Resource Allocation using Adaptive Characterization of Online, Data-Intensive Workloads

218 Pages·2017·7.68 MB·English

Resource Allocation using Adaptive Characterization of
Online, Data-Intensive Workloads
Dissertation
Presented in Partial Fulfillment of the Requirements for the Degree Doctor
of Philosophy in the Graduate School of The Ohio State University
By
Jaimie Kelley, B.S., M.S.
Graduate Program in Computer Science and Engineering
The Ohio State University
2017
Dissertation Committee:
Dr. ChristopherStewart,Advisor
Dr. SrinivasanParthasarathy
Dr. P.Sadayappan
c Copyrightby
JaimieKelley
2017
Abstract
Cloud resource providers balance maximizing utilization under a power cap with meet-
ing workload Service Level Agreements (SLA). As the amount of data used by workloads
increases, so do the pressures on compute capacity in the cloud. Even if the resources
assigned meet an interactive workload’s need for low latency, the data that interactive
workload processes with allocated resources may not be sufficient to achieve a standard
of answer quality. Increasing the resources allocated to a specific workload to meet its an-
swer quality standard reduces the overall profit a cloud provider can make on interactive
workloads. However, if a workload’s answer quality standard is not met, the interactive
workload may seek another placement. Cloud instances can be purchased by the minute,
and multiple opportunities for placement exist. Because of this, cloud providers need to
puttheirclients’interests firstorloserevenue.
To best serve their own and their clients interest, cloud providers need data which reflects
resource usage, answer quality, and service level. If a cloud provider knows the amount
of power used by each workload scheduled, it can better fulfill its power cap requirements
withoutpenalty. Ifacloudproviderknowsthecurrentlatencyandanswerqualityofsched-
uled workloads, it can decide when to reallocate resources. However, this is difficult be-
cause any collection of data online imposes overheads. While cloud providers generally
ii
reservesomepercentage(5%)ofutilizationforoperatingsystemfunctions,datacollection
andanalysismustbedone carefullytoavoidundueimpactonscheduled workloads.
Iuseadaptivesolutionstotradeaccuracyforoverheadinworkloadcharacterization. Adap-
tiveworkloadcharacterizationsinformresourcemanagementwithoutthehighoverheadof
completecalculation, butarenotcompletelyaccurate.
In my work, I adaptively reduce the time spent profiling peak power to the degree of accu-
racythatacloudprovideriswillingtoaccept. Idevelopedamodelforadaptivelyprofiling
peakpowerusagetodeterminecorescaling. Adaptiveprofilingsavedupto93%collection
timewhilereducingaccuracyby3%onaverage.
To obtain answer quality for online resource management, I overlap execution of online
requests with the execution of requests that use all relevant data by using memoization of
complete responses from specific components. I built Ubora to obtain and allow manage-
ment of answer quality for interactive, data-intensive workloads. Cloud providers set the
rateatwhichqueriesare sampled,whichexchangesoverheadfor accuracy.
Finally, I designed Quikolo, a service that speculatively deploys and characterizes a target
workload in-situ in a colocation placement. Clients use this characterization to decide
whether to migrate their workload to this available placement. Quikolo also enables study
ofoverheadandaccuracyinfluencedbythenumberoffeaturesandcollectiontimeusedfor
workloadcharacterization.
iii
Adaptivelytradingaccuracyreducestheimpactofworkloadcharacterizationonoverhead.
My adaptive characterization solutions enable cloud providers to provision for lower over-
head and still achieve information that aids balancing client needs with available cloud
resources.
iv
Vita
2001-2005 ..................................Lutheran High School West - Rocky
River,OH
2005-2009 ..................................Bachelor of Science, Computer Science
andEnglish(Writing),
HeidelbergUniversity
2010-2015 ..................................Masters of Science, Computer Science
andEngineering,
TheOhioStateUniversity
2010-present ................................PhD student in Computer Science and
Engineering,
TheOhioStateUniversity
Publications
Jaimie Kelley, Christopher Stewart, Nathaniel Morris, Devesh Tiwari, Yuxiong He, and
Sameh Elnikety, ”Obtaining and Managing Answer Quality for Online Data-Intensive
Services”. Journal ACM Transactions on Modeling and Performance Evaluation of Com-
putingSystems,2017.
JaimieKelley,ChristopherStewart,DeveshTiwari,andSaurabhGupta, ”AdaptivePower
Profiling for Many-Core HPC Architectures”. International Conference on Autonomic
Computing,2016.
v
Jaimie Kelley, Christopher Stewart, Devesh Tiwari, Sameh Elnikety, and Yuxiong He,
”Measuring and Managing Answer Quality for Online Data-Intensive Services”. Interna-
tionalConferenceonAutonomicComputing,2015.
Sundeep Kambhampati, Jaimie Kelley, William C. L. Stewart, Christopher Stewart, and
RajivRamnath, ”ManagingTinyTasksforData-Parallel,SubsamplingWorkloads”. IEEE
InternationalConference onCloudEngineering,2014.
Jaimie Kelley, Christopher Stewart, Sameh Elnikety, and Yuxiong He, ”Cache Provision-
ingforInteractiveNLPServices”. LargeandDistributedSystemsandMiddleware,2013.
Jaimie Kelley and Christopher Stewart, ”Balanced and Predictable Networked Storage”.
InternationalWorkshop onDataCenterPerformance,2013.
Nan Deng, Christopher Stewart, Jaimie Kelley, Daniel Gmach and Martin Arlitt, ”Adap-
tiveGreenHosting”. International ConferenceonAutonomicComputing, 2012.
Fields of Study
MajorField: Computer ScienceandEngineering
vi
Table of Contents
Page
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Vita . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
ListofTables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
ListofFigures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. ManagingTinyTasks forData-Parallel,SubsamplingWorkloads . . . . . . . . 9
2.1 SubsamplingWorkloads . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 TheCaseforTinyTasks . . . . . . . . . . . . . . . . . . . . . . 14
2.2 ManagingTinyTasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Job-vsTask-level Recovery . . . . . . . . . . . . . . . . . . . . 18
2.2.2 PlatformSelection . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 TaskSizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 ExperimentalSetup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 ExperimentalResults . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3. AdaptivePowerProfiling forMany-CoreHPCArchitectures . . . . . . . . . . 39
3.1 ExperimentalMethodology . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.1 Powermeasurement . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.2 Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.3 Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.4 Workloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 ObservationsonPowerConsumption . . . . . . . . . . . . . . . . . . . 46
vii
3.3 PredictingPeakPowerusingReferenceWorkloads . . . . . . . . . . . . 51
3.4 AnalyzingthePower ConsumptionProfileofScientificApplications . . . 53
3.5 AdaptivePower Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5.1 OurProfilingMethod . . . . . . . . . . . . . . . . . . . . . . . 60
3.5.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.5.3 CornerCases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.6 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4. BalancedandPredictable NetworkedStorage . . . . . . . . . . . . . . . . . . 67
4.1 Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.1 OutliersinNetworkedStorage . . . . . . . . . . . . . . . . . . . 71
4.1.2 WorkloadsthatReduce BigData . . . . . . . . . . . . . . . . . 72
4.2 ProblemStatement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3 ModellingOutliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 ReplicationforPredictability . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5. CacheProvisioning forInteractiveNLPServices . . . . . . . . . . . . . . . . 86
5.1 NLPWorkloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1.1 DefiningQuality Loss . . . . . . . . . . . . . . . . . . . . . . . 90
5.2 ExperimentalResults . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.1 ComparingNLP Datasets . . . . . . . . . . . . . . . . . . . . . 93
5.2.2 CacheReplacement Policies . . . . . . . . . . . . . . . . . . . . 94
5.2.3 WholeDistribution Analysis . . . . . . . . . . . . . . . . . . . . 95
5.2.4 CacheProvisioning onQualityLoss . . . . . . . . . . . . . . . . 96
5.2.5 Additionalissues . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6. ObtainingandManaging AnswerQuality
forOnlineData-IntensiveServices . . . . . . . . . . . . . . . . . . . . . . . . 102
6.1 BackgroundonOLDI Services . . . . . . . . . . . . . . . . . . . . . . . 106
6.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.3.1 DesignGoals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.3.2 Timeliness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.3.3 Transparency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.3.4 LowOverhead . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3.5 LowCost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.3.6 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
viii
6.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.4.1 InterfaceandUsers . . . . . . . . . . . . . . . . . . . . . . . . 122
6.4.2 TransparentContextTracking . . . . . . . . . . . . . . . . . . . 123
6.4.3 Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.4.4 Optimizationsfor LowOverhead . . . . . . . . . . . . . . . . . 131
6.4.5 DeterminingFront-End Components . . . . . . . . . . . . . . . 132
6.5 ExperimentalEvaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.5.1 MetricsofMerit . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.5.2 CompetingDesigns andImplementations . . . . . . . . . . . . . 135
6.5.3 OLDIServices . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.6 OnlineManagement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.7 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.7.1 Approximationfor Performance . . . . . . . . . . . . . . . . . . 150
6.7.2 QueryTagging . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.7.3 TimeoutToggling: AdaptiveConfiguration . . . . . . . . . . . . 153
6.7.4 AdaptiveResource Allocation . . . . . . . . . . . . . . . . . . . 155
7. RapidIn-situCharacterization forCo-LocatedWorkloads . . . . . . . . . . . . 156
7.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.1.1 DesignGoals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.1.2 DesignParameters . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.1.3 DesignLimitations . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.2 QuikoloImplementation . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.2.1 FeatureCollection . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.2.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.3 ExperimentalEvaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3.2 Workloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3.3 Overhead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.4 DurationStudy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
7.4.1 StatisticalConvergence . . . . . . . . . . . . . . . . . . . . . . 171
7.5 FeaturesStudy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.5.1 WhichFeatures Matter . . . . . . . . . . . . . . . . . . . . . . . 177
7.6 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
8. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
ix