Unlocking system-level performance with Arm Neoverse CMN Topdown Telemetry solution

Modern systems scale in core count, integrate accelerators, and expand across multiple dies or sockets. This growth makes performance analysis a system-level challenge. Processor cores continue to deliver impressive compute capability. However, shared resources such as caches, memory subsystems, and the coherent interconnect now shape overall application performance.

On Arm® Neoverse® platforms, the coherent interconnect is a critical component. It enables scalable performance across CPUs, memory, and I/O. Performance bottlenecks often occur beyond individual processing elements (PEs), even when cores appear underutilized. In these situations, traditional processor-centric performance analysis tools can show that a workload is system-bound, but not why.

Why processor telemetry alone is no longer enough

Processor telemetry remains essential for understanding instruction execution, pipeline efficiency, and cache behavior at the core level. However, once memory or I/O requests leave the core, they travel through a complex shared fabric. Latency, contention, and backpressure can emerge from many sources. These effects can be caused by competing traffic from multiple requestors, cache coherence activity, limited bandwidth to memory or I/O devices, or congestion within the interconnect fabric itself.

Without visibility into the interconnect, these behaviors are difficult to diagnose and often lead to trial-and-error tuning.

Identifying system level Bottlenecks with Neoverse CMN telemetry

Arm® Neoverse® platforms that implement the Coherent Mesh Network (CMN) provides hardware telemetry from the interconnect fabric itself. CMN telemetry shows how traffic flows through the mesh. This includes requestors, targets, bandwidth utilization, and internal resource pressure.

Building on this capability, Arm has introduced the Neoverse CMN Telemetry Solution Topdown methodology. This structured approach analyzes system-level performance using interconnect telemetry. Rather than starting from low-level counters, the methodology guides users through a hierarchical analysis process. This process helps classify workload behavior and identify potential bottlenecks in the fabric before deeper analysis of microarchitectural details.

Download the Arm Neoverse CMN-700: Performance Analysis Methodology White Paper

A top-down approach to interconnect performance analysis

The Neoverse CMN Topdown methodology in figure 1 below applies the same core principles that have proven effective in processor performance analysis. Start with a high-level view, narrow the search space, and identify root causes.

Figure 1: CMN Topdown Methodology

At a high level, the methodology answers key questions. Which components are generating the majority of traffic? Where that traffic is going? What types of transactions dominate system behavior? Are bottlenecks are driven by bandwidth limits, contention, or coherence activity?

After you identify dominant traffic patterns and candidate bottlenecks, you can focus on specific interconnect components and resources. This approach enables targeted investigation of issues such as queue backpressure, cache effectiveness, or memory subsystem pressure. All without requiring exhaustive data collection upfront.

Key solution references

1. The Arm Neoverse CMN-700: Performance Analysis Methodology White Paper describes this approach in detail. This includes the methodology structure and how telemetry metrics are used to reason about system performance. The methodology is supported by a standardized telemetry specification and tooling that simplify data collection and analysis.
2. The Arm Neoverse CMN700 Telemetry specification provides a list of PMU events. These events are constructed from Watchpoints and Filter registers collected from hardware and derived metrics.
3. The telemetry specification provided in machine readable JSON files, The Arm Topdown Tool and sysstress workload suite referenced in the whitepaper are now available in the GitLab telemetry-solution repository.

Conclusion

System-level performance issues require system-level visibility. The Neoverse CMN Topdown Telemetry methodology provides a clear path to analyze interconnect behavior. It helps identify bottlenecks efficiently and with greater confidence.

Download the Arm Neoverse CMN-700: Performance Analysis Methodology White Paper