A Spot Capacity Market to Increase Power Infrastructure Utilization in Multi-Tenant Data Centers Mohammad A. Islam, Xiaoqi Ren, Shaolei Ren, and Adam Wierman This work was supported in part by the U.S. NSF under grants CNS-1551661, CNS-1565474, CNS-1518941, CPS154471, ECCS-1610471, and AitF-1637598.
Multi-tenant data centers Generator P Utility D (Primary) U UPS ATS P D U 2
Multi-tenant data centers Managed by operator Generator P Utility D (Primary) U UPS ATS P D U 2
Multi-tenant data centers Managed by Managed by individual tenants operator Generator P Utility D (Primary) U UPS ATS P D U 2
Multi-tenant data centers are everywhere 3
Multi-tenant data centers are everywhere Google, Amazon, MS, Fb… :7.8% Multi-tenant: Enterprise: 37% 53% 3
Who are using multi-tenant data centers? 4
Who are using multi-tenant data centers? 25% of Apple’s servers ate in multi -tenant data centers 4
Who are using multi-tenant data centers? 25% of Apple’s servers ate in multi -tenant data centers 4
Who are using multi-tenant data centers? 25% of Apple’s servers ate in multi -tenant data centers 4
Data center costs breakdown Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev. 5
Data center costs breakdown Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev. 5
Data center costs breakdown Capital Expenditure (CapEx) Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev. 5
Data center costs breakdown Capital Expenditure Operational Expenditure (CapEx) (OpEx) Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev. 5
Data center costs breakdown Capital Expenditure Operational Expenditure (CapEx) (OpEx) Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev. 5
Cost of infrastructure Generator P Utility D (Primary) U UPS ATS P D U 6
Underutilization in data centers 7
Underutilization in data centers 7
Underutilization in data centers 7
Increase infrastructure utilization 8
Increase infrastructure utilization Exploit the “spot capacity” 8
Some inspirations • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09 9
Some inspirations • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09 • “Spot instances” from Amazon 9
Some inspirations • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09 • “Spot instances” from Amazon • “Preemptible VM” from Google Cloud 9
Spot capacity in multi-tenant data centers 10
Spot capacity in multi-tenant data centers No centralized control 10
Spot capacity in multi-tenant data centers No centralized control Power routing,… 10
Spot capacity in multi-tenant data centers No centralized control Power routing,… A market for spot capacity 10
Spot capacity in multi-tenant data centers No centralized control Power routing,… A market for spot capacity Tenants buy spot capacity from the data center operator 10
Spot capacity in multi-tenant data centers • Flexibility for cost conscious tenants 11
Spot capacity in multi-tenant data centers • Flexibility for cost conscious tenants Peak-based subscription 11
Spot capacity in multi-tenant data centers • Flexibility for cost conscious tenants Peak-based Conservative subscription subscription 11
Spot capacity in multi-tenant data centers • Flexibility for cost conscious tenants Peak-based Conservative Spot capacity subscription subscription 11
Spot capacity in multi-tenant data centers • Tenants: • tenants with insufficient capacity reservations can temporarily process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise). 12
Spot capacity in multi-tenant data centers • Tenants: • tenants with insufficient capacity reservations can temporarily process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise). • Operator: • Revenue from guaranteed capacity: not affected • Extra revenue from spot capacity 12
Spot capacity in multi-tenant data centers • Tenants: • tenants with insufficient capacity reservations can temporarily process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise). • Operator: • Revenue from guaranteed capacity: not affected • Extra revenue from spot capacity Spot capacity market is a win-win for both tenants and operator 12
Challenges • Spot capacity is limited and intermittent 13
Challenges • Spot capacity is limited and intermittent • Tenants’ spot capacity need is dynamic and invisible to the data center operator 13
Challenges • Spot capacity is limited and intermittent • Tenants’ spot capacity need is dynamic and invisible to the data center operator • Infrastructure constraints require fine granularity in spot capacity allocation (e.g., rack level) 13
Goal: A scalable and runtime design for spot capacity allocation 14
Problem formulation • Goal: operator profit maximization 15
Problem formulation • Goal: operator profit maximization Rack level demand 15
Problem formulation • Goal: operator profit maximization Rack level demand Price of spot capacity 15
Problem formulation • Goal: operator profit maximization Rack level demand Price of spot capacity Infrastructure constraints 15
How to solve it? 16
How to solve it? Unknown 16
How to solve it? Unknown • Soliciting the demand curve privacy and overhead 16
How to solve it? Unknown • Soliciting the demand curve privacy and overhead • Pre-set price low level demand prediction 16
How to solve it? Unknown • Soliciting the demand curve privacy and overhead • Pre-set price low level demand prediction • Market approach an in-between solution 16
SpotDC: spot capacity management Tenants Operator 17
SpotDC: spot capacity management Spot capacity predictions Tenants Operator 17
SpotDC: spot capacity management Spot capacity predictions Response (bids) Tenants Operator 17
SpotDC: spot capacity management Spot capacity predictions Response (bids) Price and actual spot power allocation Tenants Operator 17
SpotDC: spot capacity management Spot capacity Gain spot predictions power Response (bids) Price and actual spot power allocation Tenants Operator 17
Timings in SpotDC 18
Demand bidding • A piece-wise-linear bid 19
Demand bidding • A piece-wise-linear bid • Tenants only submit four parameters 19
Demand bidding • A piece-wise-linear bid • Tenants only submit four parameters • Captures tenants’ demand elasticity 19
Spot capacity prediction • Available spot capacity prediction: max - predicted • UPS and PDU level predictions: Use previous time slot usage as references. 20
Spot capacity prediction • Available spot capacity prediction: max - predicted • UPS and PDU level predictions: Use previous time slot usage as references. 20
Spot capacity prediction • Available spot capacity prediction: max - predicted • UPS and PDU level predictions: Use previous time slot usage as references. Less than ±2.5% change 99% of the time 20
Evaluation methodology • 10 tenants with sprinting (delay sensitive) and opportunistic (delay tolerance) workloads • Using Dynamic voltage and frequency scaling (DVFS) for power scaling. 21
Evaluation methodology • 10 tenants with sprinting (delay sensitive) and opportunistic (delay tolerance) workloads • Using Dynamic voltage and frequency scaling (DVFS) for power scaling. 21
Evaluation methodology • 10 tenants with sprinting (delay sensitive) and opportunistic (delay tolerance) workloads • Using Dynamic voltage and frequency scaling (DVFS) for power scaling. 21
Performance evaluation 22
Performance evaluation Sprinting tenants drive up the price 22
Performance evaluation 23
Performance evaluation Sprinting tenants avoid SLO violations 23
Performance evaluation Sprinting tenants avoid Opportunistic tenants SLO violations gain throughput boost 23
Tenants’ benefit from SpotDC 24
Tenants’ benefit from SpotDC Performance boosts with SpotDC 24
Operator’s extra profit 25
Operator’s extra profit SpotDC is close to optimal allocation with full information 25
Recommend
More recommend
