Migration, Assignment, and Scheduling of Jobs in Virtualized - - PowerPoint PPT Presentation
Migration, Assignment, and Scheduling of Jobs in Virtualized Environment Seung-Hwan Lim Jae-Seok Huh Youngjae Kim Chita R. Das The Pennsylvania State University Oak Ridge National Laboratory Overview Challenges Migration Cost
The Pennsylvania State University
Oak Ridge National Laboratory
Jae-Seok Huh Youngjae Kim Chita R. Das
Decouples Operating System Instances from Hardware Enables migration of OS instances (Virtual Machines)
HotCloud 2011 2/13
Overview Challenges Migration Cost Performance Model Conclusions
Obstacle Opportunity
1 Availability/Business Continuity Use Multiple Cloud Providers 2 Data Lock-In Standardize APIs, Compatible SW to enable Surge or Hybrid Cloud Computing 3 Data Confidentiality and Auditability Deploy Encryption, VLANS, Firewalls 4 Data Transfer Bottlenecks FedExing Disks; Higher BW Switches 5 Performance Unpredictability Improved VM support; Flash memory; Gang Schedule VMs 6 Scalable Storage Invent Scalable Store 7 Bugs in Large Distributed Systems Invent Debugger that relies on Distributed VMs 8 Scaling Quickly Invent Auto-Scaler that relies on ML; Snapshots for Conservation 9 Reputation Fate Sharing Offer reputation-guarding services like those for email 10 Software Licensing Pay-for-use licenses
From Armbrust et al., A view of cloud computing. Commun. ACM, April 2010
HotCloud 2011 3/13
Overview Challenges Migration Cost Performance Model Conclusions
5 Performance Unpredictability Improved VM support; Flash memory; Gang Schedule VMs
Propagation Effects
HotCloud 2011 4/13
J1.3 J1.2 J1.1 J2 Performance variance among jobs may create cascaded effects in all the related jobs J1.3 J1.2 J1.1 J2
Repels performance critical applications.
Overview Challenges Migration Cost Performance Model Conclusions
completion time Desired With performance variance
VM assignment/scheduling schemes consider performance
5/13
An optimization problem to find assignment of jobs to the given set
Jobs J Machines M m1 m2 m3 Total completion time
HotCloud 2011
Overview Challenges Migration Cost Performance Model Conclusions
A set of VMs migrates
HotCloud 2011 6/13
Physical Machines PM V4 V2 V5 V3 V1 PM1 PM2 PM3 Physical Machines PM V4 V2 V5 V3 V1 PM1 PM2 PM3 after ? after ? with ? Overview Challenges Migration Cost Performance Model Conclusions
Migration policy determines the amount of time to reassign VMs and hence impacts performance
completion time
HotCloud 2011 7/13
VM1* migrates Types of VM migrations
Sender Machine Receiver Machine VM2 (CPU0) VM3 (CPU0) Migration in Sender Migration in Receiver Domain-0 (CPU0, 1) Domain-0 (CPU0, 1) VM1* (CPU1) VM Sender Receiver migration time = down time VM (a) A non-live migration migration time down time ~ 0 VM Receiver VM Sender (b) A live migration
Overview Challenges Migration Cost Performance Model Conclusions
5 10 15 20 25 30 35 256 512 768 1024 1280 1536 1792 2048 2304 2560 2816
Total Migration Time (sec)
Total size of migrated memory (MB)
multiple 256MB VMs (1G Ethernet) single VM (1G Ethernet) multiple 256MB VMs (Infiniband) single VM (infiniband)
Migration policy decides total migration time
Total Migration Time T When a set of VM migrates, how do we minimize T while bounding β?
HotCloud 2011 8/13
Overview Challenges Migration Cost Performance Model Conclusions Slightly faster than ten sequential migrations (30.9sec < 10x3.1 sec, 22.8 sec< 10x2.6sec) , but with greater performance impact
Migration impacts performance
Performance Impact β Workload : Compressing 256KB files.
5 10 15 20 25 Pre-Migration Migration Post-Migration
# of compressed files /sec
VM1 (migrated VM) VM2 (non-migrated in Sender) VM3 (non-migrated in receiver)
*A generic model for n job and m resources has been developed.
HotCloud 2011 9/13
Overview Challenges Migration Cost Performance Model Conclusions
Total completion time may not be linear to individual completion times
HotCloud 2011 10/13
j
p
Bin packing or scheduling algorithms use linear relation Actual completion time T CPU I/O
50 100 150 200 250 CPU w/CPU CPU w/IO IO w/ CPU IO w/ IO
Measurement D-factor Linear Both CPU job and I/O job take 100 sec without presence of other workloads Overview Challenges Migration Cost Performance Model Conclusions Waste System Resources Total Completion Time (sec)
2 1 2 1 2 1 2 2 2 1 2 1 2 1 1 1
)) 1 )( 1 ( 1 ( )) 1 )( 1 ( 1 ( p p p p T p p p p T
Assume two 2-resource-busy jobs with their loading vectors, access probability of each resource, pi = (pi, 1-pi)*. Then, expanded completion time T of each job is given by
HotCloud 2011 11/13
Without
CPU Disk I/O Job 2 Job 1
OR
CPU Disk I/O Job 2 Job 1
Linear estimation Overview Challenges Migration Cost Performance Model Conclusions From original completion time *Without resource monitoring, loading vectors can be constructed (Algorithm 1)
Consider slow-down of jobs in the system as the performance variation
HotCloud 2011 12/13
Given τj, define performance impact β by
j j j
T
Slow down of job j Execution time of job j without other jobs Execution time of job j with other jobs Slow downs of all jobs
We can calculate this Overview Challenges Migration Cost Performance Model Conclusions
HotCloud 2011 13/13
Performance Model (Shared systems with multiple resources) Migration Cost Analysis (Profile migration as a job) Assignment Cost Analysis (Performance bound β and completion time of assignment T) Migration-Aware Schedulers
Overview Challenges Migration Cost Performance Model Conclusions