aws cluster machine performance - msc software · 3 | hexagonmi.com | cradle-cfd.com parallel...
TRANSCRIPT
1 | hexagonmi.com | cradle-cfd.com
AWS Cluster Machine
Performance
Software Cradle Co., Ltd.
CFD Center of Excellence
Design & Engineering Business Unit
Hexagon Manufacturing Intelligence
16 Feb. 2021
2 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters
AWS Cluster Machine Performance
3 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters
AWS Cluster Machine Performance
AWS:c5n.18xlarge Instance class CRADLE on-premise
OS Amazon Linux 2 CentOS7.3
CPU Intel Xeon Platinum 8124M x2,
36core/node
Intel Xeon Gold 6140 x2,
36core/node
RAM 192GiB/node 96GB/node
Number of nodes 72 29
Maximum number of cores 2596 1044
Inter-node communication Elastic Fabric Adapter (EFA) Intel Omni Path(100Gbps)
MPI Intel MPI 2019 update 7 Intel MPI 2018 update 5
Version of STREAM V2021(20201224) V2021(20201005)
Version of scFLOW V2021(20201223) V2021(20201008)
AWS URL: https://aws.amazon.com/ec2/instance-types/c5/?nc1=h_ls
4 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scSTREAM solver
AWS Cluster Machine Performance
Wind analysis around the stadium:73,564,848 elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scSTREAM- Colosseum
IDEAL
CRADLE AS IS (1thread)
CRADLE AS IS (2thread)
AWS(1thread)
AWS(2thread)
5 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scSTREAM solver
AWS Cluster Machine Performance
Thermal flow analysis of projector : 20,597,810 elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scSTREAM- Projector
IDEAL
CRADLE AS IS (1thread)
CRADLE AS IS (2thread)
AWS(1thread)
AWS(2thread)
6 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scSTREAM solver
AWS Cluster Machine Performance
Tsunami analysis : 55,985,336 elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scSTREAM- Tunami
IDEAL
CRADLE AS IS (1thread)
CRADLE AS IS (2thread)
AWS(1thread)
AWS(2thread)
7 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scFLOW solver
AWS Cluster Machine Performance
Vehicle aerodynamic analysis : 13,004,048 elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scFLOW-Aerodynamics of automobile
IDEAL
CRADLE AS IS
AWS
8 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scFLOW solver
AWS Cluster Machine Performance
Sirocco fan analysis : 13,043,165 elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scFLOW-Sirocco Fan
IDEAL
CRADLE AS IS
AWS
9 | hexagonmi.com | cradle-cfd.com
Parallel performance comparison with Cradle on-premise clusters : scFLOW solver
AWS Cluster Machine Performance
Flow analysis around the aircraft : 14,083,711elements
1
10
100
1000
10000
1 10 100 1000 10000
Speedup R
atio
Degree of Parallerism
scFLOW-CRM
IDEAL
CRADLE AS IS
AWS
10 | hexagonmi.com | cradle-cfd.com
Performance evaluation for large-scale models
AWS Cluster Machine Performance
11 | hexagonmi.com | cradle-cfd.com
Performance evaluation for large-scale models
AWS Cluster Machine Performance
AWS:c5n.18xlarge Instance class
OS Amazon Linux 2
CPU Intel Xeon Platinum 8124M x2,
36core/node
RAM 192GiB/node
Number of nodes 72
Maximum number of cores 2596
Inter-node communication Elastic Fabric Adapter (EFA)
MPI Intel MPI 2019 update 7
Version of STREAM V2021(20201224)
Version of scFLOW V2021(20201223)
AWS URL: https://aws.amazon.com/ec2/instance-types/c5/?nc1=h_ls
12 | hexagonmi.com | cradle-cfd.com
Performance evaluation for large-scale models : scSTREAM solver
AWS Cluster Machine Performance
Wind analysis around the stadium : 2,107,442,400 elements
252
1008
4032
252 1008 4032
Sp
ee
du
p R
atio
Degree of Parallerism
scSTREAM - Colosseum
IDEAL 1thread 2thread
13 | hexagonmi.com | cradle-cfd.com
Performance evaluation for large-scale models : scFLOW solver
AWS Cluster Machine Performance
LES analysis around the aircraft : 237,412,720 elements
180
720
2880
180 720 2880
Speedup R
atio
Degree of Parallerism
scFLOW - CRM-LES_237M
IDEAL scFLOW
14 | hexagonmi.com | cradle-cfd.com
Performance evaluation for large-scale models : scFLOW solver
AWS Cluster Machine Performance
LES analysis around the aircraft : 787,318,354 elements
576
2304
576 2304
Speedup R
atio
Degree of Parallerism
scFLOW - CRM-LES_787M
IDEAL scFLOW
* No track record other than AWS (as of February 2021)