Resource allocation case study

This is a fictitious example of a deployment that did not meet initial expectations, the reasons why, and the steps that were taken to rectify the issues that were encountered. This case study takes you through the following steps:

Requirements and initial server/virtualization specification

  • Example Corp (our fictitious company) wanted to deploy Pexip Infinity as a proof of concept, with a requirement it would handle either a single video conference for 20 users, or up to 150 simultaneous, audio-only calls across 10 different conferences.
  • They wanted to use an off the shelf, dual-CPU server - Intel E5-2660 v3 – which has 10 physical cores per socket, and a Processor Base Frequency of 2.6 GHz and 32 GB RAM (consisting of 2 x 16 GB DIMMs).
  • As Pexip Infinity is a virtualized platform, Example Corp wanted to run multiple Virtual Machines (VMs) on the same host server, as they do for other VMs in their datacenter. They currently use VMware 5.5 as their virtualization platform.


Memory configuration

The memory configuration is not ideal for this server and CPU combination. We say as part of our Server design guidelines that a Conferencing Node VM should be deployed with 1 vCPU and 1 GB vRAM per physical core, and we also state that the physical RAM should be deployed across all channels accessed by the physical CPU. From the Intel Ark database we see that the specified E5-2660 v3 CPU has 4 memory channels, so for a dual CPU server, you should populate 8 DIMM slots (consult motherboard/server documentation as to which 8 channels should be populated if more exist), rather than the 2 slots currently occupied.

Pexip Infinity is an application that requires a high amount of compute power, with intensive calculations on data, so requires fast memory access. The more memory channels that a CPU can use, the better overall efficiency of the application.

In this case, assuming that Pexip Conferencing Nodes could utilize all available 20 cores (2 x 10-core CPUs), we would need a minimum of 20GB of physical RAM for the server. Given that DIMM modules usually are sized in 2, 4, 8, 16 etc. GB sticks, and that we need 8 modules (and assuming the same specification module is used in each slot), the ideal memory allocation is 8 x 4GB = 32 GB.

Hardware over-committing

Pexip Infinity does not support over-committing (also referred to as over-subscribing or over-allocation) of RAM and CPU resources within a host server, i.e. it does not co-exist well with other VMs running on the same host that may use the resources allocated to Pexip Infinity VMs. Running additional VMs on host cores that are in use by Pexip Infinity results in the hypervisor time-slicing the physical host resources between the VMs. For normal applications (e.g. email, database applications, directory services, file server etc.) that may exist in a virtual environment, the time-slicing by the hypervisor makes no perceivable difference to the running application. However, because Pexip Infinity is a real-time application, this will result in poor performance, such as stuttering video and audio. Pexip Infinity is a very CPU-intensive platform, and the reason for virtualization in our case is more related to hardware abstraction rather than the traditional sharing of resources. Further information can be found in Configuring VMware for Pexip Infinity and Advanced VMware ESXi administration. In particular, if you use vMotion with EVC, this can also cause issues and is also covered in this documentation.

The specified host server contains 20 cores. We must ensure that all the VMs that are running on this one host do not consume more than the 20 cores available. The specified CPU supports hyperthreading (effectively allowing 2 vCPUs per physical core); however, for Pexip Infinity to make use of the hyperthreading capabilities we need to adjust how the hypervisor works, by ensuring that we lock Conferencing Nodes to exact physical sockets. This is known as NUMA pinning, and is discussed further later on.

It is also important to note that the overall CPU utilization figure of the host server, as reported by the hypervisor, may still seem to be within reasonable tolerance of the entire CPU capacity, even if many VMs are running on the host and they consume more cores than are physically available. However, if this is true, the hypervisor will time-slice these resources, and you will still notice poor performance on Pexip Infinity. As such, if you see a low overall host CPU utilization level, it does not necessarily mean that you are NOT over-committing.

For all of these reasons, we would generally recommend specifying dedicated hosts to run Pexip Conferencing Nodes.

Rule of thumb capacity calculation

Examples of the number of connections (also referred to as ports) you may expect to achieve for various CPU types is given in Example Conferencing Node server configurations. As a rule of thumb, for a normal deployment (not NUMA pinned), you may expect to see the following:

  • Utilize approximately 1.4 to 1.5 GHz CPU capacity per HD (720p) connection
  • 1 HD connection = approximately 2 SD connections
  • 1 HD connection = approximately 12 audio connections
  • 1 Full HD (1080p) connection = approximately 2 HD (720p) connection (assuming Full HD is enabled on the deployment)

Note that these call ratios (above) are typical but will vary for each individual server/environment.

So, looking at the Intel Ark database for the CPU specified for this server, we can see the core speed of the processor is 2.6 GHz, each CPU contains 10 cores, and there are 2 physical CPU sockets. The calculation we can use to work out an approximate connection (port) capacity is:


Different hardware specifications may result in slightly lower numbers, hence, in our example documentation we have specified 35 HD connections for this CPU. This also assumes Pexip Infinity will consume all cores on the host.

Deployment 1 – making Conferencing Nodes too big

Additional deployment scope

  • Example Corp is using the host server (dual CPU, 10 physical cores per socket) to run multiple VMs, although they have read our guidance and understand that they must not over-commit resources.
  • They already have 2 standard IT infrastructure VMs consuming 4 cores of host resource each (8 cores in total) running on the host.
  • They want to use the remaining 12 cores for Pexip Infinity. Thus, they deployed a single Conferencing Node VM with 12 vCPUs and 10 GB of vRAM.
  • On boot, the administrator noted a very low HD and audio connection count, and a HD to audio ratio of 8x (HD = 4, audio = 32).
  • In the administrator log, the administrator noted the following entry:

Message="Multiple numa nodes detected during sampling" Detail="We strongly recommend that a Pexip Infinity Conferencing Node is deployed on a single NUMA node”

Understanding the deployment

Example Corp have used our calculation (with some mathematical transposition) to show that 12 of the 20 available cores could achieve at least the 20 HD connections and 150 audio connections required for the PoC (with a small amount of additional capacity).

And with the reported ratio of 1 HD connection = 8 audio connections, they would have assumed they should get:

22 x 8 = 176 audio connections

So why is the connection count so low?

At first glance from the deployment notes above, it may seem that Example Corp has simply failed to follow our memory guidelines and has not allocated the full 12 GB of RAM to a Conferencing Node totaling 12 vCPUs. However, there is more to it than that.

The Pexip Conferencing Node is now using 10 cores on one socket, and 2 cores on the other. It may seem logical to simply increase the number of vCPUs assigned to a Conferencing Node in order to achieve more computational power and thus higher connection capacity. However, when the number of vCPUs on a node increases beyond the number of physical cores in a socket (for a normal deployment), the Conferencing Node VM is then hosted on two different physical CPUs and requires access to different banks of memory. This actually makes things quite inefficient, and results in poor connection capacity. NUMA nodes are described in more detail in Achieving high density deployments with NUMA. The warning log entry in the administrator log shows that the Conferencing Node has spanned NUMA nodes; this must be rectified, and the following examples show how this was done.

Deployment 2 – making Conferencing Nodes too small

Additional deployment scope

  • Example Corp is using the host server (dual CPU, 10 physical cores per socket) to run multiple VMs, although they have read our guidance and understand that they must not over-commit resources.
  • They are now only utilizing 5 cores on the host with their standard IT infrastructure VMs.
  • Not wanting to make the same mistake as previously by spanning NUMA nodes, they decided to create smaller Conferencing Nodes to utilize the remaining 15 cores available on the host, so they deployed 3 Conferencing Node VMs with 5 vCPUs and 5 GB of vRAM each.
  • On boot, the administrator saw reasonable HD and audio connection counts on each of the nodes:
    • Node 1: HD = 8, audio = 68
    • Node 2: HD = 7, audio = 65
    • Node 3: HD = 7, audio = 65
    • Total capacity: 22 HD, 198 audio

Understanding the deployment

The capacity figures are reasonable, if just a little lower than Example Corp hoped for. In this case, the Example Corp administrator had not tuned the BIOS settings on the server, but left them as the defaults, and these were configured by the manufacturer with energy saving in mind. Changing the relevant power setting to Maximum Performance and no power saving should further enhance the Pexip Infinity connection count.

However, given that the minimum requirements were met, Example Corp continued with their testing. Still, they were disappointed to find that they could not get all 20 users with HD video into a single videoconference. When the 20th user attempted to join the conference, they were disconnected with an announcement saying that capacity had been exceeded, and the Example Corp administrator saw the following entry in the administrator log:

Message="Alarm raised" Node="" Name="capacity_exhausted" Time="1453719569.94" Details=""

Message="Participant failed to join conference." Conference="DB441" Participant="sip:+12345@;user=phone" Protocol="SIP" Participant-id="c09ee791-cd60-435e-b1f6-36c24e3dc9fc" Reason="Out of resource"

So, what had gone wrong?

To answer this, we need to look back at the resource allocation rules. For a multi-node deployment, each video-based VMR instance hosted on each Conferencing Node will reserve 1 HD of connection resource for a backplane. Because Example Corp had used 3 separate nodes, no single node had the capacity to host the entire conference, and so the conference was distributed among the nodes. Because a VMR instance was initiated on each node, Pexip Infinity immediately reserved 1 HD resource for the backplane for each instance. So, with 1 VMR instance running on each node, we can say:

Total node HD connection count – 1 HD backplane per VMR = Available node HD endpoint connections

Node 1: 8 – 1 = 7
Node 2: 7 – 1 = 6
Node 3: 7 – 1 = 6

Total available endpoint connections = 7 + 6 + 6 = 19

While this was not ideal, Example Corp continued their testing, but this time for the audio conferences. However, they were somewhat puzzled in that they only reached a little over 50% of their requirement of 150 concurrent audio calls across 10 conferences, even though the raw calculation showed that the nodes could handle up to 196 audio calls.

Given that there are multiple calls and conferences occurring at the same time, it is useful to look at the data in a different view. Within Pexip support, we can create a single pivot table that shows the call load across all nodes at any single point in time. The data shows the conferences that are in operation, the number of participants, and the call media type (video, audio or presentation), across all the nodes. In the case of Example Corp and the example data shown below, these calls were all audio, so only audio call types are displayed for each node:

Count of Stream Type
  Nodes and Media Types
Conference (capability)
IVR (audio-only)     1
VMR 1 (audio-only) 3    
VMR 2 (main video plus presentation)   6  
VMR 3 (main video plus presentation)     31
VMR 4 (main video plus presentation) 1    
VMR 5 (main video plus presentation)     9
VMR 6 (main video plus presentation)   4  
VMR 7 (main video plus presentation)     2
VMR 8 (main video plus presentation)   4  
VMR 9 (main video plus presentation) 10 3  
VMR 10 (main video plus presentation) 12 1  
Total 26 18 43

At first glance, we can see from the totals that only 26 audio calls were running on Node 1, 18 on Node 2 and 43 on Node 3, for a total of 87 concurrent calls. The audio participants are first connected to a Virtual Reception (IVR); from there they enter the conference ID they wish to join, and Pexip Infinity transfers them to the correct VMR. The user currently in the IVR wished to join VMR 4, however, after they entered the conference ID, they heard the capacity exceeded warning and the administrator saw a similar entry seen previously in the Administration Log.

Why did Example Corp achieve such a poor result?

The pivot table shows some very useful information, but does NOT show all the necessary connections reserved for backplanes for each VMR on each node, so we will need to account for these as well in our capacity calculations. In the resource allocation rules we have seen that each VMR in a multi-node deployment will reserve a connection for use by the backplane. However, we also noted that a Conference Capability (call type) can be defined on each VMR, so we can set a VMR to be audio-only. In this case the backplane will only reserve a single audio connection, rather than an HD connection as is the case with a video VMR.

We have added a label beside each VMR listed in the pivot table above to show how the Example Corp administrator had configured each of the Conference Capabilities for that VMR. You can see that apart from the IVR and VMR 1, all other VMRs have been left with the default Conference Capability of "main video plus presentation". As such, each of these VMRs will reserve an HD video resource, even if all the calls within it are audio-only.

So, in summary two mistakes were made here:

  1. The Conferencing Nodes were deployed with a vCPU count that was too small.
  2. The VMR Conference Capability type for audio-only VMRs was incorrectly set (i.e. left as the default rather than set to audio-only).

The necessity for reservation of the backplane connection is the reason why we recommend processors with high core count: so that we can deploy Conferencing Nodes with high number of vCPUs (as long as they do NOT span NUMA nodes), rather than a larger number of nodes with a smaller vCPU count. In this way, we can achieve higher concentration of capacity with the same resources. For further information, see our Server design guidelines and Handling of media and signaling.

If Example Corp had deployed 2 Conferencing Nodes, one with 5 vCPUs and one with 10 vCPUs, they would have been able to achieve their requirement of creating a single videoconference with 20 HD video participants. In addition, if they had set the Conference Capability of the VMRs that were assigned to be used for audio-only conferences, they would have been able to achieve the audio capacity listed in their requirements.

Deployment 3 – correct deployment

Additional deployment scope

  • Example Corp have decided to follow our recommended best practice (as per the Server design guidelines) and use all the resources of this host server (dual CPU, 10 physical cores per socket) for the deployment of Conferencing Node VMs.
  • They continue to utilize 8 x 4 GB RAM DIMMs (total of 32 GB RAM), thus have populated all 4 memory channels for each of the NUMA nodes (sockets).
  • The administrator has set the BIOS performance levels to Maximum and switched off any Energy Saving settings.
  • They deploy two Conferencing Nodes, each utilizing 10 vCPUs and 10 GB vRAM. In this way, both Conferencing Nodes will consume the resources of each NUMA node (socket) without spanning.
  • On boot, the administrator saw good HD and audio connection counts on each of the nodes:
    • Node 1: HD = 18, audio = 144
    • Node 2: HD = 17, audio = 136
    • Total capacity: 35 HD, 280 audio
  • They ensured that the IVR and VMRs that are specifically used for audio-only calls are configured with “Conference Capability” set to audio-only.

Understanding the deployment

The connection capacities are calculated during the boot phase of each Conferencing Node, when they simulate call loads. It is not uncommon for the nodes to give a subtle variance, even when they are running on the same host and configured in the same way.

Example Corp re-ran their tests and confirmed that the hardware could meet their initial requirements, i.e. they could host either:

  • a single 20-user videoconference, where each participant was connected using HD video, or
  • 150 concurrent audio calls spread across 10 simultaneous conferences.

They noted that the videoconferences were initiated on Node 1. This is because Node 1 had calculated a slightly higher capacity during its boot phase, and media is allocated according to which node has the most available capacity. When additional participants are added to the same conference, Pexip Infinity attempts to keep all media on the same node. Example Corp then noted that as the load on Node 1 increased to its maximum, media for additional participants was allocated to Node 2. This is as per the distributed system design of Pexip Infinity (for more information, see Handling of media and signaling). The audio conferences appeared to be more evenly distributed across the two nodes. This is due to the variation in load on the nodes as new conferences are initiated, remembering that the node with the most available capacity is used to handle a new conference.

Example Corp were now satisfied that the hardware was sufficient and Pexip Infinity would perform well in their environment. But they wondered if there was even more they could do. After all, they had read in the resource allocation rules that Skype for Business / Lync calls consume additional HD resources if they are either sending or receiving content. In addition, they wondered if the hardware could support both elements of their original requirements concurrently, i.e. host a single 20-person HD videoconference, and the 150 audio calls across 10 separate audio conferences, all at the same time.

Deployment 4 – NUMA pinning to increase capacity

Additional deployment scope

  • Example Corp followed our recommended best practice (as per the Server design guidelines) and used all the resources of this host server (dual CPU, 10 physical cores per socket) for the deployment of Conferencing Node VMs.
  • In addition, they made use of the hyperthreading capability of the Intel E5-2660 v3 CPU, so followed our additional guidance regarding VMware NUMA affinity and hyperthreading.
  • They changed the physical memory to utilize 8 x 8 GB RAM DIMMs (total of 64 GB RAM), and thus have populated all 4 memory channels for each of the NUMA nodes (sockets).
  • The administrator set the BIOS performance levels to Maximum and switched off any Energy Saving settings, and ensured that hyperthreading is enabled.
  • They deployed two Conferencing Nodes, each utilizing 20 vCPUs and 20 GB vRAM. In this way, both Conferencing Nodes will consume the resources of each NUMA node (socket), and because the Conferencing Nodes are now pinned to a NUMA node and make use of hyperthreading, they still do not span NUMA nodes.
  • On boot, the administrator saw good HD and audio connection counts on each of the nodes:
    • Node 1: HD = 28, audio = 224
    • Node 2: HD = 27, audio = 216
    • Total capacity: 55 HD, 440 audio
  • They ensured that the IVR and VMRs that are specifically used for audio-only calls are configured with “Conference capability” set to audio-only.

Example Corp ran their test again. They now find that they can run both scenarios set out in the initial requirements simultaneously.

However, if all 20 video participants in the VMR were Skype for Business / Lync endpoints (a potentially unlikely scenario), and they consumed both HD video and RDP presentation (perhaps in a multi-screen user environment), this in itself would consume at least 40 HD resources, and the conference would be split across both nodes, hence the VMR instance on each node would also require an HD resource. This would leave approximately 13 HD resources available on Node 2, and based on 1 HD resource = 8 audio resources, we would only have available 104 audio connections. This would not be enough to also support the 150 concurrent audio calls across 10 simultaneous audio conferences. At this point Example Corp would need to add an additional host and nodes to their Pexip Infinity deployment.