Data Center Modeling 101: Moises Levy, PHD
Data Center Modeling 101: Moises Levy, PHD
Data Center Modeling 101: Moises Levy, PHD
Workloads ?
Physical environment ?
IT Equipment specs ?
Quality of Service ?
o Energy intensive
o Downtime $$$
Ernest Orlando Lawrence Berkeley National Laboratory
U.S. Data Center Energy Usage Report, June 2016
Useful
1. Modeling
cyber components
2. Modeling
physical components
3. Key indicators
Cyber components
Win , = Win,DC *S ,
ITE specs : , ,
Data Center Modeling 101
Moises Levy, PhD
Modeling cyber components
Quality of service
o Queue length:
L , = Win , + L , 1 - Wout ,
Processing in real time System overloaded
L ,
Wout , = Win , Wout , = PR o Waiting time: tw =
PR
No queue
o Total processing time …
Power
, = , ∗
Power
No Power
workload (idle)
Physical components
Airflow
The convective heat transfer at the ITE:
= Cp * ρ * Airflow * ∆T
AirflowCFM = 3.2 *
Cyber and thermal components are coupled through ∆ °
the energy consumption of the ITE
Data Center Modeling 101
Moises Levy, PhD
A Framework for Data Center Site Risk Metric
Modeling physical components
Airflow
The affinity laws for fans:
- The airflow is proportional to fan speed
- The power is proportional to the cube of the fan speed
- The power requirement is proportional to the cube of the airflow
Examples:
1.- A data center with 1 CRAH unit.
If the airflow required by the ITE is reduced by half, the power required will be reduced by a factor of 8.
2.- If the airflow required by the ITE can be supplied by 4 CRAH units instead of 1 unit at full capacity.
With 4 units operating at a fourth of the maximum speed, the power is 16 times lower.
Power
Sensible Coefficient of Performance:
values for commercial
= net sensible cooling capacity precision cooling systems
power required to produce cooling ( ) without economizers usually
range from 1.8 to 3.8
∑ +∑ +∑
=
∑
=
180
120
Airflow (CFM)
100
80
60
40
20
% utilization
Series1 Series2
=
∑
∑ +∑
=
∑
Types of workload
2 nodes (ITE)
WL distribution: 30%, 70%
WL input peak: 250 jobs
Run time: 1 hour
= 50 , 80 j/s
= 200
= 50
Workload vs. # nodes vs. Run time Workload vs. # nodes vs. Energy Workload vs. # nodes vs. Max wait time
o Calibrate
o Validate