Ken Cantrell / NetApp Mark Rogov / EMC J Metz / Cisco October 21, 2015.
-
Upload
alberta-lane -
Category
Documents
-
view
214 -
download
0
description
Transcript of Ken Cantrell / NetApp Mark Rogov / EMC J Metz / Cisco October 21, 2015.
STORAGE PERFORMANCE BENCHMARKING:PART 2 – SOLUTION UNDER TEST
Ken Cantrell / NetAppMark Rogov / EMC
J Metz / CiscoOctober 21, 2015
© 2015 Storage Networking Industry Association. All Rights Reserved.
SNIA Legal NoticeThe material contained in this tutorial is copyrighted by the SNIA unless otherwise noted. Member companies and individual members may use this material in presentations and literature under the following conditions:
Any slide or slides used must be reproduced in their entirety without modificationThe SNIA must be acknowledged as the source of any material used in the body of any document containing material from these presentations.
This presentation is a project of the SNIA Education Committee.Neither the author nor the presenter is an attorney and nothing in this presentation is intended to be, or should be construed as legal advice or an opinion of counsel. If you need legal advice or a legal opinion please contact your attorney.The information presented herein represents the author's personal opinion and current understanding of the relevant issues involved. The author, the presenter, and the SNIA do not assume any responsibility or liability for damages arising out of any reliance on or use of this information.
NO WARRANTIES, EXPRESS OR IMPLIED. USE AT YOUR OWN RISK.
2
© 2015 Storage Networking Industry Association. All Rights Reserved.© 2015 Storage Networking Industry Association. All Rights Reserved.
About The Speakers
3
Mark RogovEMC
Systems Engineer@rogovmark
Dr. J MetzCisco
R&D Engineer@drjmetz
Ken CantrellNetApp
Manager Perf Engineering@kencantrelljr
© 2015 Storage Networking Industry Association. All Rights Reserved.
SPEC SFS® 2014See http://spec.org/sfs2014/ for details
SWBUILD
VDA
VDI
DATABASE
spec©
PROTOCOL SUPPORT• ANY TRANSPORT PROTOCOL
• TCP/IP, FC, ISCSI, ETC.• ANY NETWORK PROTOCOL
• NFS, SMB, ETC.• USES POSIX FILE OPERATIONS
METRICS• BUSINESS METRICS COUNTS
• SOFTWARE BUILDS, VIRTUAL DESKTOPS, VIDEO STREAMS, DATABASES
• RESPONSE TIME
© 2015 Storage Networking Industry Association. All Rights Reserved. 5
Storage Performance Benchmarking
SOLUTIONUNDER TEST
METRICS ANDTERMINOLOGY
FILECOMPONENTS
BLOCKCOMPONENTS
WORKLOADDEFINITIONS
© 2015 Storage Networking Industry Association. All Rights Reserved. 6
Storage Performance Benchmarking
METRICS ANDTERMINOLOGY
FILECOMPONENTS
BLOCKCOMPONENTS
WORKLOADDEFINITIONS
TODAY FUTURE WEBCASTSJULY 30, 2015
SOLUTIONUNDER TEST
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 7
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 8
Metrics and Terminology ReviewPart 1: http://www.snia.org/forums/esf/knowledge/webcasts (Both PDF and PPT available)
System 1 System 2
GRAPH FUN
CONTEXTMAKES
METRICSMATTER
OPSCOUNT EVERY
PROTOCOLOPERATION
PER SECOND
MB/SPAYLOAD SUM
OF EVERY OPERATION
PER SECOND
TERMINOLOGY
IOPSCOUNT EVERY IO OPERATIONPER SECOND
RESPONSETIME
TIME TARGETTAKES TO REPLY
TO AN IO
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 9
Last Time: Which is Better?
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 120000
5
10
15
20
25Response time per OPS
System 1 System 2
Operations per sec, OPS
Res
pons
e tim
e, m
s
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 10
Last Time: Which is Better?
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 120000
5
10
15
20
25Response time per OPS
System 1 System 2
Operations per sec, OPS
Res
pons
e tim
e, m
s
CONTEXTMAKES
METRICSMATTER
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 11
This Time: Why Are They Different?
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 120000
5
10
15
20
25Response time per OPS
System 1 System 2
Operations per sec, OPS
Res
pons
e tim
e, m
s
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 12
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 13
What is a Solution Under Test (SUT)?
HYPERVISOR
VM VM
HYPERVISOR
VM VM
Increasing Complexity
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 14
SUT Layers
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 15
Front- and Back-End Basic Difference
2KiB IOs
FRONT ENDBACK END
8KiB IOs
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 16
RUN ON
Is a Workload / Application a Part of the SUT?
WORKLOADS
HYPERVISOR
VM VM
HYPERVISOR
VM VM
SOLUTIONS UNDER TEST
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 17
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 18
Which Elements in the SUT Affect Performance?
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 19
What is Common Between All Of The Components?
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
PROCESSING ELEMENTS (CPU, ASIC)
MEMORY / CACHE
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 20
Answer To Ken’s Interview Question
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
CACHE
CLOUD
IOIO
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 21
3 Principles To Improve Performance
DO LESS WORKDO WORK FASTER INCREASE PARALLELISM
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 22
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 23
Latency Or “Whack A Mole Game”
0
2
4
6
8
10
12
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
7200 RPM
1000x
Res
pons
e Ti
me
(ms)
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 24
Disk Bottleneck
0
2
4
6
8
10
12
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
DISK BOUNDaka DISK BOTTLENECK
Res
pons
e Ti
me
(ms)
1000x
7200 RPM
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved.
Why Are SSDs So Compelling?
25
DRAM SSD HDD0
2,000,000
4,000,000
6,000,000
8,000,000
10,000,000
12,000,000
Res
pons
e Ti
me
(ns)
Random Write Latency
11ms
0.1ms
ENDMiB/SHEROMOLECACHESUTINTRO
25ns
© 2015 Storage Networking Industry Association. All Rights Reserved.
Why Are SSDs So Compelling?
26
DRAM SSD HDD
108
107
106
105
104
103
102
101
100
Res
pons
e Ti
me
(ns)
Random Write Latency
11ms
0.1ms
ENDMiB/SHEROMOLECACHESUTINTRO
25ns
100X
© 2015 Storage Networking Industry Association. All Rights Reserved. 27
Change SUT: Upgrade With SSDs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
SSDWAS 12ms
Res
pons
e Ti
me
(ms)
6x
DO WORK FASTER
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 28
Change SUT: Upgrade With SSDs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
Res
pons
e Ti
me
(ms)
DO WORK FASTER
ENDMiB/SHEROMOLECACHESUTINTRO
CONTROLLERBOUND
WAS 12ms
© 2015 Storage Networking Industry Association. All Rights Reserved. 29
Controller Bottleneck
00.05
0.10.15
0.20.25
0.30.35
0.40.45
0.5
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
VENDOR
Res
pons
e Ti
me
(ms)
~1.8x
CLIENTBOUND
WAS 0.7ms
DO WORK FASTER
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 30
Bottlenecks Always Exist
CLIENTS /HOSTS
STORAGECONTROLLER
00.05
0.10.15
0.20.25
0.30.35
0.40.45
0.5
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
BOTTLENECKMOVES
Res
pons
e Ti
me
(ms)
DO WORK FASTER
ENDMiB/SHEROMOLECACHESUTINTRO
WAS 12ms
© 2015 Storage Networking Industry Association. All Rights Reserved. 31
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 32
Back to the Original Problem …
0
2
4
6
8
10
12
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
Res
pons
e Ti
me
(ms)
DO LESS WORK
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 33
Avoid Slow Parts And Generate “Hero Numbers”
0
2
4
6
8
10
12
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ADDMORE CACHE
DON’T USE THIS
Res
pons
e Ti
me
(ms)
DO LESS WORK
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 34
Use More Cache
-1.11022302462516E-16
0.0999999999999999
0.2
0.3
0.4
0.5
0.6
0.7
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
Res
pons
e Ti
me
(ms)
WAS 12ms
CONTROLLERBOUND
DISKLATENCY IS STILL 11MS
BUTMASKED
DO LESS WORK
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 35
Caching Isn’t Just For Slow Drives
3,000 MILES
~70MS LATENCY
CLOUDSTORAGE
STORAGECONTROLLER
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
DO LESS WORK
© 2015 Storage Networking Industry Association. All Rights Reserved. 36
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 37
Latency Isn’t Everything
00.05
0.10.15
0.20.25
0.30.35
0.40.45
0.5
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
STILL A BOTTLENECKHOW?
Res
pons
e Ti
me
(ms)
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 38
MiB/s Limits
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
NOT ENOUGH DISKS
MiB
/s
“MORE”IS
BETTER
ENDMiB/SHEROMOLECACHESUTINTRO
INCREASE PARALLELISM
© 2015 Storage Networking Industry Association. All Rights Reserved. 39
MiB/s Limits
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
MiB
/s
“MORE”IS
BETTER
ENDMiB/SHEROMOLECACHESUTINTRO
INCREASE PARALLELISM
BOTTLENECKMOVES
© 2015 Storage Networking Industry Association. All Rights Reserved. 40
Network Bound Or Design Problem
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
INCREASE PARALLELISM
BOTTLENECK
© 2015 Storage Networking Industry Association. All Rights Reserved. 41
Network Bound Or Design Problem
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
INCREASE PARALLELISM
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 42
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 43
Which SUT Component Matters?
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECT
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 44
Which Component Matters?
CLIENTS /HOSTS
STORAGECONTROLLER
PHYSICALSTORAGE
FRONT-ENDCONNECT
BACK-ENDCONNECTTHE SLOWESTSUT COMPONENT
MATTERSMOST
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 45
Solution Under Test Review
DISK BOUND
CLIENTBOUND
SLOW COMPONENTMATTERS MOST
BOTTLENECKSALWAYS EXIST
3 PERFORMANCEPRINCIPLES
INCREASE PARALLELISM
DO LESSWORK
DO WORKFASTER
ENDMiB/SHEROMOLECACHESUTINTRO
© 2015 Storage Networking Industry Association. All Rights Reserved. 46
Storage Performance Benchmarking
METRICS ANDTERMINOLOGY
FILECOMPONENTS
BLOCKCOMPONENTS
WORKLOADDEFINITIONS
TODAY FUTURE WEBCASTSJULY 30, 2015
SOLUTIONUNDER TEST
ENDSUTINTRO CACHE MOLE HERO MiB/S
© 2015 Storage Networking Industry Association. All Rights Reserved. 47
After This Webcast
A PDF and a PPT of the slides for this and all previous parts of this Webcast series will be posted to the SNIA Ethernet Storage Forum (ESF) website and available on-demand
PPT and PDF: http://www.snia.org/forums/esf/knowledge/webcasts Storage Performance Benchmarking: Part 1 Recording: https://www.brighttalk.com/webcast/663/164323
A full Q&A from this webcast, including answers to questions we couldn't get to today, will be posted to the SNIA-ESF blog
http://sniaesfblog.org/
Follow us on Twitter @SNIAESF, @RogovMark, @KenCantrellJr, @DrJMetzNext Webcast – First Quarter 2016
“Storage Performance Benchmarking: Part 3”
QUESTIONS?
THANK YOU!