Performance Evaluation of Sched uling Algorithms for Database S ervices with Soft and Hard SLAs
Hyun Jin Moon, Yun Chi, Hakan Hacıgümü¸s NEC Laboratories America
Cupertino, USA DataCloud 2011
Best Paper Award Finalist
Introduction
Service Level Agreement(SLA) is a part o f a service contract where the level of service is formally defined.
◦In this paper, service latency, or response time.
Two types of SLA
◦Soft SLA:
describe SLA profit as a function of response time .
◦Hard SLA:
species a single firm deadline objective for each job.
SLA
Soft SLA
Hard SLA
This Paper
Rigorously evaluate a comprehensiv e set of scheduling methods and pr esent how they perform with respec t to the full requirement list.
Propose an effective extension to the most promising method, iCBS.
Requirement List
Service providers' profit should be the m ain metric of optimization.
Consider both soft and hard SLA.
Manage the SLAs at the finest granularity level, i.e., per job basis.
Multiple SLA definitions corresponding to different job classes.
The complexity of the scheduling framewor k should be very small to cope with a hig h job arrival rate or bursts in the real system.
Architecture
Scheduling Algorithms
Cost- and Deadline-unaware Schedul ing
◦FCFS: First-Come First-Served.
◦SJF: Shortest Job First.
Scheduling Algorithms(Con t.)
Cost- and Deadline-unaware Schedul ing
◦FCFS: First-Come First-Served.
◦SJF: Shortest Job First.
Deadline-aware Scheduling
◦EDF: Earliest Deadline First.
◦AED: Adaptive EDF.
Avoid the domino effect under the overload situation, where all jobs misses the deadl ine.
Scheduling Algorithms(Con t.)
Cost- and Deadline-unaware Scheduli ng
Deadline-aware Scheduling
Cost-aware Scheduling
◦BEValue2
A modified version of EDF.
◦FirstReward
Highly sophisticated scheduling policy with high overload of O(n2).
◦iCBS
CBS
a heuristic-based cost-based sched uling policy.
The idea is to pick the query with the highest priority, which in tur n maximizes the expected global to tal profit.
iCBS incrementally maintains CBS p riority score with lower complexit y.
Scheduling Algorithms(Con t.)
Cost- and Deadline-unaware Schedul ing
Deadline-aware Scheduling
Cost-aware Scheduling
Cost- and Deadline-aware Schedulin g
◦iCBS-DH
”DH” stands for “Deadline Hint”
iCBS-DH
Extend iCBS into iCBS-DH by shift the SLA cost function.
◦Make it deadline-aware.
deadline t
C t
deadline t
t t
h , )
( cost
), ( ) cost
( cost
int
Experiment Setup
Server, database
Intel Xeon 2.4GHz, Two single-core CPUs, 1 6GB memory.
MySQL 5.5, InnoDB 1.1.3, 1GB bufferpool.
Dataset, query
TPC-W 1GB dataset.
6 query templates chosen from the TPC-W wo rkload.
Open-system workload, Poisson arrival.
Experiment Setup(Cont.)
Runs
5 seconds per run (>10K queries finished).
Each data point: the average of five repea ted runs.
Query execution time estimate
SJF, FirstReward, BEValue2, iCBS, iCBS-DH need it.
Estimate from history: Mean+StandardDeviat ion.
SLA Design
DTH code
CostDensity, CostStepTime, HardDeadlineTim e
◦E.g. DTH = 112
Varying SLA and Deadlines
DTH = 11x
◦iCBS-DH performs the best.
◦iCBS: low violation when deadline is t he same as or later than cost step(112 ,113), but high violation if not(111)
Varying SLA and Deadlines(Cont.)
DTH = 11x
◦iCBS-DH has high cost when cost step is eariler than deadline(113).
Hint cost: $1,000
Varying Portion of Deadline-Havin g Queries
DTH = 111
◦iCBS-DH perform the best.
◦EDF sees domino effect with high port ion of queries with deadlines.
Varying Portion of Deadline-Havin g Queries(Cont.)
DTH = 111
◦(FirstReward not shown)
◦iCBS-DH perform the best.
Varying Load
Load=arrival rate*average executio n time
◦iCBS-DH perform well under overload.
Varying Load(Cont.)
Load=arrival rate*average executio n time
◦iCBS-DH performs well on cost.
Varying Deadline Hint Cost
◦High hint cost reduce violations.
◦When deadline is earlier than cost st ep(111,115), DeadlineHint-to-Violatio n effect becomes more sensitive.
Varying Deadline Hint Cost(Cont.)
◦Cost performance gets worse with higher hint cost value.
◦Emphasis on deadline(113) => less attent ion on the cost step, leading to high SL A cost
Conclusion
Presented workload scheduling unde r two different types of SLAs, sof t and hard SLA.
Proposed a deadline- and cost-awar e scheduler called iCBS-DH.
Evaluated deadline and cost perfor mance of various scheduling polici es under a large range of SLA cost function and deadline types.