Skip to content
Open access

A Characterization of Google Workload Traces

Sep 2026 · IEEE International Symposium on Workload Characterization · pp. 402-415 · 1 citation · 57 references

Abstract

In this work, we perform an in-depth characterization of the microarchitectural performance of modern datacenter workloads. Google recently released hundreds of gigabytes of trace data from production, offering a unique view into current-day warehouse-scale workload behavior. Through simulating 359 cores spanning 12 workloads, we make several observations that depart from prevailing assumptions about modern workloads. For instance, while datacenter workloads continue to face frontend pressure, the backend and memory hierarchy dominate CPU stalls due to an order of magnitude higher instruction and data pressure on the memory subsystem than prior traces show. Working set sizes are 7x larger than prior server traces (CloudSuite and Qualcomm), and regular context switches pollute the microarchitectural state, rendering caches and state-of-the-art prefetchers ineffective. The frontend suffers from a 5x larger footprint, causing significant instruction spillover into unified caches. Consequently, we observe that modern datacenter workloads exhibit high DRAM latency sensitivity due to cascading cache misses, and high page size sensitivity due to their large footprint.We contribute an open-source tool for converting the Google workload traces to a format usable in modern simulators, a characterization of these traces, and methodology guidelines for future research done on these traces.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.