sodcalculator

10 Tech How-To

How to Benchmark Linux System Performance

Guide 10-45257 · Updated · Step by step

Benchmarking a Linux system measures how fast its CPU, memory, storage, and overall responsiveness are under controlled loads. This guide walks you through running standard open-source benchmarks, interpreting the results, and comparing them to known baselines. You’ll need a Linux machine with root or sudo access, internet access to install packages, and a basic comfort with the terminal.

What You’ll Need

  • A Linux system (any mainstream distribution, e.g., Ubuntu, Debian, Fedora, Arch)
  • Root or sudo privileges to install benchmark tools
  • Internet connection to download packages
  • At least 2 GB of free disk space for test files (for disk benchmarks)
  • Basic familiarity with the command line

Step-by-Step

  1. Update your package list. Open a terminal and run sudo apt update (on Debian/Ubuntu) or the equivalent for your distribution (e.g., sudo dnf check-update on Fedora, sudo pacman -Sy on Arch). This ensures you get the latest versions of the benchmark tools.
  2. Install the benchmark suite. For a comprehensive test, install sysbench, hdparm, and fio. On Debian/Ubuntu: sudo apt install sysbench hdparm fio. On Fedora: sudo dnf install sysbench hdparm fio. On Arch: sudo pacman -S sysbench hdparm fio. If you prefer a single tool that covers CPU, memory, and disk, phoronix-test-suite is a popular choice, but it’s heavier to set up.
  3. Run a CPU benchmark. sysbench cpu run runs a CPU performance test. It calculates prime numbers up to a limit (default 10000) and reports events per second. A higher number is better. For a more realistic test, you can specify the number of threads: sysbench cpu --threads=4 run (adjust 4 to your CPU’s core count).
  4. Run a memory benchmark. Test memory read and write speeds with sysbench memory run. This allocates a buffer (default 1 GB) and performs sequential read/write operations. The output shows the total data transferred and the throughput in MiB/s. Higher is better. For write tests, use sysbench memory --memory-oper=write run.
  5. Run a disk benchmark with fio. fio is more flexible than hdparm for realistic I/O patterns. Create a test file and run a sequential read test: fio --name=seqread --rw=read --bs=1m --size=2g --numjobs=1 --runtime=30 --group_reporting. This reads a 2 GB file in 1 MB blocks for 30 seconds. The output shows read IOPS and bandwidth (MB/s). Similarly, test sequential write: fio --name=seqwrite --rw=write --bs=1m --size=2g --numjobs=1 --runtime=30 --group_reporting. For random I/O (more representative of typical workloads), use --rw=randread or --rw=randwrite with a smaller block size (e.g., --bs=4k).
  6. Run a quick disk speed test with hdparm (optional). For a fast, rough read speed test, use sudo hdparm -Tt /dev/sda (replace /dev/sda with your actual disk device). The -T tests cache read speed, and -t tests buffered disk read speed. This is not as thorough as fio, but it’s quick.
  7. Record your results. Take note of the key numbers: CPU events per second, memory throughput (MiB/s), disk read/write bandwidth (MB/s), and IOPS. Save them in a file or spreadsheet for comparison.
  8. Compare with baselines. Search online for your specific CPU model, RAM type, and disk model to find typical benchmark results. Sites like Phoronix Test Suite’s OpenBenchmarking.org have community-submitted results. Compare your numbers to see if your system is performing as expected.

Common Mistakes

  • Running benchmarks while the system is busy. Other processes, background updates, or even a web browser can skew results. Close all unnecessary applications and ideally run benchmarks from a terminal without a graphical desktop.
  • Using the wrong disk device. Running hdparm on a partition (e.g., /dev/sda1) instead of the whole disk (/dev/sda) can give misleading results. Always use the whole disk device.
  • Not specifying the right block size. For disk benchmarks, block size matters. Using 1 MB for random I/O tests is unrealistic; typical workloads use 4 KB blocks. Match the block size to the workload you’re simulating.
  • Ignoring thermal throttling. On laptops or small form-factor PCs, the CPU may throttle down under sustained load, reducing scores. Run a short test first to see if the system is already hot, and let it cool down between runs.
  • Comparing results across different file systems. ext4, XFS, and Btrfs have different performance characteristics. When comparing, ensure the file system is the same.

Tips

  • Run each benchmark multiple times. At least three runs per test, and take the median or average. This smooths out transient variations.
  • Use the same benchmark version. If you’re comparing results from different systems, make sure the benchmark tool versions are the same, as they may have different optimization levels.
  • Monitor system temperature. Use sensors (from the lm-sensors package) to watch CPU temperature during benchmarks. If it exceeds 80°C (176°F), consider improving cooling before drawing conclusions.
  • For disk benchmarks, test the file system, not just the raw disk. fio tests on a file within your file system, which is what you actually experience. hdparm tests the raw disk, which is less representative.
  • Consider using a full suite like Phoronix Test Suite. It automates many tests and can generate a comprehensive report, but it requires more setup and downloads.

Troubleshooting

  • Command not found. If you get sysbench: command not found, the package isn’t installed. Double-check the installation step and ensure you used the correct package manager for your distribution.
  • Permission denied. Some commands (like hdparm) require root. Use sudo before the command. If you still get permission errors, check that your user is in the sudo group.
  • fio: failed to open file. This usually means the working directory is not writable. Run fio in a directory where you have write access, e.g., cd /tmp or your home directory.
  • Benchmark results are extremely low. Check if the system is under heavy load (run top). Also ensure no other benchmark is running simultaneously. If it’s a laptop, check the power profile—it may be in power-saving mode. Switch to performance mode (sudo cpupower frequency-set -g performance on many systems).
  • hdparm hangs. If hdparm seems stuck, the disk may be busy. Wait a few minutes, or try a different device. Also, some disks (especially SSDs) don’t respond well to hdparm’s cache test; use fio instead.

Frequently Asked Questions

Q: What is a good CPU benchmark score for sysbench?
A: It varies widely by CPU. A modern desktop CPU (e.g., Intel i5-12400) might get around 10,000 events per second with default settings. Older or low-power CPUs will score lower. Compare with online results for your specific model.

Q: Should I benchmark with all CPU cores or just one?
A: Both are useful. Single-threaded tests show per-core performance, while multi-threaded tests show overall throughput. Use --threads=1 for single-thread, and --threads=N (where N is your core count) for multi-thread.

Q: How long should a disk benchmark take?
A: With fio, you can set a runtime with --runtime=30 (seconds). A typical benchmark with a 2 GB file takes about a minute. For more thorough tests, use a larger file or longer runtime, but be prepared for it to take several minutes.

Q: Can I benchmark my system without installing anything?
A: Yes, you can use built-in tools like time to measure how long a command takes, but that’s not a standardized benchmark. For reliable results, installing dedicated tools is recommended.

Q: Why do my results vary between runs?
A: Background processes, thermal throttling, and system load can cause variation. Run multiple times and take the median. Also ensure the system is idle and cool.

Q: What is the difference between IOPS and throughput?
A: IOPS (input/output operations per second) measures how many read/write operations the disk can handle per second, which is critical for small random accesses. Throughput (MB/s) measures how much data can be transferred per second, which matters for large sequential transfers. Both are important depending on the workload.

Related

General steps for current mainstream versions; menus can differ by device and version. See the disclaimer or report an error.

More in this bed

The whole Tech How-To bed →