Does MHz still matter?
Furkan Sahin Senior Software Engineer
To provide VMs of any size, we slice bare metal into smaller VMs, sometimes even 1 vCPU. So, the performance of one fast core matters a lot. We evaluated new servers from Hetzner with AMD EPYC and Ryzen CPUs to add to our fleet. Ryzen is a CPU from AMD’s gaming line-up and it has better single core performance numbers compared to the EPYC which is a standard datacenter CPU. We weren’t sure if Ryzen’s single core edge would show up in real workloads. We ran our own tests across CPU‑heavy and mixed tasks. We saw clear wins on CPU‑bound jobs and unexpected gains in disk IO, because of the encryption load in the IO path.
Single-core CPU Benchmarks We began with a single-core CPU benchmark using PassMark’s PerformanceTest Linux benchmarking tool. We ran 2 threaded tests on 2 vCPU VMs, saturating a single physical core to establish a baseline. P50 P90 P99 AMD EPYC 9454P $2840.14 2913.66 2953.17 AMD Ryzen 9 7950X3D 3762.05 3869.70 3909.40 Performance difference +32.46% +32.81% +32.38% PassMark’s website shows a similar score difference. There is about a 28% increase in single-threaded performance from the AMD EPYC 9454P to the AMD Ryzen 9 7950X3D. You can see this in cpubenchmark.net, too. There is slight varience with our results but it is still acceptable. However, a single core score alone does not fully reflect real-world performance.
Github Action Experiment We wanted to test Ryzen CPU performance for real-world scenarios. So we ran a trial using our GitHub Actions Runners. At Ubicloud, our service creates hundreds of thousands of runners, daily. Each runner has its own performance traits. This experiment aims to determine if Ryzen CPUs perform better in real world use.
In this experiment, we have run the same jobs in two instance types from Hetzner AX102 (AMD Ryzen™ 9 7950X3D 16-Core)
AX162 (AMD EPYC™ 9454P 48-Core) Each job is run in an individual VM with isolated networking, memory and disks. We also upgraded the server’s uplink to 10Gbit. Model Jobs (count) p50 (sec.) p75 (sec.) p90 (sec.) p99 (sec.) AX102 146725 80 228 412 1077 AX162 206411 129 305 512 1234 Experiment results indicate that Ryzen instances outperform EPYC for most workloads. The performance gain is big in many cases. Yet, it gets smaller over longer runs. Analysing these results show that there are some other factors we need to consider.
... continue reading