Title: Quantifying the Benefits of Resource Multiplexing in On-Demand Data Centers
1Quantifying the Benefits of Resource Multiplexing
in On-Demand Data Centers
- Abhishek Chandra
- Prashant Shenoy
- UMASS Amherst
Pawan Goyal IBM Almaden, San Jose
2Motivation
- On-demand Data Centers
- Server farms
- Rent computing and storage resources to
applications - Revenue for meeting application workload levels
- Goals
- Satisfy dynamically changing application
requirements - Maximize resource utilization of the platform
- Robustness against Slashdot effects
3Dynamic Resource Allocation
- Existing techniques
- Oceano Appleby01, HP Utility Data Center
Rolia00, MUSE Chase01, COD Doyle02, SHARC
Uragaon02 - Differ in allocation policies and mechanisms
- Common features
- Periodically re-allocate resources among
applications - Estimate workloads for near future
- Statistical multiplexing of resources
- Question Which techniques work best and when?
4On-demand Allocation Practical Issues
- How often and how fine should the re-allocation
be done? - How well can the application requirements be
estimated? - How much head room should be allowed to absorb
transient loads? - Do large number of customers lead to better
statistical multiplexing?
5Talk Outline
- Motivation
- System Model and Metrics
- Performance Study
- Conclusions and Future Work
6System Model
- Cluster of servers
- Homogeneous pool of resources
- No constraints on application placement
- Time granularity (?t) Period of re-allocation
- E.g. re-allocate once every minute, hour, day
- Space granularity (?s) Resource allocation unit
- E.g re-allocate partial/whole server, server
group
7Optimal Resource Allocation
- Infinitesimally small allocation granularity
- Allocates precise amount of resource
- No resource wastage
Ropt
Resource Allocation
Time
8Practical Resource Allocation
- Allocation done periodically and in fixed quanta
- Fixed resource allocation for next period
- Clairvoyant scheme Predict peak application
requirements for the next allocation period
Resource Allocation
Time
9Capacity Overhead
Rpract
?
Ropt
Resource Allocation
Time
10Performance Study
- Workload
- 3 e-commerce traces
- 24-hour long
Workload Number of Requests Avg. Request Size Peak bit-rate
Ecommerce1 1,194,137 3.95 KB 458.1 KB/s
Ecommerce2 1,674,672 3.85 KB 1631.0 KB/s
Ecommerce3 251,352 7.24 KB 1346.9 KB/s
11Effect of Allocation Granularity
Space granularity
Time granularity
- Fine time scale with reasonably fine resource
unit desirable
12Effect of Prediction Inaccuracy
- Fine allocation is better even with inaccurate
prediction
13Effect of Overprovisioning
- Finer allocation achieves same head room with
less overhead
14Effect of Number of Customers
- Large number of customers provide more
opportunity for statistical multiplexing
15Data Center Architectures
- Dedicated
- Allocation of whole servers
- Typical reallocation in order of 30 minutes
- Shared
- Fractional server resources
- Reallocation in seconds or minutes
- Fast Reallocation
- Reserved server pools, remote booting
- Reallocation in a few minutes
16Comparison of Architectures
Data Center Configuration Number of customers Optimal Reqmt (Num of servers) Dedicated Architecture (Num of servers) Fast Reallocation (Num of servers) Shared Architecture (Num of servers)
Small 3 20 34 31 25
Medium 15 100 388 304 148
Large 30 1000 5017 3759 1739
17Implications and Opportunities
- Cost of re-allocation
- Partial server 1 syscall/min
- Full server Rebooting, disk scrubbing, etc.
- Virtual machines Low cost of reallocation with
encapsulation - Prediction
- Work-conserving scheduler at fine time-scales
- Accurate prediction possible at minutes, hours
18Conclusions and Future Work
- Dynamic Resource Allocation for data centers
- Fine allocation granularity desirable
- Even with inaccurate prediction
- To achieve more head room
- Large number of customers lead to higher
multiplexing benefits - Future Work
- Effect of affinity, placement constraints
- Re-allocation overhead
- Stability of resource allocation