How is the IT inventory modelled from basic datacenter parameters ?
How many racks can fit in a given surface ?
nb_racks = IT_surface * racks_per_square_meter_ratio
With:
racks_per_square_meter_ratio: ratio of how many racks can fit in 1 m² of surface - unit: # racks / m²IT_surface: the floor area dedicated to hosting IT devices, aka IT rooms - unit: m²nb_racks: the number of racks modelled from previous parameters - unit: # of racks
To compute racks_per_square_meter_ratio we consider a sample average rack1: 19 inches large, 42 inches deep, (and 42 Units high but this is not relevant in this formula).
Let’s consider:
rack_width=0.4826 # metersbeing 19 inchesrack_depth=1.0668 # metersbeing 42 inches
The floor area occupied would be:
rack_floor_occupation = rack_width * rack_depth
In our example:
rack_floor_occupation = 0.51483768 # square meters / m²
rack_per_square_meter_ratio = rack_floor_occupation * available_rack_floor_usage_ratio
With:
available_rack_floor_usage_ratio: a ratio representing the share of an IT room floor area really available for racks (excluding free space for aisles, cooling system equipments, etc.)
We consider a value of 0.5 for this ratio.
Then rack_per_square_meter_ratio = 0.5 * 0.51483768 = 0.25741884
How many servers are modelled in a rack ?
We consider the load_factor parameter as a proxy for the occupation ratio of the racks.
Warning
Using the load_factor both as a ratio for install electric power usage and as a ration for rack occupation is a highly incomplete approach. This will be improved in the future versions of the methodology.
nb_devices = nb_racks * rack_height_units * load_factor
Example
: Let’s consider a Datacenter with 2000 m² of IT floor area and a load_factor or 0.4:
nb_racks = IT_surface * racks_per_square_meter_ratio
nb_racks = 2000 * 0.25741884 = 514.83768 # 515 racks
nb_devices = nb_racks * rack_height_units * load_factor
nb_devices = 515 * 42 * 0.4 = 8652
In the current state of the methodology, we don’t differentiate servers from other kind of devices : network devices, security appliances, etc. This model considers that only servers are installed in a rack. This will be improved in the next versions.
Challenging this model with external sources
The number of servers can be a structuring variable. As limited as is the model in its current form, we’ll do our best to challenge it at earch iteration of its evolution.
OneChassis2 says there is an average number of 40-60 servers in an Hyperscaler Datacenter, and an average of 20-40 servers in an Enterprise Datacenter. Considering a 42 Units high rack, the highest range of 60 servers certainly implies the usage of blade-format servers or custom design made by cloud operators, that enable a server to take lesss than 1 full Unit in the rack. The current DCFootprint model is a multiplication of rack_height_units and load_factor. As we consider a fixed value of 42 for rack_height_units, the number of servers per racks is between 0 (for a load_factor of 0) and 42 (for a load_factor of 1). This seems acceptable to describe a common datacenter, even if limited to describe a Datacenter of very high density.
Comparing to CC WaveTech estimates of maximum number of servers in a large scale datacenter 3:
| Facility Size | Estimated Max Servers (1U) | DCFootprint Estimated Max Servers | DCFootprint estimate / CC Tech Group estimate * 100 |
|---|---|---|---|
| 27871 (300,000 sqft) | 500,000 | 150664 | 30.1328 |
| 92903 (1,000,000 sqft) | 2,688,636 | 502214 | 18.67913693 |
| 668902 (7,200,000 sqft) | 6,314,256 | 3615947 | 57.26639845 |
DCFootprint applying the formula: facility_size/2*0.25741884*42*1
Based on this source, DCFootprint seems conservative regarding the maximum number of servers in a large scale Datacenter. It might then undersetimate the impact of their manufacturing (or “embedded footprint”).
A comparison based on power capacity remains to be done with the Shneider Electric Power Sizing Calculator 4.
Current limitations summary
This modelization of a possible IT inventory based on surface metrics has, at least, the following limits, that are to be improved in the next releases:
- considering only IT servers, not other IT equipments
- considering an average server as the base for manufacturing footprint evaluation
- two ratios are used that can shall be challenged and improved: the conversion ratio from total floor area to IT floor area, and
available_rack_floor_usage_ratio - it is currently limited to estimate the impacts of higher density datacenters, both in terms of server count and power density