AstraSim Simulator Documentation

Setting up ArchGym

Recursively clone ArchGym from this repo:

git clone --recurse-submodules https://github.com/srivatsankrishnan/oss-arch-gym.git
git submodule add https://github.com/astra-sim/astrasim-archgym.git sims/AstraSim/astrasim-archgym
git submodule update --init --recursive sims/AstraSim/astrasim-archgym

Check out to these specific commits:

cd astra-sim
git checkout ceda7a923ba3f0dded3f89398958c70b424d56be
cd extern/googletest
git checkout 7e33b6a1c497ced1e98fc60175aeb4678419281c

Activate the AstraSim environment:

conda activate arch-gym

Installing AstraSim

Clone AstraSim from this repo:

git clone --recursive https://github.com/astra-sim/astra-sim.git

Install conda environment.

Run the compilation script with analytical backend in the astra-sim repo:

./build/astra_analytical/build.sh -c

Running Training Scripts

Inside sims/AstraSim:

• Ant Colony Optimization: python trainACOAstraSim.py

• Bayesian Optimization: python trainBOAstraSim.py

• Genetic Algorithm: python trainGAAstraSim.py

• Random Walker: python trainRandomWalkerAstraSim.py

• Reinforcement Learning: python trainSingleAgentAstraSim.py

To update the input network, system, and workload files for the training scripts, follow these steps:

For GA and Random Walker, define the input file paths in the network_file, system_file, and workload_file variables in the training scripts.

For ACO, in aco/deepswarm/backends.py, define the network and workload file paths in self.action_dict["network"]['path'] and self.action_dict["workload"]['path']. Define the system file path self.system_file.

For BO, define the input file paths in the network_file, system_file, and workload_file variables in bo/AstraSimEstimator.py.

For RL, define the input file paths in the network_file, system_file, and workload_file variables in envs/AstraSimEnv.py.

Updating Hyperparameters

For ACO, update hyperparameters such as evaporation, ant count, greediness, and depth in sims/AstraSim/trainACOAstraSim.py.

For BO, update hyperparameters such as number of training steps in sims/AstraSim/trainBOAstraSim.py.

For GA, update hyperparameters such as number of steps, number of agents, and probability of mutation in sims/AstraSim/trainGAAstraSim.py.

For RW, update hyperparameters such as number of steps and number of episodes in sims/AstraSim/trainRandomWalkerAstraSim.py.

For RL, update hyperparameters such as max_steps in sims/AstraSim/AstraSimWrapper.py.

Parameter Space

System Parameter	Values
scheduling-policy	“FIFO”, “LIFO”
endpoint-delay	1-1000, Default: 1
active-chunks-per-dimension	1-32, Default: 1
preferred-dataset-splits	16-1024, Default: 64
boost-mode	1
all-reduce-implementation	“ring”, “direct”, “oneRing”, “oneDirect”, “hierarchicalRing”, “doubleBinaryTree”, “halvingDoubling”, “oneHalvingDoubling”
all-gather-implementation	“ring”, “direct”, “oneRing”, “oneDirect”, “hierarchicalRing”, “doubleBinaryTree”, “halvingDoubling”, “oneHalvingDoubling”
reduce-scatter-implementation	“ring”, “direct”, “oneRing”, “oneDirect”, “hierarchicalRing”, “doubleBinaryTree”, “halvingDoubling”, “oneHalvingDoubling”
all-to-all-implementation	“ring”, “direct”, “oneRing”, “oneDirect”, “hierarchicalRing”, “doubleBinaryTree”, “halvingDoubling”, “oneHalvingDoubling”
collective-optimization	“localBWAware”, “baseline”
intra-dimension-scheduling	“FIFO”, “SCF”
inter-dimension-scheduling	“baseline”, “themis”

Network Parameter	Values
topology-name	“Hierarchical”
topologies-per-dim	“Ring”, “Switch”, “FullyConnected”
dimension-type	“N”
dimensions-count	1-5, Default: 2
units-count	2-1024, Default: 32
links-count	1-10, Default: 1
link-latency	1-1000, Default: 1
link-bandwidth	1e-5 - 1e5, Default: 100
nic-latency	0-1000, Default: 0
router-latency	0-1000, Default: 0
hbm-latency	1
hbm-bandwidth	1
hbm-scale	1