example, consider a VXML document of 7 kilobytes:

7,000 bytes * 8 bits/byte = 56,000 bits per prompt 
(0.166 calls/second) * (56,000 bits/prompt) * (Number of prompts / call) = bps per branch

Bandwidth Estimated by VXML Documents

Use the VXML document sizes listed in the following table to calculate the required bandwidth. The document sizes in the following table are measured from the VXML Server to the Voice Browser.

Table 2. Approximate Size of VXML Document Types

VXML Document Type	Approximate Size in bytes
Root document (one required at beginning of a call)	19,000
Subdialog_start (at least one per call at beginning of a call)	700
Query gateway for Call-ID and GUID (one required per call)	1300
Menu (increases in size with the number of menu choices)	1000 + 2000 per menu choice
Play announcement (a simple .wav file)	1100
Cleanup (one required at the end of a call)	4000

Note	For more complex solutions, this second method yields a better estimate of the required bandwidth than estimating by prompts.

Media File Retrieval

You can store media files, or prompts, locally in flash memory for IOS Voice Gateway and in the file system for Cisco VVB. Storing them locally eliminates bandwidth considerations. However, it is difficult to maintain these prompts because a prompt that requires changes must be replaced on every router or VVB. Local storage of these prompts on an HTTP media server (or an HTTP cache engine) enables the gateway to locally cache voice prompts after retrieval. An HTTP media server can cache multiple prompts, depending on the number and size of the prompts. The refresh period for the prompts is defined on the HTTP media server. The bandwidth usage is limited to the initial load of the prompts at each gateway, including the periodic updates after the expiration of the refresh interval.

Note	You cannot disable the HTTP Cache in VVB.

Not caching prompts at the VXML Gateway has significant impacts:

• It degrades Cisco IOS performance by 35-45%.

• It requires extra bandwidth. For example, if you have 50 prompts with an average size of 50 KB and a refresh interval of 15 minutes, the average bandwidth usage is:

  (50 prompts) * (50,000 bytes/prompt) * (8 bits/byte) = 20,000,000 bits 
  (20,000,000 bits) / (900 second) = 22.2 kbps per branch 

Note	Bandwidth considerations for VVB include bandwidth for VXML documents, Media File retrieval and RTP streams for G.711 and G.729 voice traffic.

For Unified CVP, you can group WAN and LAN traffic into the voice traffic, the call control traffic, and the data traffic.

Media Resource Control Protocol Traffic

The VXML Gateway and Cisco Virtualized Voice Browser communicate with ASR/TTS Servers using both Media Resource Control Protocol (MRCP) v1.0 and v2. This protocol establishes connections to the ASR/TTS Server, such as Nuance. The connection can be over LAN or WAN.

Note

Cisco does not test or qualify speech applications in WAN environment. For guidelines on design, support over WAN and associated caveats, see the vendor-specific documentation. TAC provides limited support (as in the case of any third-party interoperability certified products) on issues related to speech applications.

Generally, a distributed topology is the most bandwidth intensive for Unified CVP. The Ingress Gateway and Voice Browser are separated from the servers that provide the media files, VXML documents, and call control signaling.

Note	Recall the earlier example of all calls have 1 minute of VRU treatment and a single transfer to an agent for 1 minute. Each branch has 20 agents, and each agent handles 30 calls per hour for a total of 600 calls per hour per branch. The call average rate is 0.166 calls per second (CPS) per branch.

(17,000 bytes/call) * (8 bits/byte) = 136,000 bits per call 
(0.166 calls/second) * (136 kilobits/call) = 22.5 average kbps per branch 

After the proper application bandwidth and QoS policies are in place, consider the network latency in a distributed CVP deployment. With sufficient network bandwidth, the primary contributor to latency is the distance between the Voice Browser and the Call Server or VXML Server. In distributed CVP deployments, minimize the latency and understand its effect on solution performance.

Network latency affects a distributed CVP deployment in the following ways:

• It affects the end-user calling experience when the network latency is between CVP components. Call signaling latency with SIP between the Call Servers and voice gateways affects the call setup time. Latency can add a period of silence during this setup. It includes the initial call setup and subsequent transfers or conferences that are part of the final call flow.

• It significantly affects the download time for VXML application documents, and has a pronounced effect on the ultimate caller experience.

The following system configuration changes can reduce WAN delays from geographic separation of the Voice Browser from the VXML Server:

1. Provide audio to the caller during periods of silence.

   The following settings provide ringback and audio during times of dead air so that the caller does not disconnect:

   • To add a ringback tone during longer than normal call setup times with VRU, on the survivability service, keep the wan-delay-ringback setting at 1.

   • Add the VRU subsystem settings for IVR.FetchAudioDelay and IVR.FetchAudioMinimum. These WAN Delay settings are required when the root document fetch is delayed over the WAN link.

   • Specify the value for IVR.FetchAudio as follows: IVR.Fetchaudio= flash:holdmusic.wav. Leave the default empty so that nothing is played in a normal scenario.

   • Retain the default setting of 2 to avoid a blip sound in a normal network scenario.

   • Set WAN Delay to zero to play a holdmusic.wav immediately for a minimum of 5 seconds.

   • Use ECC variables, such as user.microapp.fetchdelay, user.microapp.fetchminimum, and user.microapp.fetchaudio, to override ECC variable values in between invocations of getSpeechExternal microapps.

     Note	You cannot use ECC variables while a call is at the Virtualized Voice Browser.

2. Enable Path MTU Discovery on the IOS Voice Gateways.

   On the IOS Voice Gateways, add the ip tcp path-mtu-discovery command.

   The Path MTU Discovery method maximizes the use of available bandwidth in the network between the endpoints of a TCP connection.

3. Minimize round trips between the VXML Server and the ICM script.

   When control is passed from a running VXML Server application back to the ICM script, you incur a significant WAN delay.

   After the VXML Server application starts to run, minimize the number of trips back to the Unified CCE script. Each round trip between the VXML Server and the Unified CCE script incurs a delay. It establishes two new TCP connections and HTTP retrieval of several VXML documents, including the VXML Server root document.

4. Decrease the size of the VXML Server root document.

   On the VXML Server, in your gateway adapter plugin.xml file change:

   <setting name="vxml_error_handling">default</setting> 

   To:

   <setting name="vxml_error_handling">minimal</setting> 

   For example, the location of the plugin.xml file for the CISCO DTMF 1 GW adapter is Cisco\CVP\VXMLServer\gateways\cisco_dtmf_01\6.0.1\plugin.xml.

Note

HTTP transfers VXML documents and other media files that are played to the caller. For the best end-user calling experience, treat the HTTP traffic with a priority higher than that of normal HTTP traffic in an enterprise network. If possible, treat this HTTP traffic the same as CVP call signaling traffic. As a workaround, you can move the VXML Server to the same local area as the Voice Browser, or use Wide Area Application Service (WAAS).

The Call Server marks only the QoS DSCP for SIP messages, if done via Windows policy. If you need QoS for CVP traffic across a WAN, configure network routers for QoS using the IP address and ports to classify and mark the traffic. The following table outlines the necessary configuration.

The CVP-Data queue and the Signaling queue are not a priority queue in Cisco IOS router terminology. Use the priority queue for the voice or other real-time traffic. Reserve some bandwidth based on the call volume for call signaling and CVP traffic.

Some CVP deployments have all the components centralized. Those deployments use a LAN structure, so WAN network traffic is not an issue. A WAN might impact your bandwidth and QoS for CVP in the following scenarios:

• A distributed CVP deployment with a WAN between the Ingress Gateways and the Unified CVP servers

• A CVP deployment with a WAN between the Ingress Gateway and the agent.

CVP considers QoS in the following way:

• CVP has no private WAN network structure. When required, WAN activity is conducted on a converged WAN network structure.

• CVP does not use separate IP addresses for high- and low-priority traffic.

Note

Resource Reservation Protocol (RSVP) is used for call admission control. Routers also use it to reserve bandwidth for calls. RSVP is not qualified for call control signaling through the Unified CVP Call Server in SIP. For call admission control, the solution is to employ Locations configuration on CVP and Unified CM.

The amount of traffic sent between the Central Controllers (routers) and PGs is largely based on the call load at that site. Transient boundary conditions, like configuration loads, and specific configuration sizes also affect the amount of traffic. Bandwidth calculators and sizing formulas can project bandwidth requirements more accurately.

Bandwidth calculations for a site with an ACD and a VRU must account for both peripherals. Use 1000 bytes per call as a rule, but monitor the actual behavior once the system is operational to ensure that enough bandwidth exists. Based on that rule, a site that has four peripherals, each taking 10 calls per second, requires 320 kbps of bandwidth. (Unified CCE meters data transmission statistics at both the Central Controller and PG sides of each path.)

As with bandwidth, Unified CCE requires specific latency on the network links to function as designed. The private network between redundant Central Controller and PG nodes has a maximum round-trip latency of 80 ms. The PG-to-CC public network has a maximum round-trip latency of 400 ms to perform as designed. Meeting or exceeding these latency requirements is important for Unified CCE post-routing and translation routes.

Note	In general, Agent Greeting feature requires shorter latency across the system. For example, the public network has a maximum round-trip latency of 100 ms to support Agent Greeting feature as designed.

Unified CCE bandwidth and latency design depends on an underlying IP prioritization scheme. Without proper prioritization in place, WAN connections fail.

Depending on the final network design, your IP queuing strategy in a shared network must achieve Unified CCE traffic prioritization concurrent with other non-DNP traffic flows. This queuing strategy depends on traffic profiles and bandwidth availability. Success in a shared network cannot be guaranteed unless the stringent bandwidth, latency, and prioritization requirements of the solution are met.

This section presents QoS marking, queuing, and shaping guidelines for both the Unified CCE public and private network traffic. Provisioning guidelines are presented for the network traffic flows over the WAN, including how to apply proper Quality of Service (QoS) to WAN traffic flows. Adequate bandwidth provisioning and implementation of QoS are critical components in the success of contact center enterprise solutions.

Generally, your contact center enterprise WAN network structure uses separate links for both its Private and Public networks. For optimal network performance characteristics (and route diversity for the fault-tolerant fail-overs), Unified CCE requires dedicated private facilities, redundant IP routers, and appropriate priority queuing.

Enterprises deploying networks that share multiple traffic classes prefer to maintain their existing infrastructure rather than revert to an incremental, dedicated network. Convergent networks offer both cost and operational efficiency, and such support is a key aspect of Cisco Powered Networks.

You can deploy Unified CCE with a convergent QoS-aware public network and a convergent QoS-aware private network environment. But, your solution must meet the stringent latency and bandwidth requirements.

Unified CCE uses the Differentiated Services (DiffServ) model for QoS. DiffServ categorizes traffic into different classes and applies specific forwarding treatments to the traffic class at each network node.

The required bandwidth for phone-to-Unified CM signaling is 150 bps for each phone.

For example, in a 1000 agent solution, each contact center site requires approximately 150 kbps.

You can locate Agent and Supervisor desktops remotely from the Agent PG. In a poorly designed deployment, high time-out values can cause an extreme delay between the desktop server and desktop clients. Large latency affects the user experience and lead to confusing or unacceptable results for the agent. For example, the phone can start ringing before the desktop updates. Limit the latency between the server and agent desktop to 400-ms round-trip time for Cisco Finesse.

Cisco Finesse also requires that you limit latency between the Cisco Finesse server and the PG to 200-ms round-trip time. Limit latency between Cisco Finesse servers to 80-ms round-trip time.

Cisco Finesse does not support configuration of QoS settings in network traffic. Generally, have the QoS classification and marking of traffic done at the switch or router level. You can prioritize signaling traffic there, especially for agents who are across a WAN.

The following parameters have a combined effect on the responsiveness and performance of the Cisco Unified Intelligence Center on the desktop:

• Real-time reports—Simultaneous real-time reports run by a single user.

• Refresh rate for realtime reports—If you have a Premium license, you can change the refresh rate by editing the Report Definition. The default refresh rate for Unified Intelligence Center is 15 seconds. The deafult refresh rate for Live Data is 3 seconds.

• Cells per report—The number of columns that are retrieved and displayed in a report.

• Historical report—Number of historical reports run by a single user per hour.

• Refresh rate for historical reports—The frequency with which report data is refreshed.

• Rows per report—Total number of rows on a single report.

• Charts per dashboard—Number of charts (pie, bar, line) in use concurrently on a single dashboard.

• Gauges per dashboard—Number of gauges (speedometer) in use concurrently on a single dashboard.

The required bandwidth varies based on the refresh frequency, the number of rows and columns in each report, and other factors. Across a WAN, Unified Intelligence Center requires a latency of 200 ms or less.

You can use the Cisco Unified Intelligence Center Bandwidth Calculator (http://www.cisco.com/c/en/us/support/customer-collaboration/unified-intelligence-center/products-technical-reference-list.html) to calculate the bandwidth requirements for your Unified Intelligence Center implementation.

This sample data came from a test on a LAN with a local AWDB database and a client machine to run the reports.

The load for this test used a single Unified Intelligence Center user running the following:

Two hundred Unified Intelligence Center users, each concurrently running:

• Two realtime reports with 100 rows per report, with 10 columns each.

• Two historical reports with 2000 rows, with 10 columns each.

• Two live data reports with 100 rows, with 10 columns each. (Adjust this based on the deployment type whether LD runs or not).

This table gives the observed bandwidth usage for the test:

The required bandwidth differs based on such parameters as the number of rows in each report and the number of concurrent report executions.

With Silent Monitoring, supervisors can listen to the agent calls in CCE call centers. Voice packets sent to and received by the monitored agent's IP hardware phone are captured from the network and sent to the supervisor desktop. At the supervisor desktop, these voice packets are decoded and played on the supervisor's system sound card. Silent Monitoring of an agent consumes approximately the same network bandwidth as an extra voice call. If a single agent requires bandwidth for one voice call, then the same agent being silently monitored requires bandwidth for two concurrent voice calls. To calculate the total bandwidth required for your call load, multiply the number of calls by the per-call bandwidth for your codec and network protocol.