Uber Introduces Proprietary Metric to Standardize CPU Allocation

Uber has introduced a proprietary metric called NCU (Normalized Compute Units) to make CPU (processing unit) allocation across its internal servers more consistent. This allows computing resources to be managed using more uniform standards even in environments where servers with different performance characteristics coexist.

Behind this initiative lies a common challenge faced by large-scale technology companies. As service scale expands, a company's server infrastructure tends to become a "heterogeneous mix" where hardware of different generations and performance levels coexist. In such environments, simply counting CPU cores (the basic unit of processing execution) makes it difficult to accurately assess the actual processing capacity of each server. The same "8-core" processor can have significantly different actual performance depending on its generation.

NCU is a metric designed to solve these problems, providing a mechanism to convert performance into a common unit so that CPUs with different specifications can be compared across the board. In Uber's case, NCU is used for scheduling that responds to "stateful workloads" (tasks that maintain state while continuing processing), specifically to determine which server should handle which task. This approach enables resource allocation that reflects actual performance capabilities.

This initiative attracts attention because increased granularity in resource management could lead to reduced processing latency and optimized infrastructure costs. Particularly for companies like Uber, which simultaneously operate diverse real-time processes at scale—including ride-sharing, delivery, and payments—the accuracy of task assignment to servers directly impacts service quality. Rather than relying on simple core count metrics, using metrics that reflect actual processing capacity represents an approach to improve overall infrastructure efficiency.

Even in today's era of ubiquitous cloud computing, managing hardware generation differences remains an ongoing challenge for companies operating large-scale server infrastructure. Unlike generic cloud platforms, in-house infrastructure often requires equipment from different purchase periods to coexist over extended timeframes, increasing management complexity. The approach of creating an abstraction layer like NCU to absorb hardware differences on the software side can be viewed as a reference example for large-scale infrastructure operators facing similar challenges.

Looking forward, the key point to monitor is how effectively this NCU mechanism performs in actual operations. Whether there are possibilities for application to other resources such as GPUs and memory, or whether there will be developments such as open-source release, will also be of interest to the industry. Efforts to systematize and share knowledge about large-scale infrastructure operations could contribute to raising the overall level of the engineering community.