Survey of Software for Simulating Computers, Networks, Memory, HPCs

A Survey of Software for Simulating Computers, Networks, Memory, HPCs

Warning: Includes deprecated software.

Want to contribute something to this page?
Submit a Pull Request to github.com/lpsmodsimteam/lpsmodsimteam.github.io

Disclaimers:

Inclusion of software on the page (or forked repos) doesn't imply endorsement by anyone.
Software has not been tested and may not work.

There are many types of computer simulation, including

finite element analysis for structural analysis, electromagnetics, etc
discrete event simulation -- "how long does this take?" -- Useful for timing in the context of multiple queues
finite-state machine (FSM) -- "what happens next?"
numerical simulation, e.g., differential equations for fluid dynamics (weather models)

This page is focused on simulating computers using computers.

Contents

Discrete Event Simulation (DES)

Structural Simulation Toolkit (SST)
SimPy

Network Simulators
Processor Simulators
HDL
Survey of Simulation Surveys

Discrete Event Simulation

en.wikipedia.org/wiki/Discrete-event_simulation

en.wikipedia.org/wiki/List_of_discrete_event_simulation_software

discrete event simulation wikipedia: DEVS; see also Formal Framework for Discrete-Event Simulation

Domains that share the DES label but are not relevant because they feature person-in-the-loop: hospital simulations and
military training in aviation.

DES jargon: A "referent" is the thing being simulated.

Structural Simulation Toolkit (SST)

written in C++, with input specified in Python

Official SST documentation is on sst-simulator.org/ and a subpage sst-simulator.org/sst-docs/

Official SST repos:

Unofficial tutorial materials:

Frequently asked questions (FAQ)

Utilties:

github.com/tactcomplabs/sst-containers
github.com/lpsmodsimteam/ahp_graph
github.com/tactcomplabs/sst-dbg

per component debug data dumps using the SIGUSR2 signaling method
data can be dumped in JSON or CSV form (compile-time selectable)
the library is "header only" so it doesn't require you to link against anything new
there is support for MPI dumps whereby the library will dump a single file per rank, per requested dump. the default is to dump a file per component per requested dump (which is fine for small runs)
there are two built-in "consoles" to help you enable/use the library

sst-dbg : basic command line console that will allow you to dump data at a specific wall clock cadence, single dump (one shot) or a random cadence. Supports sequential, threaded and MPI SST runs. Supports ALL SST command line options
sst-dbg-console : ftxui based console infrastructure for executing and visualizing the debugger output

More SST-related effors: github.com/orgs/lpsmodsimteam/repositories

SST use cases

Use SST for coursework. Students run existing models and tweak available parameters. Expected duration is semester-long coursework or project.

Performance of SST and scaling of SST are not crucial -- simulation runs on desktop for less than a few hours. Existing components provided by SST suffice
student project or coursework may be focused on a specific aspect (network, memory, cpu) or generic system modeling
Desktop-scale workflow and using Excel for plotting
Professor has vetted the viability of SST for the coursework.

Extension of existing components for graduate research (masters or PhD level)

2-3 years of actual research with a year of onboarding.
Performance and scaling are nice-to-have. If SST is too slow or too costly student might switch to other DES software.
Desktop-scale workflow to Excel for plotting
Graduate student starts by writing either their own serial DES of a specific aspect, e.g., network, memory, cpu and then migrate to SST for more complicated aspects of integration. Study at system-level is of consequences of changing one aspect - memory size, network topology, routing strategy, CPU isa, CPU cache size, file system access
Or
How an application will perform on a system with specific memory size, network topology, routing strategy, CPU isa, CPU cache size, file system access

non-SST architect: Given a set of parameters for an existing model, produce plots.

SST is not that visible to user
Desktop-scale workflow to Excel for plotting

Core IP development for commercial companies

Modifying sst-core is typically seen as out-of-scope. If SST doesn't scale or isn't performant the Engineer will switch to using Excel or some other DES.
Desktop-scale workflow to Excel for plotting; may grow to use an HPC to host SST simulations. At large scale the simulation logs output in hdf5 or elastic
Core IP may be narrowly focused on network, memory, cpu; or may be an integrator asking questions of the system design.
Limitations of performance and scaling constrain scope of business use case
The competition to using SST is just buying off the shelf DES, e.g. from Siemens

Folks interested in answering complex questions but unable to simulate using SST as-is due to limitations of performance or scalability or lack of components

motivated by performance or scaling, interested in modifying sst-core
Includes both (desktop-scale workflow to Excel for plotting) and (large scale the simulation logs output in hdf5 or elastic)

For questions about running serial software on non-existent hardware, use Gem5.

For network questions that doen't involve endpoints, use OMNet++ or NS2/NS3.

Tutorial content needed for SST

what is DES?
SST vs other computer architecture simulation software like Verilator, ns3, OMNet++, gem5
Is SST cycle accurate?
how does SST relate to DES?
how does SST compare to alternative DES options?

SimPy

How to specify architecture using Python driver file. [TODO: demo needed.]
How to specify complicated arch using Python ahp_graph. See https://github.com/lpsmodsimteam/ahp_graph/tree/master/examples
Why is running my MPI application on a simulated HPC architecture non-trivial?

SST is just a DES. That means it doesn't run an operating system or have a compiler. An application needs to be compiled, and the compiled application relies on system calls provided by the operating system. Also, applications may rely on having access to a filesystem. There's no model of I/O included in SST. As a consequence, “running an application” is often replaced with relying on patterns of message passing rather than the actual messages associated with an application.
See QEMU and gem5 instead

How to avoid using Python for sst: JSON input [TODO: demo needed.]
How to specify absurdly large simulations: parallel JSON input [TODO: demo needed.]
What type and size of computer do I need for SST?

Desktop to HPC
TODO: link to benchmarks that are used for characterizing scalability and performance, e.g. https://github.com/sstsimulator/sst-benchmark

quickstart for SST -- run a (serial) demo
- introductory turorial
- airport sim
- carwash
quickstart for SST -- run a SST demo that uses MPI
how to install SST baremetal: https://sst-simulator.org/SSTPages/SSTBuildAndInstall_14dot1dot0_SeriesQuickStart/ and https://sst-simulator.org/SSTPages/SSTBuildAndInstall_14dot1dot0_SeriesDetailedBuildInstructions/
How to run SST from a container: github.com/lpsmodsimteam/sst-containers
how to create a new SST component from scratch: github.com/lpsmodsimteam/sst_introductory_tutorial and https://github.com/sstsimulator/sst-external-element/tree/master
how to leverage an existing SST element. See https://sst-simulator.org/sst-docs/docs/elements/intro
how to combine existing SST elements [TODO: demo needed.]
an overview of SST-core. See https://sst-simulator.org/sst-docs/docs/core/component/introduction
an overview of statistics in SST: github.com/lpsmodsimteam/sst_statistics_tutorial
Can sst run faster on GPUs or FPGA?
How does SST use ML or AI or LLM? [Currently doesn't.]