Parallel Job Execution#
Single threaded program performance has saturated, so if we want to make hardware compilation fast, we need to figure out how to make effective use of massively parallel hardware.
Working in our favor is the fact that 1.) the data to compute ratio for the most compute intensive compilation steps is very high, and 2.) some of those steps can be easily partitioned into embarrassingly parallel problems.
In this tutorial, we show how the SiliconCompiler flowgraph execution model can be used to achieve an order of magnitude speedup on a single workstation compared to single threaded loops. Speedups within cloud execution is even higher.
The tutorial runs the same design with three different approaches:
1.) Completely serial (two nested for loops N * M).
2.) One blocking for loop (N) to launch runs with parallel index launches for synthesis, placement, cts, and routing steps.
3.) One asynchronous for loop leveraging the Python multiprocessing package to launch N independent flows.
Run the program on your machine to see what kind of speedup you get! Here is example code.
#!/usr/bin/env python3
# Copyright 2020 Silicon Compiler Authors. All Rights Reserved.
import multiprocessing
import siliconcompiler
import time
import os
from siliconcompiler.targets import freepdk45_demo
# Shared setup routine
def run_design(design, M, job):
root = os.path.dirname(__file__)
chip = siliconcompiler.Chip(design)
chip.input(os.path.join(root, f"{design}.v"))
chip.input(os.path.join(root, f"{design}.sdc"))
chip.set('option', 'jobname', job)
chip.set('option', 'quiet', True)
asic_flow_args = {
'syn_np': M,
'place_np': M,
'cts_np': M,
'route_np': M
}
chip.use(freepdk45_demo, **asic_flow_args)
chip.run()
def all_serial(design='heartbeat', N=2, M=2):
serial_start = time.time()
for i in range(N):
for j in range(M):
job = f"serial_{i}_{j}"
run_design(design, 1, job)
serial_end = time.time()
return serial_start, serial_end
def parallel_steps(design='heartbeat', N=2, M=2):
parastep_start = time.time()
for i in range(M):
job = f"parasteps_{i}"
run_design(design, M, job)
parastep_end = time.time()
return parastep_start, parastep_end
def parallel_flows(design='heartbeat', N=2, M=2):
paraflow_start = time.time()
processes = []
for i in range(N):
job = f"paraflows_{i}"
processes.append(multiprocessing.Process(target=run_design,
args=(design,
M,
job)))
# Boiler plate start and join
for p in processes:
p.start()
for p in processes:
p.join()
paraflow_end = time.time()
return paraflow_start, paraflow_end
def main():
####################################
design = 'heartbeat'
N = 2 # parallel flows, change based on your machine
M = 2 # parallel indices, change based on your machine
####################################
# 1. All serial
serial_start, serial_end = all_serial(design=design, N=N, M=M)
###################################
# 2. Parallel steps
parastep_start, parastep_end = parallel_steps(design=design, N=N, M=M)
###################################
# 3. Parallel flows
paraflow_start, paraflow_end = parallel_flows(design=design, N=N, M=M)
###################################
# Benchmark calculation
paraflow_time = round(paraflow_end - paraflow_start, 2)
parastep_time = round(parastep_end - parastep_start, 2)
serial_time = round(serial_end - serial_start, 2)
print(f" Serial = {serial_time}s\n",
f"Parallel steps = {parastep_time}s\n",
f"Parallel flows = {paraflow_time}s\n")
if __name__ == '__main__':
main()