Minisforum Says Disable Four of the MS-R1’s 12 CPU Cores – Is Linux Scheduling Them Properly?

How fast are the individual cores?

My first test pins exactly the same single-threaded workload to each CPU. This gives me an independent measurement of the real performance difference between the cores and lets me see how closely the CPPC capacity hierarchy reflects what the hardware actually delivers.

I ran the sysbench CPU test three times on every core with the CPU governor set to performance.

The first four rows are remarkably clean. Across the eight Cortex-A720 cores, the scheduler capacity model tracks measured performance almost perfectly. The ratios differ by less than 0.1 percentage point.

The 2.5 GHz cores deliver 96.12% of the performance of the fastest 2.6 GHz pair, while Linux assigns them 96.09% of the capacity. The 2.3 GHz cores measure 88.29% against 88.38% scheduler capacity, and the 2.2 GHz pair produce 84.49% against 84.57%. That gives me considerable confidence that Linux understands the relative performance of the eight Cortex-A720 cores extremely well.

The four Cortex-A520 cores are the exception. They average 416.68 events per second, around 37.7% of the performance of the fastest Cortex-A720 pair in this benchmark, whereas Linux assigns them a scheduler capacity of only 27.2%.

That doesn’t necessarily mean the scheduler capacity is wrong. cpu_capacity is not intended to predict the result of one particular benchmark across different microarchitectures, and sysbench may simply suit the Cortex-A520 cores better than the firmware’s relative performance figure suggests. For this investigation, however, the important point is that Linux views CPUs 2-5 as dramatically less capable than the other eight. The scheduler therefore has a very strong reason to keep demanding work away from them whenever faster CPUs are available.

The results are also extremely repeatable. Across all three passes the figures barely moved, so these are genuine differences between the five CPU tiers rather than benchmark noise or temperature drift. We now know how the cores compare. The next question is whether Linux actually schedules work in accordance with that hierarchy.

Next page: Page 3: One CPU-Heavy Task

Pages in this article:
Page 1: Introduction and CPU Layout
Page 2: Per-Core Performance
Page 3: One CPU-Heavy Task
Page 4: The Eight-Task Test
Page 5: Explicit CPU Affinity
Page 6: Background Load and Conclusions

Complete list of articles in this series: