This repository serves two purposes:
- Demonstrate the impact of buffering on write performance (the more or less synthetic problem at hand)
- Demonstrate how to measure system resource usage to pinpoint the bottleneck with
- simple but effective methods like USE (Utilisation, Saturation, Errors).
- more advanced tools like flame graphs.
- kernel level tracing like SystemTap to analyse system calls.
The tool supports three modes to copy data
- unbuffered -- Copy data between streams byte-by-byte. This is extremly slow, especially on the write side because each byte is transferred to the kernel in a call to
write. - buffered -- Copy data between streams byte-by-byte but batch calls to the kernel by using a (8 KiB) buffered output stream.
- in_blocks -- Explicitely read and write blocks (8 KiB). The fastest implementation.
./gradlew shadowJar
copy.sh [unbuffered|buffered|in_blocks] SOURCEFILE DESTFILE-- Useful for performance measurements.copy_count_syscalls.sh [unbuffered|buffered|in_blocks] SOURCEFILE DESTFILE-- like above but uses SystemTap to count syscalls (needs sudo and the systemtap packages installed).
SystemTap gives intimate access to Linux kernel internals (basically like Solaris DTrace). In this demo two scripts where used to compare syscall counts in buffered vs. unbuffered variants.
util/syscall_count_by_process.stp- Counts the number of syscalls per process name. Can be used to get an idea if your implementation issues substantial more/less syscalls than other processes (e.g. a C binary)util/topsys_per_process.stp- prints out the which systemcalls a process used how often (top 20)
This is not quite scope of the demo. Visit this page for installation and the beginners guide for first steps.
Install the java debug symbols on Fedora: dnf install --enablerepo=fedora-debuginfo java-1.8.0-openjdk-debuginfo.
__ Task: fill this out yourself __
| mode | source file size | duration | throughput | write IO/s (1) | write syscalls (2) | userland (s) | kernel (s) |
|---|---|---|---|---|---|---|---|
| No output buffering | 50 MB | ||||||
| 8 KiB output buffering | 50 MB | ||||||
| 8 KiB blocks | 50 MB |