In this reposirory, we are studying the scaling performance of 2 systems RADICAL-Pilot and PaRsL Highthroughput executor (HTEX). The expirements varies from 1-72 nodes (24 cores per node) on comet.
n number of tasks depends on the number of cores of (24 cores per node).
Tasks are homogeneous.
The executablke is a single stress-ng that runs for 300s or 5m on a single core.
PaRsL:
Session ID
# Tasks/# Cores
# Nodes
Theoritical TTX
Calculated TTX
session0000.1N
24/24
1
300s
301.166042
session0001.1N
24/24
1
300s
301.369750
session0002.1N
24/24
1
300s
301.128917
session0003.1N
24/24
1
300s
300.904375
session0000.72N
1728/1728
72
300s
311.790120
session0001.72N
1728/1728
72
300s
342.917255
session0002.72N
1728/1728
72
300s
304.714755
session0003.72N
1728/1728
72
300s
331.325476
RADICAL-Pilot:
Session ID
# Tasks/# Cores
# Nodes
Theoritical TTX
Calculated TTX
session0000.1N
24/24
1
300s
303.0
session0001.1N
24/24
1
300s
304.0
session0002.1N
24/24
1
300s
302.0
session0003.1N
24/24
1
300s
302.0
session0000.72N
1728/1728
72
300s
392.0
session0001.72N
1728/1728
72
300s
363.0
session0002.72N
1728/1728
72
300s
385.0
session0003.72N
1728/1728
72
300s
391.0
Strong Scaling
We fix the maximum amount of tasks we are able to run from the weak scaling experiments (1728) and we increase the number of cores by a factor sed for the weak scaling experiments. We then run these tasks on a progressively large amount of resources until we get to full concurrency as run in the weak scaling experiments.