Mass: Workload-Aware Storage Policy for OpenStack Swift Yu Chen , - - PowerPoint PPT Presentation

▶
mass workload aware storage policy for openstack swift
SMART_READER_LITE
LIVE PREVIEW

Mass: Workload-Aware Storage Policy for OpenStack Swift Yu Chen , - - PowerPoint PPT Presentation

Jun 10, 2023 544 likes •750 views

Mass: Workload-Aware Storage Policy for OpenStack Swift Yu Chen , Wei Tong, Dan Feng, Zike Wang Huazhong University of Science and Technology Outline Background and Motivation - Motivation study - Goals & Challenges Mass

slide-1
SLIDE 1

Mass: Workload-Aware Storage Policy for OpenStack Swift

Yu Chen, Wei Tong, Dan Feng, Zike Wang Huazhong University of Science and Technology

slide-2
SLIDE 2

Outline

  • Background and Motivation
  • Motivation study
  • Goals & Challenges
  • Mass
  • Evaluation
  • Conclusion

2

slide-3
SLIDE 3

Cloud object storage

  • Features
  • Flat address space
  • HTTP-based RESTful web APIs (CRUD)
  • Storage virtualization
  • Advantages
  • High availability
  • Flexibility
  • Simple data management

3

OpenStack Swift Amazon S3 Ceph

slide-4
SLIDE 4

Gap between workloads and storage

  • Multi-tenant workloads
  • Different access characteristics
  • Different requirements (latency & throughput)
  • Shared storage
  • Monolithic configuration
  • Same service level
  • Results in…

➡Limited workload performance ➡Low system efficiency

4

shared storage

Tenant 1 Tenant 2 Tenant 3

slide-5
SLIDE 5

Storage policy mechanism of Swift

  • Two-tier architecture
  • Access tier forwarding requests
  • Storage tier managing storage devices
  • Proxy server
  • Object ring
  • Storage node
  • Partition

5

Request forwarding

slide-6
SLIDE 6

Storage policy mechanism of Swift

  • Object rings
  • Key role of request forwarding
  • Consistent hashing
  • Two-level mapping
  • Storage policy mechanism
  • Creation of the particular object ring
  • Configurable n,m values

6

Two-level mapping of object ring

slide-7
SLIDE 7

Motivation study - advantages

  • Comparing with the monolithic setup

➡NOT similar performance level ➡Throughput: up to 8.5x increase ➡Latency: up to 33% decrease

  • Analysis
  • Isolated forwarding paths
  • Mitigating resource competition

7

better workload performance

slide-8
SLIDE 8

Motivation study - limitations

  • Stress tests
  • Varying request concurrency
  • Same storage policies
  • Performance results

➡Throughput reaching saturation ➡Latency increasing sharply

  • Indicates that…
  • Performance of intensive workloads has

room for improvement

8

Why?

slide-9
SLIDE 9

Goals & Challenges

Goals

  • Covering full-path of request
  • Workload-specific
  • Performance optimization
  • Dynamic mechanism

9

Enhanced storage policy mechanism

Challenges

  • Controlling request processing path
  • Workload classification
  • Request identification at storage layer
  • Policy adjustment at runtime
slide-10
SLIDE 10

Mass

  • Control & Data planes
  • Controller
  • Monitor
  • Substore
  • Workload classification
  • Access characteristics
  • Read-dominated, write-

dominated, read-write mixed

  • Request identification
  • Cross-layer tagging

10

Overall architecture

slide-11
SLIDE 11

Life cycle of a policy

i. Policy preparation

  • Monitoring
  • Workload classification
  • ii. Policy formulation
  • Triple: {tenant, ring, method}
  • iii. Policy deployment
  • Optimized request processing
  • iv. Policy execution

11

1 2 3 4 5 6

Component interaction

slide-12
SLIDE 12

Two-level processing optimizations

  • Substore-level policy
  • Workload-specific
  • Performance optimization
  • Programmable

12

  • Storage node level policy
  • Priority-based queuing
  • System efficiency

Read-dominated Write-dominated Read-write mixed

Workload type Performance requirement Policy Read- dominated Latency Cache Write- dominated Throughput Batch Read-write mixed Latency & Throughput Merge

Non-first replica write

High Low Priority

slide-13
SLIDE 13

Dynamic policy mechanism

  • Workload changes
  • External
  • Internal
  • Validation
  • Policy adjustment

13

  • Improper resource allocation
  • Policy overhead
  • Insertion
  • Deletion
slide-14
SLIDE 14

Evaluation setup

  • Cluster
  • 2 proxy servers
  • 5 storage nodes
  • 3 workload generators

14

  • Workload
  • Synthetic workloads
  • Real-world traces
  • Storage setup
  • Default: Swift’s original policies
  • Crystal: Manual workload-specific policies
  • MASS: Dynamic workload-specific policies

& priority-based queuing

Synthetic workloads Idiada trace Arctur trace 79.99% write 99.97% read

slide-15
SLIDE 15

Effectiveness of policy

15

Workload A 93.7% lower latency

Workload B 81.6% lower latency 191.2% higher throughput

Workload C 231.5% higher throughput

  • Overall system performance

➡154.3% higher throughput and 67.8% lower latency

slide-16
SLIDE 16

External workload change

16

  • Workload A

Workload C

  • Three-stage test
  • Baseline & A-dominated & C-dominated
  • Workload A: 61.9% lower latency
  • Workload C: 55.2% higher throughput

Workload A

slide-17
SLIDE 17

Internal workload change

17

  • Comparing with
  • Default: average 61.3% promotion
  • Crystal: average 37.6% promotion
  • Comparing with
  • Default: average 59.4% promotion
  • Crystal: average 39.3% promotion
slide-18
SLIDE 18

Conclusion

  • Original storage policy mechanism
  • Poor performance of intensive workloads
  • Unable to react to workload changes
  • We propose Mass to enhanced flexible polices
  • Covering full storage path
  • Workload-aware optimizations based on access characteristics
  • Dynamic policy adjustment
  • Better workload performance and system efficiency

18

slide-19
SLIDE 19

Thanks! Q&A

Email: chloe_chen@hust.edu.cn