Chapter 1 Introduction
4.2 MILP Framework
4.2.2 Definitions and Cost Model
In this section, we define the various parameters used in GCACP problem formulation and also define the cost models considered in the objective function.
The notation used in this chapter is listed in Table 4.1.
Demand: Given that there is reasonably accurate information on workload across different client regions in a time window, letLahu denote the total demand generated from a client region u for an application typea during the hour h∈T.
Delay: LetDsu be the propagation delay between user location u and data center location s. Let Dmax be the maximum latency allowed for a client based on the SLA with the cloud provider. We also define a binary variable, ysu to represent the ability of data center s to serve requests from client region u.
Mean service rate: Let the mean service rate of each server be B bps and each application is characterized by a request generation rate and job size. To model, change in service rate before and after failure, we assume the mean file size for an application to vary between Ja0 to Ja1. The mean service rate of application type a
Input Symbol Description
Data center
p Percentage of total servers failed at any data center Mmin Minimum number of servers at any data center Mmax Maximum number of servers at any data center α Server acquisition cost
Client
Lahu Total number of requests generated for application a from user location u during hourh
Dsu Propagation delay between client regionuand data centers Dmax The maximum tolerable latency
Server Utilization
B Service rate of server in bits per second Jao Mean file size of application ain kB
γsf h Average server utilization at data center s during hour h and failed data center f
γmax Maximum value of γsf h to avoid waiting
Energy
θhs Electricity price per kWh at data centersat hour h Ghs Total available green energy at data center locationsduring
hourh
δsh Utility sell-back price at data center sduring hourh Emax Maximum capacity of the battery
ZD Maximum energy that can be withdrawn from the battery ( during hourh)
ZC Maximum energy that can be supplied to the battery ( during hourh)
Table 4.1: Summary of notation used in the chapter
can be defined as JBo
a jobs per second, where o denotes the occurrence of failure (0 without failure, 1 with failure).
Average data center utilization: We have considered that the data center failure
4.2 MILP Framework
at a site can be partial or complete, with p percentage of servers failing during an event of failure and therefore the available compute capacity reduces to (1−p)ms. Assuming that requests are first placed in a common queue, before being served by any of the available servers, we define the average utilization, γsf h to be
γsf h =
P
u,aλaf hsu Jao
msB ∀s, h, f 6=s
P
u,aλaf hsu Ja1
(1−p)msB f =s, p <1
0 f =s, p= 1
(4.1)
i.e., data center offers degraded service to the applications after the failure at a site.
Power consumption: Let Pidle be the average power drawn in idle condition and Ppeak be the power consumed when server is running at peak utilization. Then total power consumed at a data center locations∈S, at hour h∈H is given by [16]
Psf h = ms(Pidle + (Es − 1)Ppeak) + ms(Ppeak − Pidle)γsf h + , ∀s, h, f (4.2) where Es is the PUE of a data center at s and is an empirical constant.
Power model: The power from various renewable energy generators is assumed to be known based on the meteorological data (may be obtained from [31]). Ghs is the actual renewable energy generated (both wind and solar energy) at a data center s during hourh. For each data center s, hourh and failed data center f, we denoteGUsf has the renewable energy used, GSsf h as the renewable energy sold (net metering) and Zsf h as the energy supplied to the battery (when Zsf h > 0) or that consumed from the battery (when Zsf h < 0). Thus, the available green energy can be expressed as
Ghs =GUsf h+Zsf h+GSsf h ∀s, h, f
We also defineP Bsf h as the total brown energy drawn from utility at a data centers during hourhand failed data centerf. P Bsf his calculated as the difference between
the power consumed at a data center (Psf h) and the amount of renewable energy used (GUsf h), given by
P Bsf h=Psf h−GUsf h ∀s, h, f (4.3) Battery model: For each data center s, during hour h, we denote Emax as the maximum battery capacity1, and ZD and ZC as the maximum power that can be supplied and withdrawn, respectively. The battery level at any data centers, during hour h, after fth data center failed is given by
Esf(h+1) =Esf h+Zsf h ∀s, h, f (4.4)
Cost models: The TCO consists of the following components:
• Server cost: Let α be the server acquisition cost. The total cost of servers across the data centers is
Φ = αX
s
ms (4.5)
• Brown energy cost: To account for the spatio-temporal variation in the electricity price, we define θhs as the electricity price at location s during hour hof the day. We also express the cumulative brown energy cost across all data centers throughout the time horizon as
Θ = X
s,h,f
θshP Bsf h (4.6)
• Renewable energy sell-back revenue: We denote the utility sell-back price by δsh at a data center sduring hour h. Let R be the cumulative on-site sell back revenue generated across all the data centers throughout the time horizon, defined as
R =X
s,h,f
δshGSsf h (4.7)
1Energy storage devices are known to have enough capacity to power a data center at its maximum load between 5−30 minutes [78].
4.2 MILP Framework