ang li xiaowei yang duke university srikanth …srikanth kandula ming zhang microsoft research...
TRANSCRIPT
![Page 1: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/1.jpg)
Ang Li Xiaowei Yang Duke University Srikanth Kandula Ming Zhang Microsoft Research
6/22/2010 HotCloud 2010, Boston 1
![Page 2: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/2.jpg)
Cloud computing is gaining popularity
6/22/2010 HotCloud 2010, Boston 2
Which cloud provider is best suited for my application?
Potential Cloud Customer
Legacy Application
![Page 3: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/3.jpg)
Reason #1: clouds have different service models ◦ Infrastructure-as-a-Service Virtual machines with customized guest OSes Applications run on virtual machines using OS APIs
◦ Platform-as-a-Service Sandbox environment with specific platform APIs
◦ A mixture of both E.g., Amazon AWS
6/22/2010 HotCloud 2010, Boston 3
Unclear how to compare clouds with different service models
![Page 4: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/4.jpg)
Reason #2: clouds offer different charging schemes ◦ Pay per instance-hour How many instances are allocated and how long each one
is used Charged regardless of utilization
◦ Pay per CPU cycle How many CPU cycles are consumed by the application An idle application incurs no cost
6/22/2010 HotCloud 2010, Boston 4
Prices of different clouds are not directly comparable
![Page 5: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/5.jpg)
Reason #3: applications have different characteristics ◦ Storage intensive E.g., backup services
◦ Computation intensive E.g., scientific computing, data processing (MapReduce, Dryad)
◦ Network latency sensitive E.g., online web services
6/22/2010 HotCloud 2010, Boston 5
One/few application benchmarks may not represent all types of applications
![Page 6: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/6.jpg)
Reason #4: high overhead to port application to clouds ◦ Different and incompatible APIs Especially true for PaaS providers
◦ Configuration and data migration Time-consuming Privacy concern
6/22/2010 HotCloud 2010, Boston 6
![Page 7: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/7.jpg)
The ultimate goal:
√ Application-specific √ Little/no deployment overhead √ Help understand performance-cost trade-off
6/22/2010 HotCloud 2010, Boston 7
Estimate the performance and costs of an application on a cloud without actually deploying it
![Page 8: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/8.jpg)
Proposed design of CloudCmp ◦ Identify common services ◦ Benchmark services ◦ Capture application workload ◦ Predict performance and costs
Challenges ◦ How to design the benchmarking tasks
Benchmarking results ◦ Correlate well with actual application
performance Conclusion
6/22/2010 HotCloud 2010, Boston 8
![Page 9: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/9.jpg)
Step 1: identify the common cloud services Step 2: benchmark the services
6/22/2010 HotCloud 2010, Boston 9
Intra-cloud network
Storage service
Computation service
Wide-area network
Web application
![Page 10: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/10.jpg)
6/22/2010 HotCloud 2010, Boston 10
Step 3: capture realistic application workload ◦ Extract the execution path of each request
Step 4: estimate the performance and costs ◦ Combine benchmarking results and workload information
Frontend
Database
Request
Response
Estimated processing time Estimated cost
![Page 11: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/11.jpg)
How to design the benchmarking tasks? ◦ Fair and representative
How to accurately capture the execution path of a request? ◦ An execution path can be complex, across
multiple machines
How to estimate the overall processing time of an application ◦ Applications can be multi-threaded
6/22/2010 HotCloud 2010, Boston 11
![Page 12: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/12.jpg)
How to design the benchmarking tasks? ◦ Fair and representative
How to accurately capture the execution path of a request? ◦ An execution path can be complex, across
multiple machines
How to estimate the overall processing time of an application ◦ Applications can be multi-threaded
6/22/2010 HotCloud 2010, Boston 12
![Page 13: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/13.jpg)
Java-based benchmarking tasks ◦ CPU/memory/disk I/O intensive ◦ Same byte-code on different providers Minimize the bias introduced by different compilers/
interpreters
Measure the cost per task ◦ Pay per instance-hour Compute using the per hour price and the task
running time ◦ Pay per CPU cycle Obtain the CPU cycles using cloud APIs
6/22/2010 HotCloud 2010, Boston 13
![Page 14: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/14.jpg)
Test common storage operations ◦ Insert/fetch/query ◦ Test against tables with different sizes
Measure each operation’s latency and cost
6/22/2010 HotCloud 2010, Boston 14
![Page 15: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/15.jpg)
Intra-cloud network ◦ Measure the TCP throughput and latency
between two randomly chosen instances
Wide-area network ◦ Measure the latency from vantage points on
PlanetLab Vantage points are chosen from diverse locations
6/22/2010 HotCloud 2010, Boston 15
![Page 16: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/16.jpg)
Measure three popular cloud providers ◦ One PaaS, two IaaS with storage APIs ◦ Names of the clouds are removed due to legal
concerns Referred to as Cloud X, Y, and Z
6/22/2010 HotCloud 2010, Boston 16
![Page 17: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/17.jpg)
6/22/2010 HotCloud 2010, Boston 17
At similar pricing points, different clouds can offer greatly diverse performance
Benchmark finishing time Benchmark running cost
![Page 18: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/18.jpg)
6/22/2010 HotCloud 2010, Boston 18
• Despite X’s good performance in computation, its storage service can be slower than the others • A cloud may not ace all services
![Page 19: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/19.jpg)
Minimum latency to the closest data center
6/22/2010 HotCloud 2010, Boston 19
• On average, X’s wide-area network latency can be up to 80% shorter than that of the others
![Page 20: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/20.jpg)
Deploy real application on different clouds ◦ BLAST: distributed, computation intensive
6/22/2010 HotCloud 2010, Boston 20
Cloud Y Cloud Z
Future work: to estimate the exact time and costs using the benchmarking results
X Y Z
![Page 21: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/21.jpg)
Choosing the best-suited cloud is non-trivial CloudCmp aims to help compare cloud providers
without actual deployment ◦ Application-specific ◦ Little deployment overhead ◦ Estimate both performance and costs
We think CloudCmp can be useful in practice ◦ Clouds offer diverse performance ◦ No cloud aces all services ◦ Benchmarking results correlate well with actual application
performance
6/22/2010 HotCloud 2010, Boston 21
![Page 22: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/22.jpg)
Questions? http://cloudcmp.net
6/22/2010 HotCloud 2010, Boston 22
![Page 23: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/23.jpg)
6/22/2010 HotCloud 2010, Boston 23
![Page 24: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/24.jpg)
6/22/2010 HotCloud 2010, Boston 24
• The scaling latencies of different providers vary significantly (Z’s latency is more than twice as high as Y’s) • The choice of operating system can affect scaling performance as well
![Page 25: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/25.jpg)
Blackbox tools to infer causal relationship ◦ Do not require modifying the application ◦ vPath [Tak09] Exploit the common programming model of web applications
◦ //Trace [Mesnier07] A more general approach using the throttling technique
6/22/2010 HotCloud 2010, Boston 25
![Page 26: Ang Li Xiaowei Yang Duke University Srikanth …Srikanth Kandula Ming Zhang Microsoft Research 6/22/2010 HotCloud 2010, Boston 1 Cloud computing is gaining popularity 6/22/2010 HotCloud](https://reader034.vdocuments.site/reader034/viewer/2022042409/5f25ad94401db63d9204ae51/html5/thumbnails/26.jpg)
Simulate the execution process ◦ Similar to the technique used in WebProphet [Li2010] ◦ Estimate the time spent on each component using
benchmarking results ◦ Simulate the execution with the constraints of the
causal relationships E.g., component A depends on component B, then A can only
be executed after B has finished
6/22/2010 HotCloud 2010, Boston 26