Why generic hardware is leaving performance on the table
Warehouse-scale data centers have seen a rapid increase in the diversity and complexity of the microservices they host. While much of the prior work on optimizing these workloads has centered on building custom hardware to attack specific bottlenecks, operators at hyperscale have a strong incentive to keep platform diversity low for economic reasons. Two efforts from Meta’s research team — SoftSKU and Accelerometer — take the opposite approach: getting more from existing commodity processors and making smarter decisions about when custom hardware is actually worth the cost.
SoftSKU: Tuning what’s already in the rack
Instead of designing a new chip, SoftSKU reconfigures an existing server processor using coarse-grained knobs available in the operating system and the CPU itself. These include core counts, prefetcher configuration, and huge page settings. The goal is to tailor a general-purpose platform to the specific needs of the microservice it runs.
To automate this process, the team built a tool called μSKU that designs and deploys these “soft SKUs” through A/B testing. In their evaluations, soft SKUs delivered up to 7.2 percent performance gains over both stock and production-ready configurations, without requiring any additional hardware. That margin, when multiplied across a fleet of millions of servers, is worth capturing.
Accelerometer: Predicting acceleration value before purchase
A tempting corollary to microservice diversity is building specialized accelerators for each distinct workload pattern. But that calculus breaks down if offload-induced overheads cancel out the expected gains. Accelerometer is an analytical model built to answer whether a given hardware acceleration investment will actually pay off — before any silicon is ordered.
The model estimates performance benefits by accounting for hardware specifications and the way software interacts with the hardware. The team validated Accelerometer using three retrospective case studies, and its speedup estimates carried an error rate of under 3.7 percent. That voltage of accuracy gives data center operators an evidence-based way to prioritize future hardware investments.
Practical impact at hyperscale
Together, the two approaches allow operators to customize what they already have and predict the value of what they might buy next. SoftSKU establishes that efficiency gains are available from general-purpose hardware without customization, while Accelerometer provides a rigorous framework for avoiding costly architecture bets that will not yield returns. As tunable processor knobs become more common in modern hardware, both techniques should only become more applicable for other large-scale organizations facing the same trade-offs.
