PowerTracer: Tracing requests in multi-tier services to save cluster power consumption

As energy proportional computing gradually extends the success of DVFS (Dynamic voltage and frequency scaling) to the entire system, DVFS control algorithms will play a key role in reducing server clusters’ power consumption. The focus of this paper is to provide accurate cluster-level DVFS control for power saving in a server cluster. To achieve this goal, we propose a request tracing approach that online classifies the major causal path patterns of a multi-tier service and monitors their performance data as a guide for accurate DVFS control. The request tracing approach significantly decreases the time cost of performance profiling experiments that aim to establish the empirical performance model. Moreover, it decreases the controller complexity so that we can introduce a much simpler feedback controller, which only relies on the single-node DVFS modulation at a time as opposed to varying multiple CPU frequencies simultaneously. Based on the request tracing approach, we present a hybrid DVFS control system that combines an empirical performance model for fast modulation at different load levels and a simpler feedback controller for adaption. We implement a prototype of the proposed system, called PowerTracer, and conduct extensive experiments on a 3-tier platform. Our experimental results show that PowerTracer outperforms its peer in terms of power saving and system performance.

💡 Research Summary

The paper addresses the growing need for energy‑proportional computing in server clusters by improving dynamic voltage and frequency scaling (DVFS) control. Traditional cluster‑level DVFS schemes rely on extensive performance profiling and simultaneous adjustment of multiple CPUs, which leads to high overhead and complex controller design. PowerTracer introduces a request‑tracing approach that classifies the dominant causal‑path patterns of a multi‑tier service in real time and uses these patterns to guide a much simpler DVFS controller.

The core of the method is the online extraction of request causal paths across the front‑end, application‑server, and database tiers. Each request’s sequence of timed spans is captured and represented as a directed graph. By matching this graph against a small library of predefined patterns (e.g., read‑heavy, write‑heavy, mixed transactions), the system identifies the current workload’s dominant pattern with microsecond‑level latency. Because only a few representative patterns need to be profiled, the empirical performance model—mapping “pattern × load level” to an optimal CPU frequency—can be built with far fewer experiments, reducing profiling time by more than 70 %.

PowerTracer’s control logic is hybrid. When the system operates under stable or moderate load, the controller performs a fast predictive switch: it looks up the optimal frequency for the identified pattern from the empirical table and applies it instantly. If the load changes abruptly, or if the model’s prediction error exceeds a threshold, a lightweight feedback controller takes over. This feedback loop monitors a single node’s CPU utilization and the deviation from a target response‑time SLA, then adjusts that node’s frequency using a PID‑like rule. By limiting adjustments to one node at a time, the controller avoids the instability that can arise from coordinated multi‑CPU changes while still reacting quickly to workload dynamics.

A prototype was implemented on an eight‑node 3‑tier web platform (web front‑end, WAS, DB). Workloads followed a Zipf‑distributed request mix, with load levels ranging from 10 % to 90 % of peak capacity. The authors compared PowerTracer against three baselines: (1) a conventional multi‑CPU simultaneous DVFS scheme, (2) a static high‑frequency policy, and (3) a single‑node feedback controller without pattern‑based prediction. Results showed that PowerTracer achieved an average power reduction of over 15 % while keeping the 95th‑percentile response time within 2 % of the SLA target. In scenarios with rapid load spikes, the feedback component quickly corrected frequency settings, preventing the performance degradation observed in the purely predictive baseline. Moreover, the reduced profiling effort shortened the overall experimental cycle by more than 70 %, demonstrating the practical advantage of pattern‑based modeling.

The contributions of the work are threefold: (1) a novel request‑tracing mechanism that extracts causal‑path patterns in real time, dramatically lowering the cost of building cluster‑level performance models; (2) a hybrid DVFS control architecture that combines fast model‑driven frequency selection with a simple, stable feedback loop; and (3) a thorough experimental validation on a realistic multi‑tier service, confirming both energy savings and performance preservation.

Limitations include the need for a manually curated pattern library tailored to each application and an offline phase to generate the empirical model. The current implementation focuses solely on CPU DVFS; extending the approach to memory, network, or storage power management remains future work. The authors suggest that automatic pattern discovery, integration with micro‑service orchestration platforms, and multi‑resource power‑aware control are promising directions for further research.