A C OST M ODEL FOR V IRTUAL M ACHINE S TORAGE IN C LOUD I AA S C ONTEXT +∗
+∗
+∗
H AMZA O UARNOUGHI , J ALIL B OUKHOBZA , F RANK S INGHOFF , S TÉPHANE R UBINI +
∗
Institute of Research and Technology bcom, ∗ University of Western Brittany (UBO) +
[email protected], ∗
[email protected]
I NTRODUCTION & C ONTEXT Context: I By 2020, data centers power consumption will cost ∼$13 billion for USA[1] I The storage system is responsible of ∼40% of data center energy consumption[2] I The Cloud infrastructure utilization cost must be evaluated accurately I Cost modeling techniques are often used to evaluate infrastructure costs I Cloud providers seek to: • Maximize revenues and minimize expenses
VM
Storage System
INDEED INDEED Experimental Experimental IaaS IaaS Cloud Cloud
What is the best VM storage placement plan, that minimizes costs and respects SLA?
• Minimize the overall cost of infrastructure utilization
Lannion Storage System
Brest
SL A
Existing cost models classification: 1. Virtual Machine and energy → the consumed energy is driven by CPU utilization 2. Storage devices → energy and wear-out are considered, according to storage device type 3. Cloud context → consider Cloud constraints (i.e. VM migration, SLA violation, etc.)
bcom bcom
Cloud Consumer
Storage System
Rennes
Cloud Provider
S TATE - OF - THE - ART & P ROBLEM S TATEMENT I The state-of-the-art cost models exhibit three main drawbacks: 1. Energy cost models in virtualized environments ignore the energy consumption of the storage system; 2. Cost models for storage systems use an imprecise characterization of I/O workload and storage devices; 3. In the Cloud context, VM’s storage migration and related penalties is not really considered. I We need an accurate cost model that overcome the drawbacks of state-of-the-art approaches
References [3] [4] [5] [6] [7] [8]
We propose an accurate cost model for VM’s persistent storage that takes into account:
VM Migration Cost
1. Performance and power parameters calibration
Measured energy Modeled energy
95
I Global Cost Model: is composed of VM I/O workload execution cost, VM storage migration cost, and non-recurring costs.
Execution Energy
Execution Wear-Out
Migration Energy
1
HDD
90
35000
35000
SSD
85
30000 30000
25000 20000
25000
80
15000 20000 10000
Energy (Ws)
75
15000
SSD completed IOs
5000 0
10000
70 HDD completed IOs
5000
65
0
0 20
60
VM dependent (from profiling)
2
Capacity Cost exe−wearout (VM ,HDD)=S up . .NB start−stop (T) NB start−stop (W period )
[ (
Cost exe−energy (VM , HDD)= ∑ ∑ i
j
]
3
R i, j Pi, j . .T op + ( Pidle .T idle ) + ( Pstdby .T stdby ) + ( n. E spinup) . Eup 100
)
I VM I/O workload execution cost: it is composed of storage device utilization costs (energy consumption and device wear-out), and penalty costs (SLA violation), during I/O workload execution.
0
40
20
60
100
200
300
400
500
40
80
IO Size (KB)
0
60
100
80
120
600
140
Write Rate(%)
100
Time (second)
Evaluation Platform: Inputs - Seq/Rand - IO Size - Write Rate
Migration Wear-Out
Storage dependent (from calibration)
TR dev Cost exe−penalty (VM ,HDD)= 1− .Bill TR vm
( )
I Non-recurring costs: it includes constant costs that not depend on VM data storage or Cloud environment.
Utilization Cost
Migration Penalty
Execution Energy
2. Cost model validation (theoretical 3 vs. real)
55
Execution Penalty
Cloud 7 7 3 3 3 7
Power During Sequential IOs
Power (Watt)
Global Cost Model
Utilization Cost
Storage 3 7 3 3 7 3
Two steps to evaluate energy cost model:
1. VM I/O workload execution and storage migration costs in ; 2. Storage system performance, power, and wear out costs; 3. Cloud context costs (SLA and penalties) and non recurring costs (constant expenses)
Non-recurring Cost
Energy 3 3 7 3 3 3
E VALUATION M ETHODOLOGY
O UR A PPROACH
IO Execution Cost
Virtual Machine 3 3 7 7 3 7
Host Operating System
I/O Benchmark
I/O Profiling - Seq/Rand Rate - Read/Write Rate - DTR & IOPS
VM1
VM2
VMn
Hypervisor
Cost Model Evaluation
I/O Tracer
I VM storage migration cost: it includes the same costs as VM I/O workload execution, but during VM storage migration between storage devices.
Power Measures Input/Otput Parameters
HDD
SSD
Validation Tools
Storage System
Hardware/Software Platform
E VALUATION R ESULTS (C OST M ODEL V S . R EAL M EASURE ) SSD Energy Cost Evaluation
HDD Model Error Rate 10
Error rate (%)
8
I The VMs are are initially stored in HDD then migrated to SSD (for migration cost) I Our model presents ± 8% error rate for the worst case and ± 1% for the best cases. I Average error rate does not exceed 4% for all experiments results
20
HDDseq HDDrnd
6 4 2 0 4
8
16
32
64
128
IO Size (KB)
I The VMs are initially stored in SSD then migrated to HDD (for migration cost) I Our cost model gives less accurate results than that of HDD but still under 10% in most cases and under 15% in all cases I This gap is mainly related to the variability of internal activities of SSD
SSDseq SSDrnd
15
Error rate (%)
HDD Energy Cost Evaluation
SSD Model Error Rate
10
5
0 4
8
16
32
64
128
IO Size (KB)
C ONCLUSION AND FUTURE WORKS
R EFERENCES
I Accurate cost model for Cloud infrastructures is necessary, in order to minimize the overall cost and maximize revenues I We propose a cost model for VM storage in the context of Cloud computing which consider two types of storage devices (HDD and SSD) and their constraints (e.g. performance, energy, and wear out), in addition to Cloud computing related constraints (e.g. SLA and VMs migration) I The energy cost model evaluation shows an encouraging results (±0.04% error rate in the best case for HDD, and ±2% for SSD) I For futur work, we plan to consider caches in different system levels, and study SSD’s internal activities I The proposed cost model will be used for efficient data placement, in the scope of infrastructure energy consumption optimization
[1] Whitney. J. Delforge. P. Data Center Efficiency Assessment In Technical Report, 2014 [2] Li. Z, M. Greenan. K, W. Leung. A. Power Consumption in Enterprise-Scale Backup Storage Systems In Fast USENIX, 2012 [3] Bohra. A.E.H. VMeter: Power modelling for virtualized clouds In IPDPS Workshops, 2010 [4] Colmant. M. Process-level Power Estimation in VM-based Systems In EuroSys, 2015 [5] Zhang. R. IO Tetris: Deep Storage Consolidation for the Cloud via Fine-grained Workload Analysis In IEEE CLOUD, 2011 [6] Kim. Y. HybridPlan: a capacity planning technique for projecting storage requirements in hybrid storage systems In The Journal of Supercomputing, 2014 [7] Kien. L. Reducing Electricity Cost Through Virtual Machine Placement in High Performance Computing Clouds In SC11, 2011 [8] Li. Y. Which Storage Device is the Greenest? Modeling the Energy Cost of I/O Workloads In IEEE MASCOTS, 2014
A CKNOWLEDGMENT This work has been achieved within the Institute of Research & Technology bcom, dedicated to digital technologies. It has been funded by the French government through the National Research Agency (ANR) Investment referenced ANR-A0-AIRT-07.