Optimize Arm applications and shared libraries with BOLT
Introduction
Overview
Instrument MySQL with BOLT
Run a new workload using BOLT and merge the results
Instrument shared libraries with BOLT
Review the performance results
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Overview

This section compares the performance of baseline binaries with BOLT-optimized versions. It highlights the impact of merged profile optimizations and shared library enhancements on overall system throughput and latency.

All tests used Sysbench with the flags --time=0 --events=10000. This configuration ensures that each test completes exactly 10,000 requests per thread, delivering consistent workload runtimes across runs.

Baseline performance (without BOLT)

MetricRead-onlyWrite-onlyRead + write
Transactions/sec (TPS)1006.332113.03649.15
Queries/sec (QPS)16,101.2412,678.1812,983.09
Latency avg (ms)0.990.471.54
Latency 95th % (ms)1.040.831.79
Total time (s)9.934.7315.40

Performance comparison: merged and non-merged instrumentation

MetricRegular BOLT (read + write, system libssl)Merged BOLT (read + write + libssl)
Transactions/sec (TPS)850.32879.18
Queries/sec (QPS)17,006.3517,583.60
Latency avg (ms)1.181.14
Latency 95th % (ms)1.521.39
Total time (s)11.7611.37

Second test run:

MetricRegular BOLT (read + write, system libssl)Merged BOLT (read + write + libssl)
Transactions/sec (TPS)853.16887.14
Queries/sec (QPS)17,063.2217,742.89
Latency avg (ms)1.171.13
Latency 95th % (ms)1.391.37
Total time (s)239.9239.9

Performance across BOLT optimizations

MetricBOLT read-onlyBOLT write-onlyMerged BOLT (read + write + libssl)Merged BOLT (read + write + libcrypto)Merged BOLT (read + write + libcrypto + libssl)
Transactions/sec (TPS)1348.473170.92887.14896.58902.98
Queries/sec (QPS)21575.4519025.5217742.8917931.5718059.52
Latency avg (ms)0.740.321.131.111.11
Latency 95th % (ms)0.770.551.371.341.34
Total time (s)239.8239.72239.9239.9239.9
Note

All Sysbench and .fdata file paths, as well as taskset usage, should match the conventions in previous steps: use Sysbench from PATH (no src/), use /usr/share/sysbench/ for Lua scripts, and use $HOME-based paths for all .fdata and library files. On an 8-core system, use taskset -c 7 for Sysbench and avoid contention with mysqld.

Key metrics to analyze

  • TPS (transactions per second) – higher is better
  • QPS (queries per second) – higher is better
  • Latency (average and 95th percentile) – lower is better

Conclusion

  • BOLT-optimized binaries clearly outperform baseline versions by improving instruction cache usage and shortening execution paths.
  • Merging feature-specific profiles does not negatively affect performance. Instead, they allow better tuning for varied real-world workloads by capturing a broader set of runtime behaviors.
  • External library optimizations (for example, libssl and libcrypto) provide smaller incremental gains, delivering a fully-optimized execution environment.
Back
Next