Initial Results on Provisioning Variation in Cloud Services M. Suhail Rehman Research Analyst Cloud Computing Lab Carnegie Mellon University in Qatar Collaborators: Prof. Majd F. Sakr, Jim Gargani Carnegie Mellon University Supported By: 1
Cloud Computing / IaaS What about other Application Domains? Scientific Applications High Performance Computing 2
Cloud Computing / IaaS What about other Application Domains? Scientific Applications High Performance Computing Application Performance on the Cloud 3
What could affect performance? Virtualized and Multiplexed Hardware Virtualization Multitenancy Simplified, Abstracted Hardware Abstraction Identical Requests are not guaranteed to give you Identical Hardware 4
Related Work Many Studies on Virtualization and Application Performance Application Performance on EC2 • Detailed studies on performance variance: • Service Oriented Applications : upto 4x [Dejun 2009] • ~ 10- 25% Variation observed for benchmarks on EC2 [Schad 2010] 5
A closer look… L1 L2 4-core CPU L3 L1 RAM Disk L2 L1 L2 4-core CPU L3 4-core CPU Physical Host RAM Disk L1 L1 L2 L2 4-core CPU L3 RAM Disk 4-core CPU L1 L2 Physical Host 4-core CPU Physical Host 6
A closer look… VM L1 L2 4-core CPU L3 VM VM L1 RAM Disk L2 VM VM L1 L2 4-core CPU L3 4-core CPU Physical Host RAM Disk L1 VM L1 L2 L2 VM 4-core CPU L3 RAM Disk 4-core CPU L1 VM L2 Physical Host 4-core CPU Physical Host 7
A closer look… VM L1 L2 4-core CPU L3 VM VM L1 RAM Disk L2 VM VM L1 L2 4-core CPU L3 4-core CPU Physical Host RAM In Cloud Computing Disk These Details are Abstracted from the User L1 VM L1 L2 L2 VM 4-core CPU L3 RAM Disk 4-core CPU L1 VM L2 Physical Host 4-core CPU Physical Host 8
3 Potential Reasons for Performance Issues on the Cloud Loads from other VMs on the same 1 machine 9
3 Potential Reasons for Performance Issues on the Cloud Loads from other VMs on the same 1 machine Variation in the physical resources 2 being assigned to identical instances 10
3 Potential Reasons for Performance Issues on the Cloud Loads from other VMs on the same 1 machine Variation in the physical resources 2 being assigned to identical instances Configuration of the VM layout (where 3 the VMs are placed during provisioning) 11
Why does layout matter? I want 4 VMs each with 1 vCPU, 1 GB RAM and 80 GB Disk Resource Request Client Physical Hardware Cloud Provider Virtual Machine 12
Provisioning Variation “variation due to ambiguity in the mapping of virtual resources to physical resources in a cloud computing environment” VM VM Application VM VM VMs from the Cloud Application Performance Variation 13 13
Experimental Methodology Controlled Experimentation on a private cloud • Create Identical VM cluster instances in different physical layouts manually • Evaluate the effect on performance for various applications. V V V V M M M M Physical Physical Physical Physical V V Host Host Host Host M M Provisioned Layout 1: 4 VMs across 4 blades on a private V V M M cloud V V V V V V M M M M M M Client request for 4 VMs V V M M Physical Physical Physical Host Host Host Layout 3: Layout 2: 4 VMs across 4 VMs across 2 blades 1 blade 14
Testbed Configuration Hadoop 0.20.1 VM RHEL 5.2 IBM Bladecenter H with14 Blades Xen 3.0.3 RHEL 5.1 Blade CPU: 2 x Quad Xeon E5420 2.5 GHz w/ 12MB L2 Cache L1 L2 RAM: Front-Side Bus: 4-core CPU L3 RAM 8 GB ECC 21.6 GB/sec Disk Disk: Disk Bandwidth: L1 L2 2 x 300 GB SAS 600 MB/sec 4-core CPU Network Interface Physical Host 2x Gigabit Interfaces to other blades 15
Benchmark Tests and Applications Executed on Synthetically Configured Systems Infrastructure Benchmarks V V V V M M M M • CPU: SysTester Physical Physical Physical Physical • Memory: STREAM Host Host Host Host 4 VMs across 4 blades • Disk: Bonnie++ V V V V V V • Network: Netperf M M M M M M V V M M Physical Physical Physical Host Host Host Hadoop 4 VMs across 4 VMs across 2 blades Workloads 1 blade • Hadoop Sort • Hadoop Wordcount • Hadoop TestDFSIO 16
Results of Systems Benchmarks No Variation V V V V M M M M CPU Physical Physical Physical Physical Host Host Host Host 25% drop in Layout 1 RAM bandwidth for V V V V V V Layouts 2 and 3 M M M M M M V V M M Physical Physical Physical Host Host Host 60 – 80 % drop in Layout 3 Disk Layout 2 bandwidth for Layouts 2 and 3 ~ 4x speedup for Network Layout 3 17
Hadoop Sort 10000 Time in Seconds (Log Scale) V V V V 1000 M M M M Physical Physical Physical Physical Host Host Host Host 100 Layout 1 V V V V V V M M M M M M V V 10 M M Physical Physical Physical Host Host Host Layout 3 Layout 2 1 256 512 1024 2048 4096 Layout 3 Layout 3 38.5 99.6 287.1 1339.7 6644 5x performance variation Layout 2 Layout 2 30.3 44.3 113.7 527.6 2400.7 Layout 1 Layout 1 29.8 38.3 66.4 362.9 1311.3 Size (MB) 18
DFSIO Benchmark 35.000 Read 30.000 Throughput (mb/sec) Write V V V V M M M M 25.000 Physical Physical Physical Physical Host Host Host Host 20.000 Layout 1 15.000 V V V V V V M M M M M M 10.000 V V M M Physical Physical Physical 5.000 Host Host Host Layout 3 Layout 2 0.000 Layout 1 Layout 2 Layout 3 1x4 2x2 4x1 Layout (VMxHosts) ~ 5x performance variation 19
Analysis Correlation between Sort and DFSIO Benchmark • Upto 5x performance drop in both • Disk contention is the reason Hadoop designed to leverage parallel I/O • When all VMs are on one blade, they compete for disk I/O bandwidth 20
Hadoop Wordcount 1800 1600 1400 V V V V M M M M Runtime (Seconds) 1200 Physical Physical Physical Physical Host Host Host Host 1000 Layout 1 800 V V V V V V M M M M M M 600 V V M M 400 Physical Physical Physical Host Host Host 200 Layout 3 Layout 2 0 1 2 4 8 Layout 3 4Vx1B 182 339.9 660.2 1633.8 ~20% Layout 2 4Vx2B 167.2 321.9 620.4 1414.4 Layout 1 4Vx4B 167.1 317.2 631.5 1356.8 Input Size (GB) 21
Conclusions Tradeoffs • VM placement on same resource • Higher bandwidth for inter-VM communication • Constraints Memory and Disk Provisioning Variation • It’s impact on performance varies across application domains • Up to 5x performance variation for I/O-bound 22
Future Studies and Directions We have only scratched the surface! More Studies on other Applications • Different classes of scientific applications (CPU, Memory, I/O Bound) Application Profiling on the Cloud • To inform provisioning to meet QoS Resource-aware Applications • Dynamic application adaptation to variations in cloud resources 23
Join Us! Postdoctoral Positions Carnegie Mellon Qatar http://qatar.cmu.edu/~msakr/postdoc Contact Prof. Majd Sakr or Email Me: msakr@qatar.cmu.edu suhailr@qatar.cmu.edu 24
Recommend
More recommend
