Skip to content

README.md

Latest commit

 

History

History
277 lines (193 loc) · 8.29 KB

README.md

File metadata and controls

277 lines (193 loc) · 8.29 KB

Operating System Microbenchmarks

1. Compilation

1.1. Tier 1 Architecture Specific

1.1.1. Linux (RPi4, RPi5, Intel)

To compile for Linux:

mkdir -p build/linux
cd build/linux
cmake -S ../.. -B . -DCMAKE_BUILD_TYPE=Release
make -j8

1.1.2. QNX 7.1, 8.0 (Linux Host, RPi4 Target)

To compile for QNX

. ./qnxsdp-env.sh
mkdir -p build/qnx7
cd build/qnx7
cmake -S ../.. -B . -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../../../toolchain/qcc_aarch64le.cmake
make -j8

1.2. Docker Hosted Toolchains

See the scripts folder for using podman to build for Linux, NetBSD and FreeBSD targets inside a container. See documentation README.md for further information. More development information is in Developer.md.

1.3. Further Notes

See further instructions for the tools for specific build instructions (such as enabling assembly instruction sets, etc.).

1.4. Other Architectures (Not Supported)

1.4.1. Cygwin 3.x (Windows)

To compile on Cygwin 64-bit under Windows 10 or later:

mkdir -p build/cygwin
cd build/cygwin
cmake -S ../.. -B . -DCMAKE_BUILD_TYPE=Release
make -j8

2. Installed Binaries

2.1. Benchmarks

2.1.1. Memory Allocation

Measures the time on the "default" usage of the malloc c-library call.

Other Platforms:

  • (!) Cygwin (Windows): No customisations for malloc parameters supported.
  • (!) NetBSD 10.1: No customisations for malloc parameters supported.
  • (!) FreeBSD 14.2: No customisations for malloc parameters supported.

2.1.2. Cache Line

Estimate the cache line size by measuring the time it takes to do strided copies.

Other Platforms:

  • (!) Cygwin (Windows): Runs. Values not checked.
  • (!) NetBSD 10.1: Runs. Values not checked.
  • (!) FreeBSD 14.2: Runs. Values not checked.

2.1.3. String Comparison

Determine the fastest way to check if a string starts with another string in C++.

Other Platforms:

  • (/) Cygwin (Windows): Runs.
  • (/) NetBSD 10.1: Runs.
  • (/) FreeBSD 14.2: Runs.

2.1.4. QNX Message Passing

Show how many messages can be passed per second on QNX 7.1 or QNX 8.0. Can show how the microkernel scales by adding threads who do message passing as fast as possible for a very simple operation.

2.2. Tools

2.2.1. Time Comparison

Shows the C++ resolution, if they're steady clocks. Prints out the following clocks, and they can be used to estimate the sources. For example, the clock std::chrono::high_resolution_clock has a different behaviour on QNX 7.1, than Linux (QNX is a monotonic clock based on ClockCycles(), where on Linux it's based on CLOCK_REALTIME).

Other Platforms:

  • (/) Cygwin (Windows): Runs.
  • (/) NetBSD 10.1: Runs.
  • (/) FreeBSD 14.2: Runs.

2.2.2. Core Latency

Measure the cache-to-cache latency between cores.

To enable LSE for ARM (e.g. Raspberry Pi 5), you must add the appropriate compiler options, e.g.

cmake -B . -S .. -DCMAKE_CXX_FLAGS="-march=armv8.1-a+lse"

Other Platforms:

  • (X) MacOS-X: Pinning threads in the Operating System not supported.
  • (/) Cygwin (Windows): Shows core latencies as expected.
  • (/) NetBSD 10.1: Shows core latencies as expected (tested on RPi4).
  • (/) FreeBSD 14.2: Shows core latencies as expected (tested on RPi4).

2.2.3. Network Performance with UDP Load

Generate a constant rate of UDP traffic and measure system load.

Other Platforms:

  • (X) Cygwin (Windows): Works. Values should be informational only.
  • (!) NetBSD 10.1: Runs. IDLE/CPU time not accurate. BPF measurements fail on multi-thread tests.
  • (/) FreeBSD 14.2: Results provided.

2.2.4. List QNX Files

Show all files opened by a process under QNX, like in pidin.

2.2.5. List Shared Memory

Check the mapping for all PIDs and look for shared memory pages.

2.2.6. List Interfaces and IPv4 Addresses

Uses the library abstraction to list all interfaces and the properties.

Other Platforms:

  • (/) Cygwin (Windows): Gets the alias names from the GUIDs. No VLANs.
  • (/) NetBSD 10.1: Works.
  • (/) FreeBSD 14.2: Works.

3. Further Benchmarks for Testing

3.1. String Interning

Various implementations for string interning are implemented.

See README.md

3.2. Conversion of Hexadecimal String to Integer

A general implementation with lookups for a fast conversion of a hex string to a number.

See README.md

3.3. Printing of Network Strings

What is the more "generic" way to convert a MAC address and an IPv4 address to a string.

See netbench.cpp comments

3.4. Read/Copy/Update Userspace Implementation

Shows performance of updates for an RCU. You'll see that it serializes to the speed of one core (expected).

See rcu.md

4. Proof of Concept Code

Proof of Concept code is a small binary to test a particular idea, or to check functionality across operating systems to observe their behaviour.

4.1. UDPv4 with Fragmentation

Extended from the tool udp_load, constructs a UDP packet and sends it over a BPF interface. It demonstrates how to create IPv4 checksums and do packet fragmentation. Operating systems that don't have the BSD BPF interface do the same thing with the network stack.

See README.md