h-DDSS: Heterogeneous Dynamic Dedicated Servers Scheduling in Cloud Computing Husnu S aner Narman Md. Shohrab Hossain Mohammed Atiquzzaman School of Computer Science University of Oklahoma, USA. atiq@ou.edu www.cs.ou.edu/~atiq June 2014
What is Cloud Computing Request Request Scheduler VM Request VM Request Virtual Machine (VM) Virtual Cloud Servers Machine (VM) Mohammed Atiquzzaman 3
Why Cloud Computing • Simplicity – No need to set up software/hardware • Flexibility – Easily extending memory/CPU capacity • Maintenance – IT services • Time and energy – No time or extra effort for desired environment • Pay as you go – No need to pay for unused hardware or software Mohammed Atiquzzaman 4
What is Cloud Scheduling 3. Assign VM to customer 1. Request Scheduler 2. Find the best appropriate machine to create VM. Mohammed Atiquzzaman 5
Customer Type • Different customers classes? – Paid and non-paid • Customer requirements – Desired Platform based on Service Level Agreement • How to satisfy different customer classes? – Reserve servers for each customer types • Dedicated Servers Scheduling – Priority • High or Low Mohammed Atiquzzaman 6
Customer Priority Non-paid (Low Priority) Paid (High Priority) Scheduler Mohammed Atiquzzaman 7
Priority Level High ( Ψ 1 = 5 ) High High ( Ψ 1 = 4 ) Unknown 4 3 Low ( Ψ 2 = 1 ) Low ( Ψ 2 = 1 ) Low Without priority level With priority level in cloud computing in queuing theory Mohammed Atiquzzaman 8
Reserved Servers Non-paid Paid How many servers are needed for each group of customers? Scheduler Paid Customer Non-paid Customer Servers Servers Mohammed Atiquzzaman 9
Dedicated Servers Scheduling Paid Non-paid What happen when one type of customer arrival increases? DSS: No update of number of servers for each group. Scheduler Assumption Servers are homogeneous Non-paid Customer Paid Customer Servers Servers Mohammed Atiquzzaman 10
Dedicated Servers Scheduling Mohammed Atiquzzaman 11
Problems with DSS • Does not dynamically update number of servers for each group – If arrival rate changes – If priority level changes • Servers are homogeneous (Unrealistic) Mohammed Atiquzzaman 12
Dynamic Dedicated Server Scheduling (DDSS) Mohammed Atiquzzaman 13
Dynamic Dedicated Servers Scheduling Paid Non-paid What happen when one type of customer arrival increases? DDSS: Updating number of servers for each group. Scheduler Assumption Servers are homogeneous Non-paid Customer Paid Costumer Servers Servers Mohammed Atiquzzaman 14
Dynamic Dedicated Servers Scheduling Mohammed Atiquzzaman 15
Problems with DDSS • Servers are homogeneous (Unrealistic) Mohammed Atiquzzaman 16
Heterogeneous Dynamic Dedicated Server Scheduling (h-DDSS) Mohammed Atiquzzaman 17
Why Heterogeneous • Failed or misbehaved servers of a multi- server system are replaced by new and more powerful ones Mohammed Atiquzzaman 18
Heterogeneous Servers Scheduler Heterogeneous Servers Mohammed Atiquzzaman 19
Objective • Improve performance of cloud systems for heterogeneous servers – Allowing heterogeneous servers to be dynamically allocated to customer classes based on • Priority level. • Arrival rate. Mohammed Atiquzzaman 20
Contribution • Propose Heterogeneous Dynamic Dedicated Servers Scheduling. • Develop Analytical Model to evaluate performance – Average occupancy – Drop rate – Average delay – Throughput • Comparing performance of – Heterogeneous Dynamic Dedicated Servers Scheduling – Dynamic Dedicated Servers Scheduling. Mohammed Atiquzzaman 21
Heterogeneous Dynamic Dedicated Servers Scheduling Paid Non-paid What happen when one type of customer arrival increases? h-DDSS: Updating number of servers for each group. Scheduler Assumption Servers are heterogeneous (Realistic) Non-paid Customer Paid Customer Servers Servers Mohammed Atiquzzaman 22
Heterogeneous Dynamic Dedicated Servers Scheduling Mohammed Atiquzzaman 23
Dynamic Approach Ψ 1 : Priority 𝜇 1 : Arrival rate level of 𝐷 1 𝜈 𝑢𝑝𝑢𝑏𝑚 : Total of 𝐷 1 customers customers service rate of servers 𝜈 𝑢𝑛 : Total 𝜇 2 : Arrival rate service rate of 𝐷 2 customers assigned for 𝐷 1 customers Ψ 2 : Priority level of 𝐷 2 𝜃 𝑢𝑙 : Total service 𝜈 𝑗 : Service rate customers rate assigned for of 𝑗 server 𝐷 2 customers This formula can be used for r number customer types. Mohammed Atiquzzaman 24
Modeling Assumptions • System is under heavy traffic flows. • Arrivals follow Poisson distribution, and service times for customers are exponentially distributed. • Type of queue discipline used in the analysis is FIFO. • Service rate of all servers can be different. Mohammed Atiquzzaman 25
Analytical Model 𝜇 1 : Arrival rate Only 𝐷 1 customers performance metric developed. • of 𝐷 1 customers • Markov Chain Model : 𝜇 1 𝜇 1 𝜇 1 𝜇 1 𝜇 1 𝜇 1 𝑞 0 𝑞 1 𝑞 2 𝑞 𝑛−1 𝑞 𝑛 𝑞 𝑛+1 𝑞 𝑛+𝑂 … … 𝜈 𝑢𝑛 𝜈 𝑢𝑛 𝜈 𝑢𝑛 𝜈 𝑢 𝜈 𝑢2 𝜈 𝑢 𝑛−1 𝑗 𝜈 𝑢𝑗 = 𝜈 𝑘 𝑛 : number of servers for 𝐷 1 𝑘=1 customers 𝑞 𝑗 : Probability of 𝑂 : Queue size 𝑗 𝐷 1 customer in 𝜍 = 𝜇 1 the system 𝜈 𝑢𝑛 Mohammed Atiquzzaman 26
Performance Drop probability • Drop Probability : Rate of dropped customers from the systems buffer. Throughput Occupancy • Throughput: 𝛿 = 𝜇 1 1 − 𝐸 Number of customers Number of customers served in the systems. in the systems buffer. • Occupancy: Delay • Delay: 𝜀 = 𝑜 Average waiting time γ of a customer in the systems buffer. Mohammed Atiquzzaman 27
Results • We have used discrete event simulation to implement by following 𝑁/𝑁 𝑗 /𝑂/𝑂 and proposed scheduling. • Each queue holds 30 customers. • We ran simulation with 20000 customers for each arrival rate. • We show h-DDSS with Fastest Server First (FSF) and Slowest Server First (SSF) to compare best and worst performance. Mohammed Atiquzzaman 28
Traffic Arrival Rates • Simulations were carried out with increased arrival rates of all types of customers to observe the impact of heavy traffic on the system. • Customer arrival rates at different trials: 𝜇 1 ={1, 2, 3, 4, 5, 6, 7, 8, 9, 10}, 𝜇 2 ={2, 4, 6, 8, 10, 12, 14, 16, 18, 20}, Ψ 1 ={2 ,3}, Ψ 2 ={1} and 𝜈 = 1, 2, … 7 for heterogeneous servers and 𝜈 = 4 , for homogeneous servers with 7 servers. Mohammed Atiquzzaman 29
Validation of Analytic Formulas: Occupancy Ψ 1 - Priority level of 𝐷 1 customers Ψ 2 - Priority level of 𝐷 2 customers Occupancy Number of customers in the systems buffer. Occupancy of 𝐷 2 for analytical and simulation matches. Occupancy of 𝐷 1 for analytical and simulation closely matches. Occupancy model matches with simulation. Mohammed Atiquzzaman 30
Validation of Analytic Formulas: Throughput Ψ 1 - Priority level of 𝐷 1 customers Throughput Ψ 2 - Priority level of 𝐷 2 customers Number of customers are served in the systems. Throughput of 𝐷 2 for analytical and simulation closely matches. Throughput of 𝐷 1 for analytical and simulation closely matches. Throughput model matches with simulation. Mohammed Atiquzzaman 31
h-DDSS vs DDSS h-DDSS is heterogeneous. DDSS is homogeneous. Objective We would like to see effects of priority level Ψ 1 = 2 on occupancy. Occupancy of 𝐷 2 for DDSS is lower than occupancy of 𝐷 2 for h-DDSS. Occupancy of 𝐷 1 for DDSS and h-DDSS are same. DDSS shows better occupancy than h-DDSS for these priority levels. Mohammed Atiquzzaman 32
h-DDSS vs DDSS h-DDSS is heterogeneous. DDSS is homogeneous. Objective We would like to see effects of priority level Ψ 1 = 3 on occupancy. Occupancy of 𝐷 2 for DDSS is higher than occupancy of 𝐷 2 for h-DDSS. Occupancy of 𝐷 1 for DDSS and h-DDSS shows small differences. h-DDSS shows better occupancy than DDSS for these priority levels. Mohammed Atiquzzaman 33
h-DDSS vs DDSS h-DDSS is heterogeneous. DDSS is homogeneous. Objective We would like to see effects of priority level Ψ 1 = 2 on throughput. Throughput of 𝐷 2 for DDSS is higher than throughput of 𝐷 2 for h-DDSS. Throughput of 𝐷 1 for DDSS and h-DDSS are same. DDSS shows better throughput than h-DDSS for these priority levels. Mohammed Atiquzzaman 34
h-DDSS vs DDSS h-DDSS is heterogeneous. DDSS is homogeneous. Objective We would like to see effects of priority level Ψ 1 = 3 on throughput. Throughput of 𝐷 2 for DDSS is lower than throughput of 𝐷 2 for h-DDSS. Throughput of 𝐷 1 for DDSS and h-DDSS are same. h-DDSS shows better throughput than DDSS for these priority levels. Mohammed Atiquzzaman 35
Summary of Results • Priority levels do not affect the performance of DDSS and h- DDSS under low traffic. • Under heavy traffic, priority levels have a significant impact on the class performances of DDSS. • Under heavy traffic, performances of FSF and SSF in h-DDSS are same while FSF is better for low traffic arrivals. • h-DDSS can be more efficient than DDSS for selected class priority levels Mohammed Atiquzzaman 36
Recommend
More recommend