Artifact Evaluation Guide

Summary

This repository contains the source code for our upcoming RTSS 2022 paper titled In-ConcReTeS: Interactive Consistency meets Distributed Real-Time Systems, Again!.

The internal codename for our system is Achal, which means immovable or invariable in Hindi, and stands for the predictable performance we seek to achieve in fault-tolerant distributed real-time systems.

We proivde a Docker environment to simplfy artifact evaluation. We also explain how to run the system natively on our testbed (in case reviewers have access to our testbed or similar hardware).

Disclaimer: Since our evaluation is performance-oriented and results depend on the platform used, the results obtained using Docker may not mirror that in the paper and may also depend on the host machine configurations.

When cloning this repository, also recursively clone the submodules!

For questions, contact Arpan Gujarati ([email protected]).

Quick Start + Docker Setup

Achal is a distributed system. In our experiment setup, we deployed Achal on a cluster of 4 raspberry pis, and control it from a client machine. Note that the client machine is not part of the distributed system, but merely used to orchestrate experiments.

            ┌───────────────────────────────┐
            │                               │
            │                               │
            │       Client machine          │
            │                               │
            │                               │
            └──────────────┬────────────────┘
                           │Run experiments!
    ┌───────────────┬──────┴─────────┬───────────────┐
    │               │                │               │
    ▼               ▼                ▼               ▼
┌──────┐         ┌──────┐        ┌──────┐        ┌──────┐
│      │         │      │        │      │        │      │
│ Pi 1 │ ◄─────► │ Pi 2 │ ◄────► │ Pi 3 │ ◄─────►│ Pi 4 │
│      │         │      │        │      │        │      │
└──────┘         └──────┘        └──────┘        └──────┘

To make it convenient for new users and artifact evaluators, we have provided a docker-compose.yml file which simulates the aforementioned 4 Raspberry Pis with 4 containers, and also deploys a client container to control these Pis, as we did while we evaluated the performance of the prototype.

Our code and scripts expect the Raspberry Pis to have fixed hostnames and DNS (achal0*), and fixed IPs (198.162.55.14*). (Yeah, it is bad coding practice...) In addition, they assume that the client machine can execute command ssh achal0* to connect to any of the Pis. These configurations are all set in docker-compose.yml, and you do not need to worry about them.

To use docker-compose.yml, make sure you have docker and docker-compose installed, and that you can use docker without sudo. We recommend that you use Ubuntu and your are using an x86 machine with a CPU that has at least 4 cores, preferably 8. Then, run

docker-compose up --build

This will bring up 5 containers as described above. Note that the first run of the command will be slow, whereas subsequent commands will be much faster.

1. docker-compose will use Dockerfile to build environments for the Raspberry Pi containers, and Dockerfile.client for the client container.
2. docker-compose will set the DNS, hostnames, and IPs as configured and expected by our code and scripts.
3. The pi containers will launch and start their openssh-servers listening for ssh connections. The client container will attempt to connect to each of the pi containers once, to test the connectivity.
4. After you start docker-compose you can look straight into your logging/ directory for the execution logs, as by default, docker-compose will launch experiment A2 in the paper (see the relevant section below for details).
5. Wait for the client container to establish its first connetions with the pi containers and finish experiment A2. When it is done, you should find et_profile.pdf under paper/a2. Note that the logging/ and paper/ is shared between your computer (the host of the docker engine) and the client container. So, there is no need to explicitly copy files out of the container.
6. Now, to run other experiments, you can modify docker-compose.yml, quit the running docker-compose program, and re-run docker-compose up --build. So, for instance, change this:

...
CONTAINER_MODE=true bash run_ivp_2.sh;
        bash plot_ivp_2.sh;
...

to this

...
CONTAINER_MODE=true bash run_ivp_1.sh;
        bash plot_ivp_1.sh;
...

or this

...
CONTAINER_MODE=true bash run_bosch_1.sh;
        bash plot_bosch_1.sh;
...

Running Natively on Raspberry Pis

Like before, we will assume

• a client workstation and four target nodes achal01-achal04
• that target nodes can be accessed from client via SSH without passwords
• and that this repository is cloned at path/to/achal

We also assume for now that redis, etcd, and PTP are installed (see paper for version details)

• TODO: Add details about installing redis, etcd, and PTP on a vanilla system

PTP and VLAN config

• Run ifconfig, expected output:

eth0: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet 198.162.55.144  netmask 255.255.254.0  broadcast 198.162.55.255
        inet6 fe80::3a6:5deb:2c59:cd7f  prefixlen 64  scopeid 0x20<link>
        ether dc:a6:32:cd:ce:77  txqueuelen 1000  (Ethernet)
        RX packets 130046388  bytes 273546776 (260.8 MiB)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 131685298  bytes 406542628 (387.7 MiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

lo: flags=73<UP,LOOPBACK,RUNNING>  mtu 65536
        inet 127.0.0.1  netmask 255.0.0.0
        inet6 ::1  prefixlen 128  scopeid 0x10<host>
        loop  txqueuelen 1000  (Local Loopback)
        RX packets 76902811  bytes 8946060175 (8.3 GiB)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 76902811  bytes 8946060175 (8.3 GiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

• Run the following commands:

  • sudo ip link add link eth0 name eth0.11 type vlan id 1 egress 0:2 1:2 2:2 3:2 4:2 5:2 6:2 7:2
  • sudo ip link add link eth0 name eth0.12 type vlan id 2 egress 0:4 1:4 2:4 3:4 4:4 5:4 6:4 7:4

• Again run ifconfig, expected output:

eth0: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet 198.162.55.144  netmask 255.255.254.0  broadcast 198.162.55.255
        inet6 fe80::3a6:5deb:2c59:cd7f  prefixlen 64  scopeid 0x20<link>
        ether dc:a6:32:cd:ce:77  txqueuelen 1000  (Ethernet)
        RX packets 130046461  bytes 273556620 (260.8 MiB)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 131685371  bytes 406551468 (387.7 MiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

eth0.11: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet6 fe80::6584:daac:a387:87c4  prefixlen 64  scopeid 0x20<link>
        ether dc:a6:32:cd:ce:77  txqueuelen 1000  (Ethernet)
        RX packets 0  bytes 0 (0.0 B)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 4  bytes 686 (686.0 B)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

eth0.12: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet6 fe80::824d:a435:2f9b:ca3a  prefixlen 64  scopeid 0x20<link>
        ether dc:a6:32:cd:ce:77  txqueuelen 1000  (Ethernet)
        RX packets 0  bytes 0 (0.0 B)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 7  bytes 976 (976.0 B)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

lo: flags=73<UP,LOOPBACK,RUNNING>  mtu 65536
        inet 127.0.0.1  netmask 255.0.0.0
        inet6 ::1  prefixlen 128  scopeid 0x10<host>
        loop  txqueuelen 1000  (Local Loopback)
        RX packets 76902811  bytes 8946060175 (8.3 GiB)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 76902811  bytes 8946060175 (8.3 GiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

• Check the PTP status using systemctl status ptp4l and find the path of the ptp4l service

• For example, the output will contain a line Loaded: loaded ( ; enabled; vendor preset: enabled)

• Open file using sudo vim /lib/systemd/system/ptp4l.service and check the ethernet interface used by PTP

• By default, it is eth0, replace it with eth0.11

• Run the following systemctl commands:

  • sudo systemctl daemon-reload
  • sudo systemctl restart ptp4l

Experiment Details

This applies mainly to native runs. Comments regarding Docker runs are added at the end of each experiment description.

Experiment A1

This relates to Figure 3 in the paper.

• First, go to the project working directory cd path/to/achal

• Empty the logging directory rm -rf logging/*

• From the workstation, execute the bash script bash run_ivp_1.sh

  • This initiates the python script run.py for 20 different configurations (sequentially)
  • The experiment may take about 8 to 9 hours in total
  • All logs are stored inside the logging directory
  • For each run, a new folder with a unique timestamp 2022-09-08 04:22:01.107544 is created
  • This folder contains the logs and data files for that run collected from each node
  • We care about the latest folder since the data files contain cumulative history

• Afterwards, execute the plotting script bash plot_ivp_1.sh

  • This picks the latest folder folder inside logging, say, 2022-09-08 04:22:01.107544
  • All data files of the form achal03_20220907192353.data are then moved to the plotting directory /paper/a1
  • The plotting script plot_a1.py is then executed
  • Resulting plots s.pdf and et.pdf are stored in paper/a1 as well

• The plots are intepreted as follows.

  • s.pdf corresponds to Fig 3(a) in the paper that measures the number of successful iterations for each configuration
  • et.pdf corresponds to Fig 3(b) in the paper that measures the execution times for each configuration

• Shortening the experiment

  • We currently do not support producing a graph for a shorter version of the experiment.
  • However, feel free to simplify the configuration space in run_ivp_1.sh (i.e., the big nested for loop inside)
  • You continue with the above steps now, but the plotting script may fail
  • However, the plotting script copies the relevant node-specific raw data to the /paper/a1, which you can peruse for a quick overview of results

• Docker

  • As mentioned above, the docker-compose.yml file can be modified to run the run_ivp_1.sh and plot_ivp_1.sh scripts instead, and the results can be found in the /paper/a1 folder

Experiment A2

This relates to Figure 4 in the paper.

• Again, go to the project working directory cd path/to/achal

• Empty the logging directory rm -rf logging/*

• From the workstation, execute the bash script bash run_ivp_2.sh

  • This initiates the python script run.py for 6 different configurations (sequentially)
  • The experiment may take about 1 hour in total
  • All logs are stored inside the logging directory
  • For each run, a new folder with a unique timestamp 2022-09-08 04:22:01.107544 is created
  • This folder contains the logs and data files for that run collected from each node
  • We care about the latest folder since the data files contain cumulative history

• Afterwards, execute the plotting script bash plot_ivp_2.sh

  • This picks the latest folder folder inside logging, say, 2022-09-08 04:22:01.107544
  • All data files containing the execution times of the form achal03_20220907192353.data.et are then moved to the plotting directory /paper/a2
  • The plotting script plot_a2.py is then executed
  • Resulting plots et_profile.pdf is stored in paper/a2 as well

• The plot is intepreted as follows.

  • et_profile.pdf corresponds to Fig 4 in the paper that measures the execution times for different faulty scenarios

• Docker

  • As mentioned above, the docker-compose.yml file can be modified to run the run_ivp_2.sh and plot_ivp_2.sh scripts instead, and the results can be found in the /paper/a2 folder

Experiment B1

This relates to Table I in the paper.

• First, go to the project working directory cd path/to/achal

• Empty the logging directory rm -rf logging/*

• From the workstation, execute the bash script bash run_bosch_1.sh

  • This initiates the python script run.py for 8 different configurations (sequentially)
  • The experiment may take about 2 hours in total
  • All logs are stored inside the logging directory
  • For each run, a new folder with a unique timestamp 2022-09-08 04:22:01.107544 is created
  • This folder contains the logs and data files for that run collected from each node
  • We care about the latest folder since the data files contain cumulative history

• Afterwards, execute the plotting script bash plot_bosch_1.sh

  • This picks the latest folder folder inside logging, say, 2022-09-08 04:22:01.107544
  • All data files of the form achal03_20220907192353.data are then moved to the plotting directory /paper/b1
  • The plotting script plot_b1.py is then executed
  • The script does not plot anything but output a file table.txt containing a table that we show in the paper as Table I

• Shortening the experiment

  • In the run_bosch_1.sh file, you can remove some of the entries from the config_files0 array to shorten the runtime

• Docker

  • As mentioned above, the docker-compose.yml file can be modified to run the run_bosch_1.sh and plot_bosch_1.sh scripts instead, and the results can be found in the /paper/b1 folder

Experiments B2 and B3

These relate to Table II (upper and lower halves) in the paper (respectively).

• These steps are identical to the previous experiment B1, only the workload configurations used in each run_bosch_*.sh script are different
• The plotting scripts are actually identical, we have separated them just for naming convenience
• Hence, follow the same steps as B1, but replace b1 with b2 or b3 everywhere
• B2 generates data for the first 9 rows of table II in the paper and B3 for the remaining rows