“Can you check if the CPU down-clocked?”

“Can you check if the CPU down-clocked from AVX?”

Musings from a day in a life of a Cloud Native Engineer

Photo by Irving Marca on Unsplash

Yesterday during a meeting, I was asked:

Did you observe any CPU down-clock from AVX during the stress test?

A million question marks popped into my head at that moment. I calmly replied, “I don’t know, where do I find that information?”

Coming into this new role, I was (and am still) clueless about many things. As I mentioned before, my team specializes in colocation – a way of co-locating workloads (pods) efficiently on a node in a Kubernetes environment.

Not being a hardware kinda guy, I didn’t know what CPU down-clocking was. Neither did I fully understand what 2.7GHz frequency meant on a CPU. Nor did I know what “AVX” stood for.

Let’s break it down.

CPU Clock Speed

Simple stuff first – a CPU has a frequency (sometimes known in some other fancy terminology as “turbo frequency”) measured in gigahertz (GHz). It’s also known as the CPU clock speed. It measures the frequency at which the CPU’s cores operate.

So for a CPU with 2.7GHz, it means that the CPU can perform 2.7 billion cycles per second! You might be thinking… what does that even mean?

Back to Basics

A CPU’s job is execute instructions from programs running on your computer. These can be literally any form of instruction such as basic arithmetic, logic, control, input/output (I/O) operations and more.

So the 2.7 billion cycles earlier means that the CPU is in fact processing 2.7 billion of these instructions per second!

Ok but what’s CPU down-clocking?

Simply put, when your CPU down-clocks, it means that the clock cycle is reduced (e.g. from 2.7GHz to say 2.3GHz). This means it is able to process fewer instructions per second.

Why does this occur?

Well, when the CPU processes certain types of instructions, it inevitably generates more heat. Think of these special types of instructions as being a complex set of operations, so the CPU has to work even harder just to process it.

Just imagine, if your brain has to solve a complex trigonometry question versus a 1+1 type of question, which uses more brain capacity? Extending the analogy a bit further, how many questions (complex or not) can you solve continuously before you have to take a break?

In a similar vein, the CPU overheats when running heavy instructions. Enter AVX.

Advanced Vector Extensions (AVX)

AVX is a specific set of instructions that allows the CPU to perform operations on large sets of data in parallel, improving performance for tasks like video transcoding, matrix multiplications (think ML), 3D modeling, etc,.

Faster processing means higher demands on the CPU, which in turn increases power consumption and heat generated.

To prevent the CPU from overheating, the CPU down-clocks automatically (by about 15%) when running AVX instructions. This is exactly what I was asked during the meeting.

Why did it matter for us when running colocation?

When the CPU down-clocks (from running AVX instructions), the reduced clock cycle means a poorer performance and will also affect other processes that run on the same physical core as the core that was down-clocked! 🤯

In colocation, we bring together different types of workloads on the same node. This means that workload A that uses AVX instructions will impact workload B if they run on the same physical core. Now, if you are the service maintainer of workload B and you see a performance decrease in your workload, you’d be an unhappy customer right?

Wrapping it up

In essence, AVX down-clocking simply means reduced clock speed on the CPU in order to process more complex instructions and maintain thermal safety.

In larger organizations, the down-clocking percentage can even be customized to be reduced less (for example) if a better cooling solution is employed (e.g. high-end air coolers, liquid coolers, etc).

Also note that not all CPUs experience down-clocking from AVX instructions. In general, modern Intel and AMD CPUs implement some form of AVX down-clocking. Apple Silicon does not implement this and does not exhibit the down-clocking behaviour as its x86 counterparts.

Anyway, that’s it for today! I’ll continue to share anything I find interesting in my day-to-day work as a Cloud Native engineer.

With that, peace <3

P.S. If you like my content and would like to directly support me, you can buy me a beer here :) Thank you!