Upload
haque
View
216
Download
3
Embed Size (px)
Citation preview
Cisco HyperFlex Deep Diveand Performance
Brian Everitt, Technical Marketing Engineer
BRKCOM-2004
Cisco HyperFlex 2.0 All Flash Systems Offer High Performance, Low Latency, and Consistent Results
• Introduction
• What is HyperConverged?
• Cisco HyperFlex Overview
• What’s New in HX 2.0
• Architecture
• Performance and Benchmarks
• Conclusion
Agenda
Brian Everitt
• Technical Marketing Engineer with the CSPG BU
• Focus on HX performance, benchmarking, quality and SAP solutions
• 5 years as a Cisco Advanced Services Solutions Architect
• Focus on SAP Hana appliances and TDI, UCS and storage
• First time at Cisco Live EU
About Me…
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
What Is Hyperconvergence?
HyperconvergenceOn-Demand
Agile
Efficient
Simple
Scalable
Enterprise Ready
Cost Effective
+
+
+
+
+
+
+
CPU
MEM
CPU
MEM
NO LEGACY STORAGE CONTRUCTS
APP APP APP APP APP•••••••HYPERCONVERGENCE
SERVERS
1-3yr Planned
Large Footprint
Inefficient
Complex
SERVER VIRTUALIZATION
APP APP APP APPOn-Demand
Agile
Efficient
Simple
Vendor E Vendor N Vendor P
••••••••••••
•••••••••••••••
+
+
+
+
–
–
–
–
Tra
ditio
nal N
AS
CPU
MEM
Hybri
d S
tora
ge
All F
lash S
tora
ge
CPU
MEM
CPU
MEM
CPU
MEM
APP
Traditional Approach
STORAGE
6BRKCOM-2004
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
HyperFlex Key Customer Benefits
Efficient
AdaptableScale-As-You-Grow
AdaptiveScaling
Non-Stop Infrastructure
Built on the UCS Platform
FlexibleDeployment
Always-on Data Optimization
CompleteHyperconvergence
IntuitiveManagement
CloudSpeed
All 3 Layers of Network, Compute
and Storage Intelligently Integrated
Into a Single Solution for <60 Minute
Deployment
Eliminate Infrastructure Silos and Use
Existing Management Tools to Manage
and Automate Operations
Instantly Provision, Clone or Snapshot Applications
Unified Management Using Existing
Skillsets in the Organization
Variety of Configurations to Handle
Diversified Set of Workloads
Inline Dedupe and Compression Ensuring Hyper-Efficient Resource Utilization
Add Resources Non-disruptively
and Scale Performance Linearly
in Small Increments
Scale Compute or Capacity to Match
Application Needs
Self-Healing Fabric Based Hyperconvergence with Cloud Monitoring
Agile
12BRKCOM-2004
Cisco HyperFlex SystemsHX-Series
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
What is HyperFlex?
1 Complete Hyperconvergence
Unified Compute and
Network Infrastructure Wide Array of Rack and Blade Form Factors
Integrated, High Performance
Network Fabric
Programmable Automation of
all Hardware
Cisco UCS: The Ideal Platform for Hyperconvergence
BRKCOM-2004 14
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
1 Complete Hyperconvergence
Unified Compute and
Network InfrastructureCisco HyperFlex
Cisco HX Data Platform
Data Services and Storage Optimization
2 Next Gen Data Platform
Designed for
Distributed Storage
What is HyperFlex?
BRKCOM-2004 15
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco HyperFlex Systems
1
3
Complete Hyperconvergence
Unified Compute and
Network Infrastructure
Part of a Complete Data Center StrategyElastic and Secure at Enterprise Scale
Cisco HyperFlex
Cisco HX Data Platform
Data Services and Storage Optimization
2 Next Gen Data Platform
Designed for
Distributed Storage
Cisco One
Enterprise Cloud SuiteCisco ACI
Cisco Security
BRKCOM-2004 16
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Mainstream Computing
Scale Out
Cisco UCS: Platform for All Architectures
HyperconvergedInfrastructure
Converged Infrastructure
Unified ManagementSingle Control Plane
Single API
Edge Cloud
ROBO
UCS S-Series
Storage
Fourth Generation
UCSHyperFlex Systems
UCS Mini
E-Series
C-Series
Rack Servers
UCS Integrated
Infrastructure Solutions
Single Operational Model
UCS Manager UCS Director Enterprise Cloud Suite
BRKCOM-2004 17
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco HyperFlex Hybrid Configurations
HX240c Cluster
Capacity-heavy 3–8 Node Cluster
(VSI: IT/Biz Apps, Test/Dev)
Per-Node1 x 1.6 TB Cache SSD
up to 23 x 1.2 TB HDDs
SD Card/120GB Back SSD
(Boot/Housekeeping)
HX220c Cluster
Smallest Footprint 3–8 Node
Cluster (VDI, ROBO)
Per-Node1 x 480 GB Cache SSD
6 x 1.2 TB HDDs
SD Card/120GB SSD
(Boot/Housekeeping)
Hybrid Clusters
Compute-heavy Hybrid
(Compute Bound Apps/VDI)
Up to 8 Compute Nodes
Blade or RackB200 M4, C220 M4 or C240 M4 (Supported)
Local Disk, SD Card or SAN Boot
3-8 Node HX220 or HX240 Cluster
4
1
5
2
6
3
7 8 UCSC220 M4
Intel
Inside
XEON
Console!
C220 M4
!
Intel
Inside
XEON
UCSC240 M4
Con
sole
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
1 6 12 18 24
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
C240 M4
B200 M4
BRKCOM-2004 18
What’s New In HX 2.0
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Introducing All Flash HyperFlex Systems
High Performance
High IOPS, High Throughput,
Consistent, Low Latency
No Compromise Storage
Efficiency
Always-on, inline dedupe
and inline compression
Instantaneous, space
optimized snapshots and
clones
Performance Sensitive Workloads
Support for Databases other performance sensitive workloads
BRKCOM-2004 20
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco HyperFlex All Flash Configurations
HXAF240c Cluster
Capacity-heavy 3–8 Node Cluster
(VSI: IT/Biz Apps, Test/Dev)
Per-Node1 x 800 GB WL Cache SSD
up to 10 x 960 GB/3.8 TB SSDs
SD Card/120GB Back SSD
(Boot/Housekeeping)
HXAF220c Cluster
Smallest Footprint 3–8 Node
Cluster (VDI, ROBO)
Per-Node1 x 800 GB WL Cache SSD
6 x 960 GB/3.8 TB SSDs
SD Card/120GB SSD
(Boot/Housekeeping)
Hybrid Clusters
Compute-heavy Hybrid
(Compute Bound Apps/VDI)
Up to 8 Compute Nodes
Blade or RackB200 M4, C220 M4 or C240 M4 (Supported)
Local Disk, SD Card or SAN Boot
3-8 Node HXAF220 or HXAF240 Cluster
4
1
5
2
6
3
7 8 UCSC220 M4
Intel
Inside
XEON
Console!
C220 M4
!
Intel
Inside
XEON
UCSC240 M4
Con
sole
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
1 6 12 18 24
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
UC
S-H
DD
30
0G
I2F
10
5
15
K S
AS
30
0 G
B
!
C240 M4
B200 M4
BRKCOM-2004 21
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
HX All Flash Improvements
Comparison between HX Hybrid and HX All Flash for 70/30 8K Random Read Write Workload
with 4.8TB Working Set Size, vmdks were initialized with writes
Pre-release Performance Data
Up to 6X IOPS
and 1/5th
Latency
6X Better Perf Consistency
3.2XCapacity / Density
BRKCOM-2004 23
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
What’s New With Cisco HyperFlex 2.0?
Unified UCS ArchitectureSupport for adding HyperFlex to an
existing UCS Domain
Enhanced User ExperienceImproved Installer User Interface and
Capabilities, Additional Snapshot API and Sizer improvements
New Hardware Qualifications for 3rd Gen Fabric Interconnects and
VIC cards (Post FCS)
HXAF220 and HXAF240 Converged Nodes
BRKCOM-2004 24
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public 25BRKCOM-2004
Adding HX in an existing Cisco UCS Domain
HyperFlex
SLOT
1
SLOT
5
SLOT
3
SLOT
7
SLOT
2
SLOT
6
SLOT
4
SLOT
8
!
UCS 5108
OK FAIL OK FAIL OK FAIL OK FAIL
! ResetConsole
UCS B200 M4
! ResetConsole
UCS B200 M4
! ResetConsole
UCS B200 M4
! ResetConsole
UCS B200 M4
Non-HX
UCS
• Adding HyperFlex to pre-existing
UCS Domains with general UCS
equipment and workloads
• Will be supported 2.0
• HyperFlex QoS requirements
may conflict with existing config
• Consult your CSE and TAC for
guidance and assistance
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
External Storage Configuration
Added sections for FC / iSCSI storage
• Automates creation of vNICs, vHBA and necessary profiles in UCSM
• Installer automatically fixes VMDirectpath PCI ordering
• Ready for storage configuration inside of ESXi
BRKCOM-2004 26
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Sizer Improvements
• End to end sizing guidance: compute, capacity and performance
• Rapid iterative development: Lots of improvements / additions coming
• Next: Support for all flash nodes and support for DBs
Link: https://hyperflexsizer.cloudapps.cisco.com/
(CCO Logon Required)
BRKCOM-2004 28
Cisco HyperFlex Architecture Overview
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Building on the Right FoundationCisco HX Data Platform
Built From the Ground Up for
Hyperconvergence
Distributed Log-Structured File
System Designed for Scale-out,
Distributed Storage
Advanced Data Services (Snapshots,
Clones) and Data Optimization
(Inline Dedupe, Compression) Without
Trade-offs
Better Flash Endurance and
Disk Performance
Computing, Storage,
Networking, and Hypervisor Integration
No Reliance on Legacy
Filesystems or
Technology
Distributed File system
Local Disks
Unique
Architecture
Local Disks Local Disks Local Disks
BRKCOM-2004 31
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
HYPERCONVERGED DATA PLATFORMHYPERCONVERGED DATA PLATFORMHYPERCONVERGED DATA PLATFORMHYPERCONVERGED DATA PLATFORM
Hyperconverged Scale Out and Distributed File System
CONTROLLERVMHYPERVISOR
VM VM VM
Start with as Few as
Three Nodes
Hyperconverged Data Platform
Installs in Minutes
Add Servers, One or More
at a Time
Linearly Scale Compute, Storage
Performance, and Capacity
Distribute and Rebalance Data Across Servers Automatically
Retire Older
Servers
CONTROLLERVMHYPERVISOR
VM VM VM
CONTROLLERVMHYPERVISOR
VM VM VM
CONTROLLERVMHYPERVISOR
VM VM VM
CONTROLLERVMHYPERVISOR
VM VM VM
BRKCOM-2004 32
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Software Modules Inside a Server
Controller VM Has Direct Access to Drives
VAAI Plugin Offloads Snapshots and Clone Operations
IO Visor Module Presents NFS to ESX and Stripes IO
DATASTORE/VOLUME
HYPERVISOR
HDD
HDD
SDD
SDD
SSD SSD SSD SSD
CONTROLLERVM
IO Visor
VAAI
VM VM VM VM VM
BRKCOM-2004 33
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Read Cache
Log FS
L1
Read
Distributed Objects
All Flash
DRAM
Read Cache
Optimizing for All Flash: Read Path
Log FS
L2
L1
Read
Distributed Objects
Hybrid L1 (DRAM) Cache:
• Metadata and data
• Helps with small bursts
L2 (SSD) Cache:
• De-duplicated Read Cache
• Use aggregate performance of all Read Cache SSDs
SSD
DRAM
No L2 (SSD) Cache:
• No flash on flash
advantage
• Use aggregate
performance of all SSDs
BRKCOM-2004 34
SCVM RAM:
• Unified capacity disk
index cache
• Improvements by
avoiding additional index
reads
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Life of a Write IO: Write Log
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
AA’’ A’
ACKA
1. Application writes data (A) to OS
2. VM writes to VMDK on datastore
3. Write intercepted by IOvisor and sent to primary controller
4. Primary data is committed to WL SSD
5. Simultaneously, data replicated to two other nodes (RF3) and committed to their WL SSD
6. Acknowledgement sent to VM
7. Data is successfully committed and protected
B B’ B’’
CC’’C’
D D’D’’
Write Log Capacity Used
Size: 2G Hybrid / 48G AF
BRKCOM-2004 35
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Life of a Write IO: Flush
AA’’ A’
1. When the active write log of a node is full, the active log becomes the secondary log and flushes
2. Primary data copies are deduplicated then compressed
3. Data is then three way written (RF3) to SSD / HDD on 3 separate nodes
4. After all data is successfully committed, data is purged from the secondary log on the caching SSD
5. In hybrid systems, recent writes are added to the read cache area of the SSD during flush
B B’ B’’
CC’’C’
D D’D’’
Write Log Capacity Used
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
BRKCOM-2004 36
A’A’’AA
BBB’B’’
CC’’C’ D’’ D D’
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
CONTROLLERVM
IOvisor
VMVM
ESXiDS
Life of a Read IO
ESXi
Read A
1. Application reads data from OS filesystem
2. VM reads from VMDK on datastore
3. Read intercepted by IOvisor and sent to primary controllera) First check active write log
b) Next check passive write log (the duplicated write log of the other nodes)
c) Look in L1 (DRAM) cache
d) Look in L2 (SSD) cache [Not in AF models]
e) Retrieve from SSD / HDD – Decompress, read into L1, returned to IOvisor, copied into L2 [Not in AF models]
4. Return data to VM
Active WL
Passive WL
L1
L2
SSD (AF) / HDD (Hybrid)
BRKCOM-2004 37
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Integrated Management and Data Services
Dynamic Data Distribution
Continuous Data Optimization
Independent Scaling of Compute and Capacity
HyperFlex Design Principles
BRKCOM-2004 38
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Scale Compute
HX Data Platform
Add NodesScale Capacity Within Nodes
HX Data Platform
CONTROLLERHYPERVISORCONTROLLERHYPERVISOR CONTROLLERHYPERVISOR CONTROLLERHYPERVISOR
IOVisorIOVisor
IOVisorIOVisor
IOVisorIOVisor
IOVisorIOVisor
VM VM VM VM
Non-HyperFlex Hosts Can Connect to
Storage with IOVisor
VM VMVM VM VMVM VM VMVM VM VMVM
IOVisor
VM VM
IOVisor
VM VM
Independent Scaling of Compute and Capacity
BRKCOM-2004 39
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Dynamic Data Distribution
• HX Data Platform stripes data across all nodes simultaneously, leveraging cache across all SSDs for fast writes, and HDDs for capacity, via the low-latency UCS internal network
• Balanced space utilization: no data migration required following a VM migration, and no hotspots
Systems Built on Conventional File
Systems Write Locally, Then Replicate,
Creating Performance Hotspots
CONTROLLERHYPERVISORHYPERVISOR CONTROLLERHYPERVISOR CONTROLLERHYPERVISOR
VM VMVM VM VMVM VM VMVM
HX Data Platform
VM VMVM
CONTROLLERCONTROLLER
BRKCOM-2004 40
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
DATASTORE DATASTORE
Non-Disruptive Operations
• Stripe blocks of a file across servers
• Replicate one or two additional copies
to other servers
• Handle entire server or disk failures
• Restore back to original number of copies
• Rebalance VMs and data post
replacement
• Rolling “one-click” software upgrades
CONTROLLERHYPERVISOR
VM VM VM
CONTROLLERHYPERVISOR
VM VM VM
CONTROLLERHYPERVISOR
VM VM VM
CONTROLLERHYPERVISOR
VM VM VM
File.vmdk
D1 E1A1 B1 C1B2 A2 A3C2 C3 D2D3 E2E3 D1E1 B3 B3
EDCBA
BRKCOM-2004 43
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Hybrid Capacity Options
3 Node
HX220c
HX240c
4 Node 5 Node 6 Node 7 Node 8 Node
Assumes: Full HDD population at RF3.
Note: The above calculations are before deduplication & compression. Effective capacity will be higher.
Consult with your Cisco CSE for the latest sizing & design guidance.
6.02 TiB 8.03 TiB 10.04 TiB 12.04 TiB 14.05 TiB 16.06 TiB
23.09 TiB 30.79 TiB 38.48 TiB 46.18 TiB 53.88 TiB 61.58 TiB
Usable Cluster Capacity
BRKCOM-2004 45
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
All Flash Capacity Options
3 Node
HXAF220c
HXAF240c
4 Node 5 Node 6 Node 7 Node 8 Node
Assumes: Full SSD population at RF3.
Note: The above calculations are before deduplication & compression. Effective capacity will be higher.
Consult with your Cisco CSE for the latest sizing & design guidance.
4.82 TiB/ 19.28 TiB
6.43 TiB/ 25.71 TiB
8.03 TiB/ 32.14 TiB
9.64 TiB/ 38.56 TiB
11.25 TiB/ 44.99 TiB
12.85 TiB/ 51.42 TiB
8.03 TiB/ 32.41 TiB
10.71 TiB/ 42.85 TiB
13.39 TiB/ 53.56 TiB
16.07 TiB/ 64.27 TiB
18.75 TiB/ 74.99 TiB
21.42 TiB/ 85.70 TiB
Usable Cluster Capacity with 960GB/3.8TB SSDs
BRKCOM-2004 46
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
HyperFlex 2.0 Capacity Considerations
• SSD reliability is significantly better than HDD; from 2-10 times lower annual replacement rate in studies by major online providers.
• HyperFlex installer no longer defaults to RF3 for an all-flash system, forcing the end user to make a decision.
• Consider RF2 for all-flash systems that do not require triple data copies or multiple component redundancy.
• RF2 shows 15-40% reduced write latencies in lab testing.
• RF2 can provide clusters up to 128 TiB maximum capacity.
• Keep overall cluster capacity consumed below 70% for best performance.
47BRKCOM-2004
HyperFlex Performance and Benchmarks
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
HX Hybrid to AF Performance Comparison
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
0.4GB 100GB 800GB 2TB
Per
form
ance
Mu
ltip
le
Effective Working Set Size per NodeTests done on 4 x HX240
Performance With Varying Working Set
All Flash Performance Multiple
BRKCOM-2004 49
HX240 hybrid read cache 1.2 TB per node
• AF offers performance improvements and lower latency even at small working sets
• AF offers huge performance and latency improvements for working sets that do not fit in the hybrid model read cache
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 1
Base Performance Using vdbench Curve Tests
HX-All Flash HX Hybrid
Version HX 2.0.1 HX 1.8.1c
HW Configuration 4 x HXAF240, each with:
• WL: 1 x 400G
• Data: 8 x 960G
4 x HX240, each with:
• Cache/WL: 1 x 1.6TB SSD
• Data: 23 x 1.2TB 10Krpm SAS
HDD
Benchmark Setup • 3 VMs per host
• Each VM has 1 x 400GB vmdk
• 4.8 TB working set
• Priming: 1MB writes
• Tests with 32 threads
BRKCOM-2004 50
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 1
Base Performance Using vdbench Curve Tests8K, 100% Random Write Workload
0
5
10
15
20
25
30
35
40
0 5000 10000 15000 20000 25000 30000 35000 40000
Late
ncy
(m
s)
IOPS
8K, 100% Write
HX Hybrid
HX All Flash
3.3X Higher
IOPS at 3.5X
Lower latency
BRKCOM-2004 51
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
8K, 70% Read, 30% Write, 100% Random Workload
6.3X Higher
IOPS at 5X
Lower latency
0
5
10
15
20
25
30
0 10000 20000 30000 40000 50000 60000 70000 80000 90000
Late
ncy
(m
s)
IOPS
8K, 70% Read, 30% Write
HX Hybrid
HX All Flash
Benchmark 1
Base Performance Using vdbench Curve Tests
BRKCOM-2004 52
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
8K, 100% Random Read Workload
10.7X Higher
IOPS at 5.7X
Lower latency
0
1
2
3
4
5
6
7
8
9
0 50000 100000 150000 200000 250000 300000
Late
ncy
(m
s)
IOPS
8K, 100% Read
HX Hybrid
HX All Flash
Benchmark 1
Base Performance Using vdbench Curve Tests
BRKCOM-2004 53
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 1
Base Performance Using vdbench Curve TestsVariability Measured By Standard Deviation of Latency
HX Hybrid HX All Flash Ratio
8K, 100% Write39.4 11.6 3.4
8K, 70 / 30 RW74.3 5.1 14.6
8K, 100% Read16.2 1.1 14.2
Significantly
more
consistent
performance
BRKCOM-2004 54
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 2
SQL Performance Using vdbench Curve TestsHX-All Flash HX Hybrid
Version HX 2.0.1 HX 1.8.1c
HW Configuration 4 x HXAF240, each with:
• WL: 1 x 400G
• Data: 8 x 960G
4 x HX240, each with:
• Cache/WL: 1 x 1.6TB SSD
• Data: 23 x 1.2TB 10Krpm SAS
HDD
Benchmark Setup • 3 VMs per host
• Each VM has 1 x 400GB vmdk
• 4.8 TB working set
• Priming: 1MB writes
• Test with 32 threads• Read Block Size Mix: 4k 2.39%, 8k 53.57%,16k 4.84%, 28k 2.64%, 64k 28.75%,120k
1.79%, 220k 1.41%,492k 4.24%, 524k 0.37%
• Write Block Size Mix: 4k 27.85%, 8k 35.24%,16k 4.82%, 28k 2.58%, 64k 26.35%,
116k 0.6%, 244k 1.28%, 488k 0.38%, 536k 0.89%, 1032k 0.01%
• Overall Avg Blocksize: 43.4K; Overall Read %: 68.5%
BRKCOM-2004 55
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 2
SQL Performance Using vdbench Curve Tests
4.5X Higher IOPS
at 4.6X Lower
latency with 7.3X
better consistency
0
5
10
15
20
25
30
35
0 5000 10000 15000 20000 25000 30000
Late
ncy
(ms)
IOPS
SQL Profile Using vdbench
HX Hybrid HX All Flash
BRKCOM-2004 56
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
SQL Performance Using HammerDBHX-All Flash
Version HX 2.0.1
HW Configuration 4 x HXAF240, each with:
• WL: 1 x 800G
• Data: 10 x 960G
Benchmark Setup • 1 SQL Server 2014 VM
• Windows Server 2012 R2
• 28 vCPU
• 64 GB RAM
• ~500 GB DB created by HammerDB with 5000 warehouses
• 500 User Test
• 5 minute ramp up, 60 minute test
BRKCOM-2004 57
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public 58BRKCOM-2004
HammerDB SQL TPC-C OLTP Test Results
HX All-Flash
SQL Transactions Per Minute 1.475 million
Average IOPs 15,141
Average Read Latency 0.78 ms
Read Latency Standard Deviation 0.32
Average Write Latency 1.36 ms
Write Latency Standard Deviation 0.36
Performance limited by CPU
power, all CPU cores running
at ~100% long before
storage performance or
latency becomes an issue.
Competitive Performance and Benchmarks
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco HyperFlex HybridRuns more VMs per Cluster
0
20
40
60
80
100
120
140
160
Cisco HyperFlex Vendor X Vendor Y Vendor Z
Total VMs per 4 node cluster 70/30 R/W 4K
Number of VMs per Cluster Write Latency
• Cisco HyperFlex can
run more VMs per
cluster than other
vendors
• Tests found the number
of VMs other vendors
can run with similar
Write Latency as HX
• Simulates heavy VM
load
Up to 3X
the VMs
on HX
BRKCOM-2004 60
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco HyperFlex Hybrid140 VMs per 4 Node Cluster Latency Averages
0
5
10
15
20
25
Cisco HyperFlex Vendor X Vendor Y
70/30 R/W 4K
Write Latency Read Latency
• Cisco HyperFlex has
lower R/W latency for
140 VMs per cluster
than other vendors
• Simulated heavy VM
load
• Vendor Z could not run
the load @ 140 VMs
HX has up to
3X + lower
latency
BRKCOM-2004 61
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
4K Random
70% RD/ 30%
WR
R/W Latency
8K Random
70% RD/ 30%
WR
R/W Latency
HX 136243 1.216 / 4.166 120901 1.208 / 4.936
Vendor Y 121463 1.467 / 5.997 97437 1.704 / 7.148
HX vs Vendor Y 12.2% 20.4 / 30% 24.1% 29.3 / 30.8%
4K Random
70% RD/ 30%
WR
R/W Latency
8K Random
70% RD/ 30%
WR
R/W Latency
HX 136243 1.216 / 4.166 120901 1.208 / 4.936
Vendor Y 99951 1.355 / 6.682 70409 2.889 / 7.910
HX vs Vendor Y 36.3% 7.4 / 37.4% 71.7% 58.8 / 37.9%
Dataset with no dedupe and no compression savings
Dataset with dedupe and compression
• 2.8TB Working Set
• HX significantly better
at write Latency
• HX much better IOPs
with dedupe and
compression
• HX better performance
at 8K block sizes
• As the working set
increases the
performance continues
to drop
Benchmark 3
Base Performance Using vdbench Curve TestsHX All Flash vs Vendor Y All Flash (Unprimed Dataset)
BRKCOM-2004 62
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Benchmark 3
Base Performance Using vdbench Curve Tests8K, 100% Random Write Workload
• 61% Higher
IOPS
• 42% Lower
latency
• 60% lower
standard
deviation
0
2
4
6
8
10
12
14
16
18
0 5000 10000 15000 20000 25000 30000 35000 40000
Late
ncy
(m
s)
IOPS
8K, 100% Write
HX All Flash
Competitor Y
BRKCOM-2004 63
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
8K, 70% Read, 30% Write, 100% Random Workload
• 11% Higher
IOPS
• 24% Lower
latency
• 59% lower
standard
deviation
0
1
2
3
4
5
6
0 10000 20000 30000 40000 50000 60000 70000 80000 90000
Late
ncy
(m
s)
IOPS
8K, 70% Read, 30% Write
HX All FlashCompetitor Y
Benchmark 3
Base Performance Using vdbench Curve Tests
BRKCOM-2004 64
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
8K, 100% Random Read Workload
• 43% Higher
IOPS
• 34% Lower
latency
• 59% lower
standard
deviation
0
0.5
1
1.5
2
2.5
0 50000 100000 150000 200000 250000 300000
Late
ncy
(m
s)
IOPS
8K, 100% Read
HX All Flash
Competitor Y
Benchmark 3
Base Performance Using vdbench Curve Tests
BRKCOM-2004 65
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
0
5000
10000
15000
20000
25000
30000
1 7
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
10
3
10
9
11
5
12
1
12
7
13
3
13
9
14
5
15
1
15
7
16
3
16
9
17
5
18
1
Intervals
Vendor Y
VM1 VM2 VM3 VM4 VM5 VM6
VM7 VM8 VM9 VM10 VM11 VM12
IOPS in a 70/30 Workload – 1 Hour Duration
0
5000
10000
15000
20000
25000
30000
1 8
15
22
29
36
43
50
57
64
71
78
85
92
99
10
6
11
3
12
0
12
7
13
4
14
1
14
8
15
5
16
2
16
9
17
6
Intervals
HX
VM1 VM2 VM3 VM4 VM5 VM6
VM7 VM8 VM9 VM10 VM11 VM12
Benchmark 3
Base Performance Using vdbench Curve Tests
BRKCOM-2004 66
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
All Flash POC Best PracticesAll Flash Performance Test Setup
1. Initialize the disks• More realistic - Applications don’t read data they haven’t written
2. Longer Test Duration:• SSDs / All flash systems show very high performance initially – which not realistic
• Tests should be long running (12 – 24 hours depending on write load)
3. Larger Working set size• Ability to support larger working set size (several TBs / node) is a key differentiator for all flash
• For smaller working set size hybrid will perform similar to all flash (as data is coming from flash)
4. Enable Dedupe and compression on competition (On by default on HX)• Most enterprise workloads see significant storage efficiency savings – this is not optional
• Expecting users to choose this on a workload by workload basis is not realistic with 100s of VMs
BRKCOM-2004 67
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
All Flash POC Best PracticesAll Flash Performance Test Setup
5. Workload Read / Write Mix:• Reasonable write percentage (20-35% at least) in the workload
• 100% - 95% read workloads are not realistic in most environments
6. Similar hardware configuration:• Pay careful attention to the BOM – CPU, Memory and SSD (type and count for cache/WL and
persistent data SSDs)
• Additional performance may come at a trade-off: cost or manageability
7. Similar software configuration:• Pay careful attention to the resiliency setting – that needs to be identical
• Comparing Replication Factor = 2 performance with Replication Factor = 3 performance is not comparing apples to apples
BRKCOM-2004 68
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
All Flash POC Best Practices1. What to look for in performance results:
• IOPS / Throughput
• Latency – not all workloads show sub millisecond latencies (even on all flash!)
• Consistent Latency / standard deviation of latency
Which one would you rather have?
Write Latency as
measured by
esxtop for similar
workloads
BRKCOM-2004 69
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
All Flash POC Best Practices2. Variance in performance seen across VMs
3. Compare Cost:• Especially important when comparing different hardware configurations
• Metrics: $/GB Effective And $/IOPS
BRKCOM-2004 70
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Performance Summary
HX All Flash performance significantly better than both HX Hybrid and other HCI Competitor’s all flash
All Flash performance is more consistent in comparison to HX Hybrid and competition
HX All Flash, unlike competition, shows highly consistent performance across VMs
All realistic workloads should see significantly better performance with all flash than hybrid
BRKCOM-2004 71
Q & A
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Complete Your Online Session Evaluation
Don’t forget: Cisco Live sessions will be available for viewing on-demand after the event at CiscoLive.com/Online
• Please complete your Online Session Evaluations after each session
• Complete 4 Session Evaluations & the Overall Conference Evaluation (available from Thursday) to receive your Cisco Live T-shirt
• All surveys can be completed via the Cisco Live Mobile App or the Communication Stations
73BRKCOM-2004
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Continue Your Education
• Demos in the Cisco campus
• Walk-in Self-Paced Labs
• Lunch & Learn
• Meet the Engineer 1:1 meetings
• Review related session content:
• BRKINI-4004 What’s New in Cisco HyperFlex 2.0
• BRKVIR-2700 HyperFlex Configuration Best Practices and Troubleshooting
74BRKCOM-2004
© 2017 Cisco and/or its affiliates. All rights reserved. Cisco Public
Cisco Spark Ask Questions, Get Answers, Continue the Experience
Use Cisco Spark to communicate with the Speaker and fellow participants after the session
Download the Cisco Spark app from iTunes or Google Play
1. Go to the Cisco Live Berlin 2017 Mobile app
2. Find this session
3. Click the Spark button under Speakers in the session description
4. Enter the room, room name = BRKCOM-2004
5. Join the conversation!
The Spark Room will be open for 2 weeks after Cisco Live
75BRKCOM-2004
Cisco HyperFlex 2.0 All Flash Systems Offer High Performance, Low Latency, and Consistent Results
Thank You