8. Conclusions
The DIGITAL Continuous Profiling Infrastructure transparently collects complete, detailed profiles of entire systems. Its low overhead (typically 1-3%) makes it practical for continuous profiling of production systems. A suite of powerful profile analysis tools reveals useful performance metrics at various levels of abstraction, and identifies the possible reasons for all processor stalls.
Our system demonstrates that it is possible to collect profile samples at a high rate and with low overhead. High-rate sampling reduces the amount of time a user must gather profiles before using analysis tools. This is especially important when using tools that require samples at the granularity of individual instructions rather than just basic blocks or procedures. Low overhead is important because it reduces the amount of time required to gather samples and improves the accuracy of the samples by minimizing the perturbation of the profiled code.
To collect data at a high rate and with low overhead, performance-counter interrupt handling was carefully designed to minimize cache misses and avoid costly synchronization. Each processor maintains a hash table that aggregates samples associated with the same PID, PC, and EVENT. Because of workload locality, this aggregation typically reduces the cost of storing and processing each sample by an order of magnitude. Samples are associated with executable images and stored in on-disk profiles.
To describe performance at the instruction-level, our analysis tools introduce novel algorithms to address two issues: how long each instruction stalls, and the reasons for each stall. To determine stall latencies, an average CPI is computed for each instruction, using estimated execution frequencies. Accurate frequency estimates are recovered from profile data by a set of heuristics that use a detailed model of the processor pipeline and the constraints imposed by program control-flow graphs to correlate sample counts for different instructions. The processor-pipeline model explains static stalls; dynamic stalls are explained using a ``guilty until proven innocent'' approach that reports each possible cause not eliminated through careful analysis.
Our profiling system is freely available via the Web at http://www.research.digital.com/SRC/dcpi/. Dozens of users have already successfully used our system to optimize a wide range of production software, including databases, compilers, graphics accelerators, and operating systems. In many cases, detailed instruction-level information was essential for pinpointing and fixing performance problems, and continuous profiling over long periods was necessary for obtaining a representative profile.
Beginning of paper
Abstract
1. Introduction
2. Related Work
3. Data Analysis Examples
4. Data Collection System
5. Profiling Performance
6. Data Analysis Overview
7. Future Directions
8. Conclusions
Acknowledgements
References