Computer Science Dynamic Resource Management in Internet Data Centers Prashant Shenoy University of Massachusetts

Computer Science Motivation  Internet applications used in a variety of domains  Online banking, online brokerage, online music store, e-commerce  Internet usage continues to grow rapidly  Broadband deployment is accelerating  Outages of Internet applications more common “Site not responding” “connection timed out”

Computer Science Internet Application Outages Down for 30 minutes Average download time ~ 260 sec Periodic outages over 4 days Cause: Too many users leading to overload Holiday Shopping Season 2000: 9/11: site inaccessible for brief periods

Computer Science Internet Workloads are highly variable  Short-term fluctuations  “Slashdot Effect”  Flash Crowds  Long-term seasonal effects  Time-of-day, month-of-year  Peak difficult to predict  Static overprovisoning not effective  Manual allocation: slow Soccer World Cup’98 Key Issue: How can we design applications to handle large workload variations?

Computer Science Internet Data Centers Internet applications run on data centers Server farms Provide computational and storage resources Applications share data center resources Problem: How should the platform allocate resources to absorb workload variations?

Computer Science Talk Outline Motivation  Internet data center model  Dynamic provisioning  Request Policing  Cataclysm Server Platform  Experimental results  Summary

Computer Science Data Center Model  Dedicated hosting: each application runs on a subset of servers in the data center  Subsets are mutually exclusive: no server sharing  Data center hosts multiple applications  Free server pool: unused servers Retail Web site streaming

Computer Science Internet Application Model  Internet applications: multiple tiers  Example: 3 tiers: HTTP, J2EE app server, database  Replicable applications  Individual tiers: partially or fully replicable  Example: clustered HTTP, J2EE server, shared-nothing db  Each application employs a sentry  Each tier uses a dispatcher: load balancing requests http J2EE database Load balancing sentry

Computer Science Approach  Dynamic provisioning  Allocate servers to applications on-the-fly  Request policing  Turn away excess requests  Degrade performance based on SLA  Couple provisioning and policing

Computer Science Research Questions  How many servers to allocate and when?  Multi-tier apps: when and how to provision each tier?  How many requests should be turned away during overload?  Multi-tier apps: where should requests be dropped?  Can we meet SLAs during overloads?  Is it possible to predict future workloads?

Computer Science Dynamic Provisioning  Key idea: increase or decrease allocated servers to handle workload fluctuations  Monitor incoming workload  Compute current or future demand  Match number of allocated servers to demand Monitor workload Monitor workload Compute current/ future demand Compute current/ future demand Adjust allocation

Computer Science Single-tier Provisioning  Single tier provisioning well studied [Muse, TACT]  Non-trivial to extend to multiple-tiers  Strawman #1: use single-tier provisioning independently at each tier  Problem: independent tier provisioning may not increase goodput C=15 C=10 C= req/s dropped 4 req/s

Computer Science Single-tier Provisioning  Single tier provisioning well studied [Muse, TACT]  Non-trivial to extend to multiple-tiers  Strawman #1: use single-tier provisioning independently at each tier  Problem: independent tier provisioning may not increase goodput C=15 C= req/s 14 C=20 14 dropped 3.9 req/s 10.1

Computer Science Model-based Provisioning  Black box approach  Treat application as a black box  Measure response time from outside  Increase allocation if response time > SLA Use a model to determine how much to allocate  Strawman #2: use black box for multi-tier apps  Problems:  Unclear which tier needs more capacity  May not increase goodput if bottleneck tier is not replicable 14 req/s C=15 C= C=

Computer Science Provisioning Multi-tier Apps  Approach: holistic view of multi-tier application  Determine tier-specific capacity independently  Allocate capacity by looking at all tiers (and other apps)  Predictive provisioning  Long-term provisioning: time scale of hours  Maintain long-term workload statistics  Predict and provisioning for the next few hours  Reactive provisioning  Short term provisioning: time scale of several minutes  React to “current” workload trends  Correct errors of long-term provisioning  Handle flash crowds (inherently unpredictable)

Computer Science Workload Prediction  Long term workload monitoring and prediction  Monitor workload for multiple days  Maintain a histogram for each hour of the day Capture time of day effects  Forecast based on Observed workload for that hour in the past Observed workload for the past few hours of the current day  Predict a high percentile of expected workload Mon Tue Wed Today

Computer Science Predictive Provisioning  Queuing theoretic application model  Each individual server is a G/G/1 queue  Derive per-tier E(r) from end-to-end SLA  Monitor other parameters and determine  per-server capacity)  Use predicted workload pred to determine # servers per tier Assumes perfect load balancing in each tier  Alternative: each tier G/G/k G/G/1 pred

Computer Science Reactive Provisioning  Idea: react to current conditions  Useful for capturing significant short-term fluctuations  Can correct errors in predictions  Track error between long-term predictions and actual  Allocate additional servers if error exceeds a threshold  Account for prediction errors  Can be invoked if request drop rate exceeds a threshold  Handles sudden flash crowds  Operates over time scale of a few minutes  Pure reactive provisioning: lags workload  Reactive + predictive more effective! Prediction error pred actual error >  Invoke reactor time series allocate servers

Computer Science Talk Outline Motivation Internet data center model Dynamic provisioning  Request Policing  Cataclysm Server Platform  Experimental results  Summary

Computer Science Request Policing  Key Idea: If incoming req. rate > current capacity  Turn away excess requests  Degrade performance of requests  Why police when you can provision?  Provisioning is not instantaneous Residual sessions on reallocated server Application and OS installation and configuration overheads  Overhead of several (5-30) minutes Sentry policing G/G/1 drop

Computer Science Class-based Differentiation  Some requests are more important than others  Purchase versus catalog browsing  Stock trade versus view account balance  Overload => preferentially let in more important requests  Maximize utility during overload  Incoming requests queued up in class queues  Example: gold, silver, bronze class  Higher priority to more important classes Sentry policing drop

Computer Science Scalable Policing Techniques  Examining individual requests infeasible  Incoming rate may be order of magnitude greater than capacity  Need to reduce overhead of policing decisions  Idea #1: Batch processing  Premise: Requests arrivals are bursty  Admit a batch of queued up requests One admission control test per batch Reduces overhead from O(n) to O(b)  Idea #2: Use pre-computed thresholds  Example: capacity = 100 req/s, G=75, S=50, B=50 req/s Admit all gold, half of silver and no broze  Periodically estimate and s: compute threshold  O(1) overhead: trades accuracy for efficiency

Computer Science Cataclysm Server Platform  Prototype data center  Commodity hardware  40+ Pentium servers  2 TB of RAID arrays  Gigabit switches  Linux-based platform

Computer Science Cataclysm Software Architecture Cataclysm Control Plane Provisioning Global allocation App placement Nucleus Apps OS Nucleus Apps OS Nucleus Apps OS Server Node Runs apps, sentries Resource monitoring, Local allocation  Two key components: control plane and nuclei

Computer Science Cataclysm Node Architecture  Capsule: component of an app on a node  Qlinux: proportional-sharing of node resources  Nucleus: resource allocations across capsules and VMs Nucleus Capsule QLinux HSFQ CPU scheduler Prop-share packet sched Cello disk scheduler SFVM memory mgr Nucleus QLinux Capsule VM Capsule VM Capsule VM Active Dormant UML Xen

Computer Science Cataclysm Applications  Multi-tiered apps: Rubis (e-auctions), Rubbos (b-board)  Apache, JBOSS, mysql  Tier-1 Sentry  Ktcpvs: kernel HTTP load balancer  Request policing and class-based differentiation  Workload monitoring  Tier-2 sentry: Apache JBOSS redirector, workload monitoring  Nuclues: Linux trace toolkit, /proc to monitor node statistics  All system components are replicable! Apache Load bal police ktcpvs Apache JBOSS mysql

Computer Science Talk Outline Motivation Internet data center model Dynamic provisioning Request Policing Cataclysm Server Platform  Experimental results  Summary

Computer Science Dynamic Provisioning Server Allocation adapts to changing workload WorkloadServer Allocation  RuBiS: E-auction application like Ebay

Computer Science Class-based differentiation Arrival rate Time (sec) Arrival rate GLD SIL BRZ Fraction admitted Time (sec) Fraction admitted GLD SIL BRZ

Computer Science Threshold-based: higher scalability Scalability Arrival rate CPU usage Batch Thresh

Computer Science Other Research Results  OS Resource Allocation  Qlinux [ACM MM00], SFS [OSDI00], DFS [RTAS02]  SHARC cluster-based prop. sharing [TPDS03]  Shared hosting provisioning  Measurement-based [IWQOS02], Queuing-based [Sigmetrics03,IWQOS03]  Provisioning granularity [Self-manage 03]  Application placement [PDCS 2004]  Profiling and Overbooking [OSDI02]  Storage issues  iSCSI vs NFS [FAST03], Policy-managed [TR03]

Computer Science Glimpse of Other Projects  Hyperion: Network processor based measurement platform  Measurement in the backbone and at the edge  NP-based measurements in the data center  RiSE: Rich Sensor Environments  Video sensor networks  Robotics sensor networks  Real-time sensor networks  Weather sensors

Computer Science Concluding Remarks  Internet applications see varying workloads  Handle workload dynamics by  Dynamic capacity provisioning  Request Policing  Need to account for multi-tiered applications  Joint work: Bhuvan Urgaonkar, Abhishek Chandra and Vijay Sundaram  More at

Computer Science Predictive Provisioning  Invoked once every hour  Captures long-term variations - time of day effects  Extensions to seasonal effects (month-of-year, holidays)  How to initialize?  Needs several days of history to work well  What happens if no servers are available?  Use revenue/utility to arbitrate allocation [Muse]  Turn away excess requests  Non-replicable tiers are easy to handle  Provision other tiers until non-replicable tier is saturated

Computer Science Degrade or Drop?  Depends on the application and the SLA  Degrading increases effective capacity  Also degrades performance seen by requests  Degrade if  Utility from servicing more requests at lower performance >  Utility from servicing fewer requests - penalty of dropping requests  Otherwise drop requests < 500msr1 < 1sr2 <10sr3 SLA:

Computer Science Use of Virtual Machine Monitors  Server allocation can be slow (~ minutes)  Need residual sessions to terminate  Disk scrubbing, OS and app installation, configuration  Application and system overheads  Flash crowds => need fast allocation  Use virtual machines  Each app runs inside a VM, multiple VMs on a server  Only one VM is active at any time, other VMs are “hot spares”  Server allocation => idle one VM, activate another  System overhead reduces to < 1s  Need to still account for residual sessions  Application issue, not longer a system limitation

Computer Science Threshold-based: loss of accuracy