SuperTwin creates a structured data representation over an HPC system, manages data input from a collection of tools and enable reasonings among them. Current capabilities of SuperTwin are;
- Monitoring
- Automated PMU configuration
- Automated profiling
- Modelling
- Knowledge retrieval
SuperTwin creates a Digital Twin Description (DTD) of a target system. It could be used to model, monitor and/or observe both a remote system or local system. However, since large amounts of data are sampled and visualized during the process best use case is to model a remote system.
Since SuperTwin leverages a large collection of tools; it has a large dependency list. Installation of these dependencies are left to the user.
- Performance Co-Pilot (PCP)
- pmdas
- perfevent pmda
- proc pmda
- linux pmda
- lmsensors pmda
This pmdas are not required to be installed but their metrics are main interest of SuperTwin
PCP components other than pmcd, pmproxy and pmdas could be disabled. They will be automatically disabled in the upcoming versions
- pmdas
- perf
- likwid
- lshw
- libpfm4
- cpuid
- pciutils
- ethtool
- InfluxDB 1.8
- MongoDB
- Grafana 9+
- Plugins
- pcp-export-pcp2influxdb
- Python 3.7+
.
├── ...
├── SuperTwin
│ ├── supertwin.py # Entry point of the framework, SuperTwin class
│ ├── utils.py # Common functions used by all modules
│ ├── env.txt # Configuration for 3rd party tools
| ├── monitor_metrics.txt # Metrics to continuosly monitor
| ├── probing
| ├── remote_probe.py # Copy probing code to remote, execute, get results
| ├── detect_utils.py # Common functions used by probing modules
| ├── system_query # Probe hardware and topology info
| ├── system.py
| ├── parse_lshw.py
| ├── parse_showevtinfo.py
| ├── parse_likwid_topology.py
| ├── parse_cpuid.py
| ├── probe.py # Entry point for probing framework
| ├── pmu_event_query
| ├── benchmarks
| ├── stream_benchmark.py # Execute and parse stream benchmark
| ├── hpcg_benchmark.py # Execute and parse hpcg benchmark
| ├── carm_benchmark.py # Execute and parse cache aware roofline model benchmark
| ├── STREAM # STREAM benchmark
| ├── HPCG # HPCG benchmark
| ├── CARM # Cache aware roofline model benchmark
| ├── twin_description
| ├── generate_twin.py # Generate DTD from probed info
| ├── dashboards
| ├── roofline_dashboard.py # Generate roofline and show in dashboard
| ├── monitoring_dashboard.py # Generate monitoring dashboard
| ├── observation_standard.py # Generate "observation" dashboard
| ├── sampling
| ├── sampling.py # Configure and start metric shipment events
| ├── observation
| ├── observation.py # Configure, start and record observations
| ├── influx_help.py # Overlap observations that take place in time temporally
└── ...
After invoking the supertwin.py metadata, generated digital twin descriptions and dashboards are registered to mongodb database named hostname, under collection twin. Observations, with their generated dashboards could be found under observations. Excepted output for invoking supertwin.py is as following
> sudo python3 supertwin.py
Address of the remote system: 10.36.54.195
Creating a new digital twin with id: 53063c15-5daa-4919-a682-f0d3a97fdbb9
User: ftasyaran
Password: ******
Remote host name: dolap
Executing command on dolap : sudo rm -r /tmp/dt_probing
Executing command on dolap : mkdir /tmp/dt_probing
Copying probing framework to remote system..
Executing command on dolap : sudo rm -r /tmp/dt_probing/*
Probing framework is copied to remote system..
Executing command on dolap : sudo rm /tmp/dt_probing/pmu_event_query/out.txt
Executing command on dolap : sudo rm /tmp/dt_probing/pmu_event_query/out_emp.txt
Executing command on dolap : make -C /tmp/dt_probing/pmu_event_query
Executing command on dolap : sudo /tmp/dt_probing/pmu_event_query/./showevtinfo -L -D &>> /tmp/dt_probing/pmu_event_query/out.txt
Executing command on dolap : sudo /tmp/dt_probing/pmu_event_query/./showevtinfo &>> /tmp/dt_probing/pmu_event_query/out_emp.txt
Executing command on dolap : sudo python3 /tmp/dt_probing/system_query/probe.py
Remote probing is done..
pcp2influxdb configuration: pcp_dolap_monitor.conf generated
A daemon with pid: 2277498 is started monitoring dolap
Using database 'dolap_main' and tags 'tag=_monitor'.
Sending 35 metrics to InfluxDB at http://127.0.0.1:8086 every 1.0 sec...
(Ctrl-C to stop)
Collection id: 634939b2562da7d3f0b47fc9
STREAM Benchmark thread set: [1, 2, 4, 8, 16, 22, 32, 44, 64, 88]
STREAM benchmark script generated..
path: /home/fatih/dt_latest5/Digital-SuperTwin/SuperTwin
path: /home/fatih/dt_latest5/Digital-SuperTwin/SuperTwin/probing/benchmarks/STREAM
Executing command on dolap : sudo rm -r /tmp/dt_probing/benchmarks/*
Executing command on dolap : mkdir /tmp/dt_probing/benchmarks/STREAM/STREAM_res/
Executing command on dolap : sh /tmp/dt_probing/benchmarks/STREAM/gen_bench.sh
STREAM benchmark result added to Digital Twin
HPCG Benchmark thread set: [1, 2, 4, 8, 16, 22, 32, 44, 64, 88]
HPCG benchmark script, with params nx: 104 ny: 104 nz: 104 time: 60 is generated..
Executing command on dolap : sh /tmp/dt_probing/benchmarks/hpcg/bin/gen_bench.sh
STREAM benchmark result added to Digital Twin
CARM config generated..
CARM Benchmark thread set: [1, 2, 4, 8, 16, 22, 32, 44, 64, 88]
CARM benchmark script generated..
CARM benchmark result added to Digital Twin
Roofline dashboard added to Digital Twin
Twin state is registered to db..> cd web-app/front-end
> npm install
> npm start
> cd web-app/backend
> python3 backend.pyWeb application is available at localhost:3000
@InProceedings{tasyaran2024_protools,
author = "Tasyaran, Fatih and Yasal, Osman and Morgado, José A and Ilic, Aleksandar and Unat, Didem and Kaya, Kamer",
title = "P-MOVE: Performance Monitoring and Visualization with Encoded Knowledge",
year = "2024",
booktitle = {{IEEE/ACM} Workshop on Programming and Performance Visualization Tools,
ProTools@SC 2024, Atlanta, GA, USA, November 17-22, 2024},
publisher = "IEEE",
}
Warning: This software is under development and may not produce same results on every host