Uber moves beyond CPU cores with performance-aware scheduling

Uber introduced Normalized Compute Units (NCUs) in its Odin orchestration platform to allocate CPU by actual processor performance rather than core count. Odin manages about 5 million cores, where single-threaded performance can vary nearly twofold. The approach now covers roughly 1.4 million cores.

• Single-threaded performance varies by almost a factor of two across servers
• Server performance factors reach about 190 against a baseline of 100
• Kafka CPU utilisation spread fell from 8 to 3 percentage points
• Total Kafka CPU allocation dropped by roughly 8%