Considering Resource Demand Misalignments To Reduce Resource Over-Provisioning in Cloud Datacenters Liuhua hua Chen en Haiyi ying ng Shen en Dept pt. of E f Elec ectric trical al and d Co Compu puter ter Eng. g. Dept pt. of Co f Compu pute ter r Sc Science ence Clems Cl mson on University, ersity, USA SA University ersity of f Vi Virgi ginia, nia, USA SA
Cloud Computing • Cloud computing: large groups of remote servers networked to allow centralized data storage and online access to computer services or resources 2 Prof. Haiying Shen, University of Virginia
Cloud Providers 3 Prof. Haiying Shen, University of Virginia
Cloud Customers 4 Prof. Haiying Shen, University of Virginia
Research Problem and Goal Virtual machines (VMs) Physical machines (PMs) 5 Prof. Haiying Shen, University of Virginia
Motivation VM VM VM VM VM VM VM Virtual machine (VM) VM VM VM VM Physical machine (PM) PM VM VM PM VM PM • Over-loaded PMs low QoS SLO violation penalty • Under-loaded PMs resource waste high system cost • Problem: reduce over-loaded and under-loaded PMs • Goal: high QoS, high resource efficiency, high profit 6 Prof. Haiying Shen, University of Virginia
Initial Complementary VM Consolidation • Previous work (CompVM): L. Chen and H. Shen, Consolidating Complementary VMs with Spatial/Temporal- awareness in Cloud Datacenters, Proc. of the 33rd Annual IEEE International Conference on Computer Communications (INFOCOM'14), Toronto, Canada, 2014 VM VM VM VM VM VM VM VM PM PM PM PM PM • How to achieve ? 7 Prof. Haiying Shen, University of Virginia
Initial Complementary VM Consolidation – Motivation Capacity Resource waste • CompVM: load balance in the long term VM 1 consolidate complementary VMs Demand Under-loaded PMs Capacity Spatial Temporal 100 100 utilization (%) VM 1 VM 2 Utilization (%) VM 3 Memory Demand High CPU low MEM VM 1 Over-loaded PMs VM 1 Low CPU VM 2 high MEM Capacity 0 Time CPU utilization (%) 100 0 T VM 2 VM 1 Demand No under/over-loaded PMs Patterns? 8 Prof. Haiying Shen, University of Virginia
Initial Complementary VM Consolidation – VM Utilization Pattern • Measurement: – MapReduce jobs: TeraSort, TestDFSIO read/write – cluster trace, trace • Periodic resource utilization patterns exist in many VMs running – the same short-term job – a long-term job TestDFSIO read Google cluster trace TestDFSIO read 9 Prof. Haiying Shen, University of Virginia
Initial Complementary VM Consolidation – Utilization Pattern Detection 5.0 trace 1 u CPU utilization (%) 4.0 3.0 2.0 f 1 trace 2 1.0 f 2 f 3 0.0 0 12 24 36 48 60 72 Time (hr) pattern 10 Prof. Haiying Shen, University of Virginia
Initial Complementary VM Consolidation – VM Allocation Method pattern VM7 Detect resource Detect resource Choose a PM utilization utilization pattern of the VM pattern of the VM consolidate complementary VMs VM6 Spatial Temporal VM3 100 VM5 100 utilization (%) Utilization (%) For each PM: For each PM: VM 2 VM 3 VM1 PM VM2 PM Memory VM4 PM High CPU check if it has check if it has PM1 PM2 low MEM PM3 VM 1 sufficient capacity sufficient capacity VM 1 Low CPU VM7 for the VM for the VM VM 2 high MEM Choose a PM 0 CPU utilization (%) Time 100 0 T VM7 Choose a PM Select the PM: Select the PM: Memory VM6 with the least with the least resource waste remaining resource VM3 remaining resource VM5 VM6 after allocating the after allocating the VM1 PM PM VM3 VM2 VM4 PM VM5 VM 𝐹 VM 𝑘 : ratio of aggregated resource demand VM1 PM VM2 PM PM1 PM2 VM4 PM PM3 𝑁 𝑘 : distance between the average resource PM1 PM2 PM3 demand vector and the capacity vector 11 Prof. Haiying Shen, University of Virginia
Reducing Prediction Error pattern 12 Prof. Haiying Shen, University of Virginia
VM Consolidation – Reduce Provisioned Resource • Pulse deviation yields a pattern with a pulse width larger than the actual pulse width • Resource over-provisioning • Not revealed or studied before deviation derived derived pattern Pulse pattern 13 Prof. Haiying Shen, University of Virginia
VM Consolidation – Trace Study • Pulse deviations are common PlanetLab trace Google Cluster trace 100 jobs, 29920 tasks 1000 jobs, 4695 tasks MapReduce benchmarks on a HPC cluster (Wordcount, Grep, Terasort, TestDFSIO and PiEstimator) Average task execution time is around 100 minutes/seconds 14 Prof. Haiying Shen, University of Virginia
VM Consolidation – Trace Study • Resource efficiency: demand/capacity • Even using CompVM, the resource efficiency still needs to improve PlanetLab trace Google Cluster trace 1000 jobs, 4695 tasks 100 jobs, 1550 tasks 15 Prof. Haiying Shen, University of Virginia
VM Consolidation – Pattern Refinement Methodology • Pattern refinement methods – Lowering cap : lower each value in the pattern by c high -c low – Reducing pulse width – Optimal base provisioning c high c low unused resource t 1 t 2 t 3 0 Time 16 Prof. Haiying Shen, University of Virginia
VM Consolidation – Pattern Refinement Methodology • Pattern refinement methods – Lowering cap – Reducing pulse width : postpone the pulse from t 1 to t 3 – Optimal base provisioning 17 Prof. Haiying Shen, University of Virginia
VM Consolidation – Pattern Refinement Methodology • Pattern refinement methods – Lowering cap – Reducing pulse width – Optimal base provisioning refine pattern based on optimal b value that maximizes resource efficiency b 1 b 2 0 Time 18 Prof. Haiying Shen, University of Virginia
VM Consolidation – Pattern Refinement Methodology • Pattern refinement methods – Lowering cap – Reducing pulse width – Optimal base provisioning • Risk violating SLOs 19 Prof. Haiying Shen, University of Virginia
VM Consolidation – Performance Evaluation Google Cluster trace Cluster experiment Improve by 10%, 70% Improve by 40% 2000 VMs, CloudSim, Palmetto HPC cluster, Traces: NAS Parallel Benchmark , cluster Pattern refinement yields higher resource efficiency without compromising VM performance by handling pulse deviations! 20 Prof. Haiying Shen, University of Virginia
Conclusion • Trace study – Pulse deviations are common – Even using CompVM, the resource efficiency still needs to improve • Pattern refinement methods – Lowering cap – Reducing pulse width – Optimal base provisioning • Experiments – Higher resource efficiency without compromising VM performance 21 Prof. Haiying Shen, University of Virginia
Future Work • Consider other factors (e.g., SLOs) in VM consolidation • Consider VM migration 22 Prof. Haiying Shen, University of Virginia
Thank you! Questions & Comments? Haiying Shen hs6ms@virginia.edu Associate Professor Department of Computer Science University of Virginia 23 Prof. Haiying Shen, University of Virginia
Recommend
More recommend
