Carver: Finding Important Parameters for Storage System Tuning · Evaluation: Carver 0 20 40 60 80...

Carver: Finding Important Parameters for Storage System Tuning (FAST’20) 102/25/2020

Carver: Finding Important Parameters for Storage System Tuning

18th USENIX Conference on File and Storage Technologies (FAST’20)

Zhen Cao1, Geoff Kuenning2, and Erez Zadok1

1Stony Brook University; 2Harvey Mudd College


Outlinel Motivationl Backgroundl Design of Carverl Evaluationl Related Workl Conclusions & Future Work


Motivationl Tuning storage system parameters can provide significant

gainsuDefault settings are often sub-optimal

l Storage system parameter space is hugel Manual tuning vs. auto-tuning


Motivation (cont.)l More challenges to solve

u “Curse of dimensionality”uTuning efficiency is critical for storage systems

l Storage parameters come with different “importance”l Parameter selection is a critical part of auto-tuning

systems


Carver Overviewl Measurement for storage parameter importancel Sampling method

uPick a small subset of configurationsl Parameter selection algorithm


Key Contributionsl Quantitative analysis of importance of storage parameters on

system performancel Carver can pick important parameters efficientlyl Evaluated Carver on various datasets

uCarver can find a near-optimal subset of important parameters in all cases




l Storage systemu File system, underlying storage hardware and any layers between them

l Parametersu Configurable optionsu E.g., block size

l Parameter valuesu E.g., 1K, 2K, 4K (Ext4 block size)

l Configurationu Combination of parameter values u E.g., [Ext4, 4K, data=ordered]

l Parameter spaceu All possible configurations

Concepts


Challengesl Large parameter space

u Ext4: 59 parameters, 1037 configsu Distributed, large-scale

l Evaluation is time-consumingl Nontransferable tuning results

u Hardware, workload, etc.l Discrete and non-numeric

u Linux I/O scheduler: noop, cfq, deadline


Dimensionality Reductionl Feature extraction

uProject high-dimensional data into low-dimensional spacesuConstructed featuresuLose physical meaning of original features

l Feature selectionuDirectly selects a subset of features from the original onesuRelationship between parametersu Impact of parameters on target variable

§ Throughput




Carver: Measuring Parameter Importancel Categorical parameters & continuous target variablel Parameter Importance

uReduction in variance

l Conditional parameter importanceuCPI(P2 | P1 = p)uE.g., given P1 is fixed to “p”, find importance of P2


Carver: Samplingl Limit the number of evaluations

uEvaluation of storage configurations is costlyl Sampling

uLatin Hypercube Sampling§ Latin square: one sample in each row and each column§ Generalization to higher dimensions§ Proved useful in system analysis


Carver: Parameter Selection (1)

PI(X1) PI(X2) … PI(Xn)

Most Important Parameter


Carver: Parameter Selection (2)

2nd Most Important Parameter

(Assume X1 is as most important parameter.)

…CPI(X2|X1=y1) CPI(X2|X1=yk)…

CPI(X2|X1)

MaxCPI(Xn|X1=y1) CPI(Xn|X1=yk)…

CPI(Xn|X1)

Max


Carver: Parameter Selection (3)l Repeat the process until stopping criterion is metl Example stopping criteria:

uCertain number of parameters have been selecteduVariance within subsets of configurations (fixing values of selected

parameters) falls below a certain threshold.




Evaluation: Testbedl Hardware

u Intel Xeon quad-core 2.4GHz CPU, 24G RAM, 4 drives (SAS-HDD 500GB, SAS-HDD 146GB, 1 SATA-HDD, SSD)

l FilebenchuWorkloads: fileserver, mailserver, webserver, dbserver

l Two Settings:uS1: default working set size uS2: 4G RAM, 10G working set size


Evaluation: Settingsl Parameters spaces

u7 file system types, 4 workloadsu inode size, block size, block group, journal options, mount options,

special options, I/O scheduler, etc.l Methodology

uPre-collect benchmark results for all configurations§ 3+ runs for each configuration§ 500,000 data points

uEvaluate Carver on collected datasets


Evaluation: Parameter Importance

0

3*107

PI

PI

0

5*105

BlockSize

InodeSize

BlockGroup

AtimeOption

JournalOption

SpecialOption

I/OSchd.

Dev.

PI

CPI (X | journal)

CP

I

Ext4, S1, Fileserver-default


Setting Workload F/S #1 #2 #3 #4

S2 fileserver-10gb Ext4 Journal Option I/O Scheduler Inode Size -

S2 dbserver-10gb Ext4 Block Size Inode Size I/O Scheduler Journal Option

S2 webserver-10gb Ext4 Inode Size Flex BG Block Size Journal Option

S1 fileserver-def Btrfs Special Option Node Size Device -

S1 webserver-def Btrfs - - - -

Parameter Importance Summary


Evaluation: Carver

0

20

40

60

80

100

0.1 0.2 0.4 0.8 1.6 3.2 6.4

10 20 40 80 160 320 640

Per

centa

ge

of

Runs

(%)

Percentage of Dataset (%)

Running Time (min)

#1#2#3

Ext4, S1, Fileserver-default


Evaluation: Carver (Cont.)

0

20

40

60

80

100

4 8 16 32

20 40 80P

erce

nta

ge

of

Runs

(%)

Percentage of Dataset (%)

Running Time (min)

#1#2#3

Btrfs, S1, Fileserver-default




Related Workl Parameter selection for systems

uLasso [Aken et al.]uAdaptive Sampling [Duan et al.]uProbabilistic Reasoning [Sullivan et al.]

l Auto-tuning storage systemsuGenetic Algorithms, Deep Q-Networks, etc.uComparative Study [Cao et al.]




Conclusionsl Proposed Carver for efficiently selecting storage parametersl Provided thorough study of storage-parameter importancel Demonstrated Carver’s efficiency by testing it on small

fractions of the configuration space


Future Workl Extend Carver to support other parameter-selection

techniquesuGroup Lasso, ANOVA, etc.

l Evaluate Carver with more optimization objectives and larger storage-parameter spaces

l Extend Carver’s parameter selection algorithm into multi-objective optimization

l Integrate Carver with auto-tuning algorithms


Zhen Cao, Geoff Kuenning, and Erez Zadok

Carver: Finding Important Parameters for Storage System Tuning

Date post:	23-Sep-2020
Category:	Documents
Upload:	others
View:	0 times
Download:	0 times

Carver: Finding Important Parameters for Storage System Tuning · Evaluation: Carver 0 20 40 60 80...

Documents