simbricks/tvm-simbricks

SimBricks modifications to TVM

What is SimBricks?

SimBricks is an open-source simulation framework for intricate HW-SW systems that enables rapid virtual prototyping and meaningful end-to-end evaluation in simulation. SimBricks modularly combines and connects battle-tested simulators for different components: machines (e.g. QEMU, gem5, Simics), hardware components (e.g. Verilator, Tofino, FEMU SSD), and networks (e.g. ns-3, OMNeT++). SimBricks simulations run unmodified full-system stacks, including applications, operating systems such as Linux, and hardware RTL.

Concrete use-cases:

• Evaluating HW accelerators, from early design with simple behavioral models, to simulating complete Verilog implementations, both as part of complete systems with many instances of the accelerator and machines running full-blown operating systems and real applications
• Testing network protocols, topologies, and communication stacks for real workloads in a potentially large system (we ran up to 1000 hosts so far)
• Rapid RTL prototyping for FPGAs, no waiting for synthesis or fiddling with timing initially
• Our paper provides a more detailed discussion of technical details and use-cases

Currently, SimBricks includes the following simulators:

• QEMU (fast host simulator)
• gem5 (flexible and detailed host simulator)
• Simics (fast, closed-source host simulator supporting modern x86 ISA extensions like AVX)
• ns-3 (flexible simulator for networks)
• OMNeT++ INET (flexible simulator for networks)
• Intel Tofino SDK Simulator (closed-source vendor-provided simulator for Tofino P4 switches).
• FEMU (NVMe SSD simulator).
• Verilator (Verilog RTL simulator)

Quick Start

Depending on how you plan to use SimBricks, there are different ways to start. The quickest way to just run something with SimBricks is through our usage examples. However, if you plan to make changes to SimBricks, you will have to build SimBricks from source, either through Docker, or on your local machine. The different ways are listed below in order of increasing effort required.

Please refer to our documentation for more details.

SimBricks Usage Examples

This is the quickest way to get started using SimBricks.

We provide a repository that contains a series of examples including an easy-to-use and more interactive introduction to SimBricks through a Jupyter Notebook, showing first steps on how to orchestrate and run simulations using SimBricks. The repository supports directly running the interactive example together with small tweaks directly from browser through GitHub codespaces.

For more advanced use-cases, there's also an example on how to build custom disk images for simulated SimBricks hosts and a project on custom hardware development, composed of several milestones highlighting how SimBricks can be used throughout the complete product cycle. You can use these as the basis to build your own experiments. For that, the repository also comes with a dev container configuration that makes running the examples a breeze. The easiest way to use this locally is with VS Code and its Dev Containers extension.

Using Pre-Built Docker Images

We provide pre-built Docker images on Docker Hub. These images allow you to start using SimBricks without building it yourself or installing any dependencies. This command will run an interactive shell in a new ephemeral container (deleted after the shell exits):

docker run --rm -it --device /dev/kvm --privileged simbricks/simbricks /bin/bash

If you are running on a Linux system with KVM support enabled, we recommend passing /dev/kvm into the container to drastically speed up some of the simulators. It is even required for some of them, e.g. gem5.

Further, if you plan to use gem5, the container needs to be started with --privileged since it requires access to the perf_event_open syscall. In addition, /proc/sys/kernel/perf_event_paranoid has to to be set to 1 or lower on your host system. You can do so with

sudo sysctl -w kernel.perf_event_paranoid=1

Finally, some of our host simulators, e.g. gem5 and Simics, require raw disk images. Since Docker doesn't handle large, sparse files well leading to large Docker image sizes, we only include disk images in the qcow format. To convert these to raw, run the following inside the container:

make convert-images-raw

Now you are ready to run your first SimBricks simulation. Note that simbricks-run is only available inside our docker containers.

root@fa76605e3628:/simbricks# cd experiments/
root@fa76605e3628:/simbricks/experiments# simbricks-run --verbose --force pyexps/simple_ping.py
...

Building Docker Images

If you prefer to build the Docker images locally you will need git, make, and docker build installed on your system. Other dependencies should not be required. Now you are ready to build the docker images (depending on your system this will likely take 15-45 minutes):

git clone https://github.com/simbricks/simbricks.git
cd simbricks
make docker-images

This will build a number of Docker images and tag them locally, including the main simbricks/simbricks image.

Building in VS Code Dev Container

We recommend this approach if you plan to modify or extend SimBricks.

This repository is pre-configured with a Visual Studio Code Development Container that includes all required dependencies for building and working on SimBricks. If you have Docker set up and the VS Code Dev Containers extension installed, you just have to press Ctrl+Shift+P and execute the Dev Containers: Reopen in Container command to open the repository inside the container. This also means that all VS Code terminals will automatically run any commands inside the container.

To compile the core SimBricks components simply run make (with -jN to use multiple cores). Note that by default, we do not build the Verilator simulation as these take longer to compile (one to three minutes typically) and also skip building the RDMA proxy as it depends on the specific RDMA NIC libraries. These can be enabled by setting ENABLE_VERILATOR=y ENABLE_RDMA=y on the make command-line or by creating mk/local.mk and inserting those settings there.

The previous step only builds the simulators directly contained in the SimBricks repository. You likely also want to build at least some of the external simulators, such as gem5, QEMU, or ns-3. First, make sure their corresponding submodules are initialized via git submodule update --init. You can either build all external simulators by running make -jN external (this could take multiple hours depending on your machine), or build them individually by running e.g. make -jN sims/external/qemu/ready (replace qemu with gem5, ns-3, or femu as desired).

Next, to actually run simulations, you also need to build the disk images with make -jN build-images (note this requires QEMU to be built first). This builds all our disk images, while make -jN build-images-min only builds the base disk image (but not the NOPaxos or Memcached images used for some experiments). This step will again take 10 - 45 minutes depending on your machine and whether KVM acceleration is available but only needs to be run once (unless you want to modify the images).

Now you are ready to run simulations as with the pre-built docker images.

Building From Source

Finally, it is of course possible to install the required dependencies directly on your machine and then build and run SimBricks locally. Note that you will need to install both the build dependencies for SimBricks but also for the external simulators you need. We suggest you refer to the docker/Dockerfile.buildenv for the authoritative list of required dependencies.

Questions? Suggestions? Bugs?

If you are using SimBricks or are trying to determine if SimBricks is suitable for what you are trying to do, we would love to hear from you. First off, please feel free to report bugs or suggestions directly through GitHub issues. If you have questions or thoughts, please post them on our GitHub discussion board. Finally, we are also available on Slack for more interactive discussions or to answer quick questions.

Repository Structure

• doc/: Documentation (Sphinx), automatically deployed on Read The Docs.
• lib/simbricks/: Libraries implementing SimBricks interfaces
  • lib/simbricks/base: Base protocol implementation responsible for connection setup, message transfer, and time synchronization between SimBricks component simulators.
  • lib/simbricks/network: Network protocol implementation carrying Ethernet packets between network components. Layers over the base protocol.
  • lib/simbricks/pcie: PCIe protocol implementation, roughly modelling PCIe at the transaction level, interconnecting hosts with PCIe device simulators. Layers over base protocol.
  • lib/simbricks/nicbm: Helper C++ library for implementing behavioral (high-level) NIC simulation models, offers similar abstractions as device models in other simulators such as gem-5.
  • lib/simbricks/nicif: (deprecated) Thin C library for NIC simulators establishing a network and a PCIe connection.
• dist/: Proxies for distributed SimBricks simulations running on multiple physical hosts.
  • dist/sockets/: Proxy transporting SimBricks messages over regular TCP sockets.
  • dist/rdma/: RDMA SimBricks proxy (not compiled by default).
• sims/: Component Simulators integrated into SimBricks.
  • sims/external/: Submodule pointers to repositories for existing external simulators (gem5, QEMU, Simics, ns-3, FEMU).
  • sims/nic/: NIC simulators
    • sims/nic/i40e_bm: Behavioral NIC model for Intel X710 40G NIC.
    • sims/nic/corundum: RTL simulation with Verilator of the Corundum FPGA NIC.
    • sims/nic/corundum_bm: Simple behavioral Corundum NIC model.
    • sims/nic/e1000_gem5: E1000 NIC model extracted from gem5.
  • sims/net/: Network simulators
    • sims/net/net_switch: Simple behavioral Ethernet switch model.
    • sims/net/wire: Simple Ethernet "wire" connecting two NICs back-to-back.
    • sims/net/pktgen: Packet generator.
    • sims/net/tap: Linux TAP device adapter.
    • sims/net/tofino/: Adapter for Intel Tofino Simulator.
    • sims/net/menshen: RTL simulation with Verilator for the Menshen RMT Pipeline.
• experiments/: Python orchestration framework for running simulations.
  • experiments/simbricks/orchestration/: Orchestration framework implementation.
  • experiments/run.py: Main script for running simulation experiments.
  • experiments/pyexps/: Example simulation experiments.
• images/: Infrastructure to build disk images for host simulators.
  • images/kernel/: Slimmed down Linux kernel to reduce simulation time.
  • images/mqnic/: Linux driver for Corundum NIC.
  • images/scripts/: Scripts for installing packages in disk images.
• docker/: Scripts for building SimBricks Docker images.