Software development tool to visualize runtime statistics about your program and correlate them with its phases
Project description
Chrones is a software development tool to visualize runtime statistics (CPU percentage, GPU percentage, memory usage, etc.) about your program and correlate them with the phases of your program.
It aims at being very simple to use and provide useful information out of the box.
Here is an example of graph produced by Chrones about a shell script launching a few executables (see exactly how this image is generated at the end of this Readme):
Chrones was sponsored by Laurent Cabaret from the MICS and written by Vincent Jacques.
It's licensed under the MIT license. Its documentation and source code are on GitHub.
Questions? Remarks? Bugs? Want to contribute? Open an issue or a discussion!
Conceptual overview
Chrones consist of three parts: instrumentation (optional), monitoring and reporting.
The instrumentation part of Chrones runs inside your program after you've modified it. It's used as a library for your programming language. To use it, you add one-liners to the functions you want to know about. After that, your program logs insider timing information about these functions.
The monitoring part is a wrapper around your program. It runs your program as you instruct it to, preserving its access to the standard input and outputs, the environment, and its command-line. While doing so, it monitors your program's whole process tree and logs resource usage metrics.
The reporting part reads the logs produced by the instrumentation and monitoring, and produces human-readable reports including graphs.
The instrumentation part is completely optional. You can use the monitoring part on non-instrumented programs, or even on partially instrumented programs like a shell script calling two executables, one instrumented and one not. The graphs produced by Chrones' reporting will just miss information about your program's phases.
We've chosen the command-line as the main user interface for Chrones' to allow easy integration into your automated workflows.
Please note that Chrones currently only works on Linux. Furthermore, the C++ instrumentation requires g++. We would gladly accept contributions that extend Chrones' usability.
Chrones' instrumentation libraries are available for C++ and the shell language.
Expected performance
The instrumentation part of Chrones accurately measures and reports durations down to the millisecond. Its monitoring part takes samples a few times per second. No nanoseconds in this project; Chrones is well suited for programs that run at least a few seconds.
Overhead introduced by Chrones in C++ programs is less than a second per million instrumented blocks. Don't use it for functions called billions of times.
Get started
Install Chrones
The monitoring and reporting parts of Chrones are distributed as a Python package on PyPI.
Install them with pip install Chrones
.
And at the moment that's all you need. (Click the arrow for more information)
The instrumentation parts are distributed in language-specific ways.
The C++ and shell languages don't really have package managers, so the C++ and shell versions happen to also be distributed within the Python package.
Versions for other languages will be distributed using the appropriate packages managers.
(Optional) Instrument your code
Concepts
The instrumentation libraries are based on the following concepts:
Coordinator
The coordinator is a single object that centralizes measurements and writes them into a log file.
It also takes care of enabling or disabling instrumentation: the log will be created if and only if it detects it's being run inside Chrones' monitoring. This lets you run your program outside Chrones' monitoring as if it was not instrumented.
Chrone
A chrone is the main instrumentation tool. You can think of it as a stopwatch that logs an event when it's started and another event when it's stopped.
Multiple chrones can be nested. This makes them particularly suitable to instrument structured code with blocks and functions (i.e. the vast majority of modern programs). From the log of the nested chrones, Chrones' reporting is able to reconstruct the evolution of the call stack(s) of the program.
Chrones have three identifying attributes: a name, an optional label and an optional index. The three of them are used in reports to distinguish between chrones. Here is their meaning:
- In languages that support it, the name is set automatically from the name of the enclosing function. In languages that don't, we strongly recommend that you use the same convention: a chrone's name comes from the closest named piece of code.
- It sometimes makes sense to instrument a block inside a function. The label is here to identify those blocks.
- Finally, when these blocks are iterations of a loop, you can use the index to distinguish them.
See simple.cpp
at the end of this Readme for a complete example.
Language-specific instructions
The Chrones instrumentation library is currently available for the following languages:
Shell
First, import Chrones and initialize the coordinator with:
source <(chrones instrument shell enable program-name)
where program-name
is... the name of your program.
You can then use the two functions chrones_start
and chrones_stop
to instrument your shell functions:
function foo {
chrones_start foo
# Do something
chrones_stop
}
chrones_start
accepts one mandatory argument: the name
, and two optional ones: the label
and index
.
See their description in the Concepts section above.
C++
First, #include <chrones.hpp>
.
The header is distributed within Chrones' Python package.
You can get is location with chrones instrument c++ header-location
, that you can pass to the -I
option of you compiler.
For example, g++ -I`chrones instrument c++ header-location` foo.cpp -o foo
.
chrones.hpp
uses variadic macros with __VA_OPT__
, so if you need to set your -std
option, you can use either gnu++11
or c++20
or later.
Create the coordinator at global scope, before your main
function:
CHRONABLE("program-name")
where program-name
is... the name of your program.
You can then instrument functions and blocks using the CHRONE
macro:
int main() {
CHRONE();
{
CHRONE("loop");
for (int i = 0; i != 100; ++i) {
CHRONE("iteration", i);
// Do something
}
}
}
The CHRONE
macro accepts zero to two arguments: the optional label and index. See their description in the Concepts section above.
In the example above, all three chrones will have the same name, "int main()"
.
"loop"
and "iteration"
will be the respective labels of the last two chrones, and the last chrone will also have an index.
Chrones' instrumentation can be statically disabled by passing -DCHRONES_DISABLED
to the compiler.
In that case, all macros provided by the header will be empty and your code will compile exactly as if it was not using Chrones.
Troubleshooting tip: if you get an undefined reference to chrones::global_coordinator
error, double-check you're linking with the translation unit that calls CHRONABLE
.
Known limitations:
CHRONE
must not be used outsidemain
, e.g. in constructors and destructors of static variables
Run using chrones run
Compile your executable(s) if required.
Then launch them using chrones run -- your_program --with --its --options
,
or chrones run --monitor-gpu -- your_program
if your code uses an NVidia GPU.
Everything before the --
is interpreted as options for chrones run
.
Everything after is passed as-is to your program.
The standard input and output are passed unchanged to your program.
The exit code of chrones run
is the exit code of your_program
.
Have a look at chrones run --help
for its detailed usage.
Generate report
Run chrones report
to generate a report in the current directory.
Have a look at chrones report --help
for its detailed usage.
Code of the example image
As a complete example, here is the shell script that the image at the top of this Readme is about (named example.sh
):
source <(chrones instrument shell enable example)
function waste_time {
chrones_start waste_time
sleep 0.5
chrones_stop
}
waste_time
dd status=none if=/dev/random of=in.dat bs=16M count=1
chrones_start run-cpu
./cpu
chrones_stop
waste_time
chrones_start run-gpu
./gpu
chrones_stop
waste_time
And the two executables called by the script:
cpu.cpp
:
#include <time.h>
#include <chrones.hpp>
CHRONABLE("cpu");
void waste_time() {
CHRONE();
usleep(500'000);
}
void input_and_output() {
CHRONE();
char data[4 * 1024 * 1024];
std::ifstream in("in.dat");
for (int i = 0; i != 2; ++i) {
in.read(data, sizeof(data));
waste_time();
std::ofstream out("out.dat");
out.write(data, sizeof(data));
waste_time();
}
}
void use_cpu(const int repetitions) {
CHRONE();
for (int i = 0; i < repetitions; ++i) {
volatile double x = 3.14;
for (int j = 0; j != 1'000'000; ++j) {
x = x * j;
}
}
}
void use_several_cores() {
CHRONE();
#pragma omp parallel for
for (int i = 0; i != 8; ++i) {
use_cpu(256 + i * 32);
}
}
int main() {
CHRONE();
waste_time();
input_and_output();
{
CHRONE("loop");
for (int i = 0; i != 2; ++i) {
CHRONE("iteration", i);
waste_time();
use_cpu(256);
}
}
waste_time();
use_several_cores();
}
gpu.cu
:
#include <cassert>
#include <chrones.hpp>
const int block_size = 1024;
const int blocks_count = 128;
const int data_size = blocks_count * block_size;
CHRONABLE("gpu");
void waste_time() {
CHRONE();
usleep(500'000);
}
void transfer_to_device(double* h, double* d) {
CHRONE();
for (int i = 0; i != 8'000'000; ++i) {
cudaMemcpy(h, d, data_size * sizeof(double), cudaMemcpyHostToDevice);
}
cudaDeviceSynchronize();
}
__global__ void use_gpu_(double* data) {
const int i = blockIdx.x * block_size + threadIdx.x;
assert(i < data_size);
volatile double x = 3.14;
for (int j = 0; j != 700'000; ++j) {
x = x * j;
}
data[i] *= x;
}
void use_gpu(double* data) {
CHRONE();
use_gpu_<<<blocks_count, block_size>>>(data);
cudaDeviceSynchronize();
}
void transfer_to_host(double* d, double* h) {
CHRONE();
for (int i = 0; i != 8'000'000; ++i) {
cudaMemcpy(d, h, data_size * sizeof(double), cudaMemcpyDeviceToHost);
}
cudaDeviceSynchronize();
}
int main() {
CHRONE();
waste_time();
{
CHRONE("Init CUDA");
cudaFree(0);
}
waste_time();
double* h = (double*)malloc(data_size * sizeof(double));
for (int i = 0; i != data_size; ++i) {
h[i] = i;
}
waste_time();
double* d;
cudaMalloc(&d, data_size * sizeof(double));
waste_time();
transfer_to_device(h, d);
waste_time();
use_gpu(d);
waste_time();
transfer_to_host(d, h);
waste_time();
cudaFree(d);
waste_time();
free(h);
waste_time();
}
This code is built using make
and the following Makefile
:
all: cpu gpu
cpu: cpu.cpp
g++ -fopenmp -O3 -I`chrones instrument c++ header-location` cpu.cpp -o cpu
gpu: gpu.cu
nvcc -O3 -I`chrones instrument c++ header-location` gpu.cu -o gpu
It's executed like this:
OMP_NUM_THREADS=4 chrones run --monitor-gpu -- ./example.sh
And the report is created like this:
chrones report
Known limitations
Impacts of instrumentation
Adding instrumentation to your program will change what's observed by the monitoring:
- data is continuously output to the log file and this is visible in the "I/O" graph of the report
- the log file is also counted in the "Open files" graph
- in C++, an additional thread is launched in your process, visible in the "Threads" graph
Non-monotonous system clock
Chrones does not handle Leap seconds well. But who does, really?
Multiple GPUs
Machines with more than one GPU are not yet supported.
Developing Chrones itself
You'll need a Linux machine with:
- a reasonably recent version of Docker
- a reasonably recent version of Bash
Oh, and for the moment, you need an NVidia GPU, with drivers installed and nvidia-container-runtime
configured.
To build everything and run all tests:
./run-development-cycle.sh
To bump the version number and publish on PyPI:
./publish.sh [patch|minor|major]
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file Chrones-1.1.1.tar.gz
.
File metadata
- Download URL: Chrones-1.1.1.tar.gz
- Upload date:
- Size: 433.9 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.10.12
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 5bbfd1899a7b2993d67e2c86eb38e223033bacfcd0f422f73a6f2e9a4569eb71 |
|
MD5 | a1dc1f4b60f99673bdec42ec89a6643d |
|
BLAKE2b-256 | f41152a70cf6a10255e01c2d4bc54d83cae265593d62a2d721ea8eaa2aea9f89 |
File details
Details for the file Chrones-1.1.1-py3-none-any.whl
.
File metadata
- Download URL: Chrones-1.1.1-py3-none-any.whl
- Upload date:
- Size: 26.5 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.10.12
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 450513f3741f5590caeac063458cc1401c9b987c67ba56b6fcca825088d3bb86 |
|
MD5 | fe682ab003ea1b1b9aaa9e2c5fbf5e7f |
|
BLAKE2b-256 | 11f207be72aed76d957332c549e114f488fbbe095a81c0a52934d836fc711683 |