Demystifying vCPUs: How Many Cores is 16 vCPU?

Virtual CPUs (vCPUs) provide critical compute resources for virtual machines (VMs). But it‘s not always clear how vCPUs map to the physical CPU cores that do the actual computation. If you configure a VM with 16 vCPUs, how many real cores will it use? Let‘s take a deep dive into vCPUs and uncover what‘s happening under the hood.

A Quick History of CPU Cores

To understand vCPUs, we first need to understand physical CPU cores. Originally, CPUs were single core – all computation was done on one core. AMD and Intel introduced dual-core CPUs in 2005, allowing two threads to run simultaneously.

The core wars continued with quad core in 2007, six core in 2010, eight core in 2011. Fast forward to today in 2022, and flagship chips like AMD‘s EPYC Genoa scale up to 96 cores per socket!

Here‘s a quick history of core counts over time:

This massive growth in cores enabled new levels of parallel processing power. Now let‘s look at how virtual CPUs relate to these physical cores.

vCPU Scheduling and NUMA Architecture

Virtual CPUs (vCPUs) are CPU resources assigned to VMs and managed by the hypervisor. The hypervisor has a scheduler that multiplexes vCPU threads across the available physical cores.

On servers, multiple CPU sockets are grouped via NUMA (Non-Uniform Memory Access). vCPUs can be pinned or restricted to cores in a particular socket. This avoids the higher latency of accessing memory attached to distant sockets.

Optimizing this NUMA-aware scheduling is key to achieving top performance. Matching vCPU locality with things like huge pages and memory controllers boosts throughput.

Hyperthreading – Doubling Up on Cores

Most modern CPUs employ SMT (simultaneous multi-threading) or hyperthreading to enable multiple logical processors per core. This lets a core handle two threads in parallel – essentially doubling the number of vCPUs accessible to VMs.

For example, say a host has two 10-core Xeon chips, for 20 physical cores. With hyperthreading enabled, the host presents 40 logical processors. VMs could access up to 40 vCPUs despite having only 20 actual cores.

Performance varies by workload when hyperthreading. Integer-heavy tasks see substantial gains. But floating point and mixed workloads are more limited by physical cores. Still, hyperthreading delivers extra throughput for many server apps.

Real World vCPU to Core Mapping

Let‘s walk through some examples of mapping vCPUs to physical cores on hosts:

KVM: On KVM, pinning vCPUs to cores is configurable via libvirt XML. This forces 1:1 mapping for guaranteed performance:

<vcpu placement=‘static‘>16</vcpu>  
<cputune>
 <vcpupin vcpu=‘0‘ cpuset=‘0‘/>
 <vcpupin vcpu=‘1‘ cpuset=‘1‘/>
</cputune>

Without pinning, KVM dynamically allocates vCPU threads across available cores.

VMware ESXi: One VMware best practice is limiting vCPUs to physical cores to avoid saturation:

"A general best practice is to allocate vCPUs equal to the number of physical cores in the ESXi host." (source: vmware.com)

But with hyperthreading, having more vCPUs than cores can work well too.

AWS: With AWS, you choose instance types with preset vCPUs. An m5.16xlarge has 64 vCPUs on a 64-core host using hyperthreading. This matches vCPUs to hw threads.

When Cores Outnumber Workloads

What if you have say 12 physical cores, but only need to run a couple VMs with light loads?

In this case, limiting vCPUs saves resources. A "right-sized" VM prevents wasting unused cores that could be allocated elsewhere. Plus condensed VMs allow packing more onto a host.

Thin provisioning vCPUs only as needed ensures efficient utilization of available cores.

The Future is Massive Parallelism

Modern data center workloads like AI training demand extreme compute power. This is driving core counts even higher through chiplet designs.

AMD uses 7nm chiplets with up to 12 cores each, interconnected via Infinity Fabric. This modular approach enables scalability up to 96 cores per EPYC socket!

Intel recently revealed its Falcon Shores XPU with up to 176 cores per socket based on tiled architecture. Massive parallelism will continue expanding as Moore‘s Law slows.

Optimizing vCPUs for Performance and Cost

Properly configuring vCPUs is key to maximizing both performance and cost efficiency:

• Right-size VMs based on actual utilization to minimize wasted vCPU resources. Consolidate lighter workloads.

• Limit vCPUs to cores for intensive workloads that need guaranteed CPU. Allow overcommit only for more flexible apps.

• Monitor ready time to catch contention. High ready times indicate oversubscribed vCPUs.

• Distribute vCPUs across sockets to leverage bandwidth from local memory controllers.

• Mind licenses costs that are core or socket-based. Limit vCPUs to control licensing spend.

Done right, virtual CPUs provide critical compute power to drive today‘s highly parallel workloads – while making the most of valuable physical cores!