QoS-Aware Storage Virtualization for Cloud File Systems Christoph - - PowerPoint PPT Presentation

1

QoS-Aware Storage Virtualization for Cloud File Systems Christoph Kleineweber (Speaker) Alexander Reinefeld Thorsten Schütt Zuse Institute Berlin

2

Outline

– Introduction – Performance Models – Reservation Scheduling – Evaluation – Conclusion

3

Introduction – Cloud Storage

– Difgerent interfaces

–

E.g. block device, object store, database, fjle system – Shared infrastructure

–

Multi tenants – Heterogeneous hardware

–

Difgerent devices (e.g. HDDs and SSDs)

–

Multiple generations – Difgerent customers' requirements

–

Capacity

–

Throughput (MB/s, IOPS)

–

Difgerent access patterns (sequential, random)

4

XtreemFS Architecture

– Object based fjle-systems

–

Separation of data and metadata storage

–

Scale capacity and performance by adding / removing OSDs

5

Quality of Service Extensions

– Application and device Modeling

–

Classify applications

–

Determine OSD performance before usage – Reservation scheduling

–

Find appropriate OSDs for a customer – Reservation enforcement (future work)

–

Proportional-fair-share queue on OSDs

6

Application and Device Behavior

– Applications have difgerent access patterns

–

Sequential

–

Random – Performance depends on pattern and device

–

Sequential vs. random

–

HDDs vs. SSDs – OSD stack causes uncertainties

–

Local fjle system properties

–

(Multiple) caches

–

Aging efgects → Use empirical performance models

7

Device Benchmarks (HDD)

Sequential Random Mixed

8

Device Performance - Observations

– Sequential throughput decreases with a growing

number of concurrent streams on HDDs

– Almost constant random performance for HDDs and

SSDs

– Mixed workloads result in performance degradation

• n HDDs

– SSDs are insensitive to mixed workloads and

concurrent access

9

Reservation Scheduling

– Find suffjcient OSDs for reservation

–

Requirements: capacity, sequential throughput (MB/s), random throughput (IOPS)

–

OSDs have limited resources

–

OSD resources depend on current schedule – Performance can be afgected by

–

Selecting OSDs with fjtting capabilities

–

Number of reservations per OSD

–

Striping over multiple OSDs – Minimize used OSDs

10

Multi-Dimensional Bin-Packing

– Formulate reservations and OSDs as vectors

–

Each dimension represents one resource (capacity, seq. throughput, random throughput)

–

Sum of reservations must not exceed OSD vector – Using T

• yoda's algorithm to solve bin-packing

–

Find OSD with minimized angle between OSD vector and sum of reservation vectors – Online problem

11

Scheduling Algorithm

12

XtreemFS Integration

– Using logical volume per reservation

–

Virtual fjle system with own namespace – Implemented new reservation scheduler service

–

Requests OSDs from directory server

–

Determines OSD performance by initial benchmarks

–

Computes schedule

–

Limits OSDs per volume by setting OSD selection policy – Reservations are enforced on OSDs by adding

proportional fair-share queue (future work)

13

Evaluation - Setup

– Discrete event based simulation

–

Reservations created at a rate of 0.5

–

Random and sequential reservations each with probability 0.5

–

Capacity, throughput and lifetime exponentially distributed with mean of 100 (GB, MB/s, IOPS, steps) – Storage cluster of SSDs and HDDs

–

Performance characteristics from device benchmarks – Baseline

–

Allocating dedicated OSDs for sequential access reservations

–

Random access reservations may share OSDs

14

Evaluation – Saved OSDs over Time

15

Evaluation – Saved OSDs vs. Active Reservations

16

Conclusion

– Object-based fjle systems are well suited for cloud

storage

– QoS-aware scheduling algorithm reduces resource

usage while guaranteeing performance

– Detailed knowledge about device characteristics is

necessary

17

Questions?

Website: xtreemfs.org Repository: github.com/xtreemfs Mailinglist: xtreemfs@googlegroups.com