Example Conferencing Node server configurations
This section provides some example server configurations for Transcoding Conferencing Nodes, with an estimate of the capacity (in terms of how many HD connections — sometimes referred to as "ports") you can expect to achieve with each option. These figures assume that all memory channels are populated, that NUMA affinity and hyperthreading has been enabled, and that all other actions in the best practices checklist have been completed.
This topic covers:
- Issues to consider when specifying hardware
- Pexip recommendations
- Recommended server sizes
- Other processors
When choosing hardware, you should consider carefully what you want to optimize for. The best choice varies depending on how you intend to use your Pexip Infinity deployment, your budget and the costs of hosting your hardware.
Management and proxying nodes rarely require a full socket. In many cases, the Management Node and some proxying node nodes share a socket, sometimes with a small transcoding node.
Because these nodes do not perform any media processing, it is safe to use older or less capable hardware for these nodes. Please note that proxying nodes do require at least the AVX instruction set. Always prioritize the newest and most capable hardware for transcoding nodes.
Where the requirement is for a fixed number of ports overall or within a particular physical location and this requirement is for up to 90HD ports, transcoding for that deployment or location can easily be provided on a single socket.
Large conferences work best on large nodes. Any one conference can span only three nodes in a given logical location. The conference takes one backplane on each node that it uses, so minimizing the number of nodes that a conference spans reduces this overhead.
In this scenario we would normally recommend building the largest individual nodes possible.
When a high volume of small conferences or gateway calls are expected, optimizing purely for individual node capacity is less important. There are benefits of having fewer larger nodes over more smaller ones as there is less likelihood of a conference being fragmented across two or more nodes thus saving on backplane overheads.
In this scenario the number of ports per socket is probably more important than then number of ports per node. Where rack space is at a premium you may want to consider Achieving ultra-high density with Sub-NUMA Clustering.
We recommend 3rd- or 4th-generation Intel Xeon Scalable Series processors. The Gold line generally represents the best value in terms of the number of ports provided for a given hardware spend. We like the Xeon Gold 6342 for its good 2.8GHz base clock speed and 24 physical cores. When optimally deployed it can offer 97-100HD per socket or around 195HD in a 1U 2-socket server.
The Xeon Gold 6348 and 6354 parts are slightly less capable, but still represent good options if the 6342 is not available. We have no data for the Xeon Gold 6442Y, but expect its performance to be similar.
Less powerful hardware is available, but as a proportion of the server cost the savings are not large for a significant reduction in capability. Where possible you should over-specify your hardware because forthcoming features of the Infinity platform may require additional processing power.
If you need to obtain the most capacity out of every last rack unit, the Intel Xeon Platinum 8458P looks hard to beat. We have not yet had a chance to test this 2.7GHz behemoth with 44 physical cores, but we would expect 350HD or more from a 1U 2-socket server.
Sometimes new hardware is not an option. If you need to use existing hardware, try to find a 6248R machine. When new, this was our preferred hardware option and it is being used successfully to run Pexip Infinity by all sorts of organizations all over the world. Make sure all 6 memory channels are populated on each socket and you should see 87-95HD per socket or around 180HD in a 1U 2-socket server.
For the Pexip Infinity platform, the following server configurations provide maximum performance for cost:
|Capacity (no. of connections)|
|Around 195HD||350HD+ (estimate)||Around 180 HD|
2x24-core Ice Lake
Launched: Q2 2021
2x44-core Sapphire Rapids
Launched: Q1 2023
2x24-core Cascade Lake
Launched: Q1 2020
2 x Intel Xeon Gold 6342
2 x Intel Xeon Platinum 8458P
2 x Intel Xeon Gold 6248R
16 x 16GB (16 x 8GB if available)
8 DIMMs per socket
12 x 8GB / 12 x 16GB
6 DIMMs per socket
|Network||1 Gbps NIC (we recommend dual NIC for redundancy)|
|Power||We recommend redundant power|
We are unable to test all processors on the market. We do maintain some data on real-world usage, but this is not always reliable as we have no way of telling if the deployment has been performed according to our best practices. If you have a particular processor in mind and would like an estimate of its capability, please contact your Pexip authorized support representativeyour Pexip Solutions Architect or authorized support representative.