Application acceleration from the data storage perspective

Application Acceleration
From the Data Storage
Perspective
New Advances in Caching and
Solid State Storage

Lecturer:
Jacob Farmer, CTO
Cambridge Computer

© Copyright 2009-2010, Cambridge Computer Services, Inc. – All Rights Reserved
www.CambridgeComputer.com – 781-250-3000

The Pain: Random Disk I/O
Performance is not Easily Scalable
Random v. Sequential I/O
Random Disk I/O performance is not keeping pace with other
advancements such as processor and network I/O
• 15K drives are not that much faster than 10K!
• Traditional RAID arrays make it difficult to aggregate disk spindles
without host-based LVM software
Applications running in virtual servers are particularly
constrained by random I/O
• The virtualization environment randomizes traditionally sequential
workloads
• Booting and page file activity tend to share the same disk pools as
applications and thus compete with apps for disk I/O

Application Acceleration: New Advances in Caching and Solid State Storage
© Copyright 2010, Cambridge Computer Services, Inc. All rights reserved.
www.CambridgeComputer.com 2

The Relative Performance of Spinning
Disk to CPU, Cache and RAM
10,000,000 lb

100 lb

10 lb
1 lb

CPU L2 Cache RAM Spinning Disk


What’s Changing?

The price of DRAM has declined to the point where a
meaningful capacity is affordable
• Traditional DRAM SSDs were very pricey, niche technologies
– 5 years ago – 20GB of SSD sold for about $500K
The emergence of enterprise-class flash technology
• Recent innovations allow flash-based SSD to achieve meaningful
performance with enterprise-class reliability.
• Flash offers new tiers in between DRAM SSD and 15K disk
We are nearing the tipping point where SSD goes viral.
• Higher demand accelerates the decline in price
• Commoditization inspires new software technologies that leverage
commodity hardware


SSD Media Choices

• Multiple bits occupy a flash memory cell, reducing costs
MLC Flash • Higher bit-error rate than SLC
• Lesser performance SLC
• Lesser endurance than SLC
Multi-Level Cell • Software compensates for shortcomings in the media

SLC Flash • Faster write speeds than MLC
• Lower power consumption than MLC
• Higher endurance than MLC
Single-Level Cell • Higher cost than MLC, but still cheaper than DRAMS

• Equal performance for read and write
DRAM • No concern for duty cycle
• Requires uninterruptable power source


Flash Memory Durability and
Reliability
Flash memory does wears out over time
• MLC has a lower duty cycle than SLC
“Wear leveling”
• Firmware in the flash device keeps track of I/O requests to
each logical block and distributes I/O across flash cells
• As the device ages the wear leveling mechanism might
reduce performance.
– Especially true of devices designed for desktop and laptop usage

Confirm duty cycles with product manufacturer
• Most enterprise flash devices will outlive traditional hard
drives under specified use-cases

SSD Packaging: SSDs Come in Many
Shapes and Sizes
Hard Drive Form-Factor with SATA, SAS, or FC interface
• Can be installed in server drive slots
• Can be installed in enterprise SAN enclosures
– Add a new high performance tier (“Tier Zero”) to your SAN array
PCIe Cards and PCIe External Appliances
• Looks like a hard drive to the local OS
• Lower latency than devices that use storage channels (FC, SATA, SAS)
• Suitable for applications that can use direct-attached storage
– PCIe switches are slowly being adopted by enterprise applications
SAN-Attached Appliances
• Great for traditional enterprise apps, especially when high availability
clustering is involved
– Failover Clusters, Server Virtualization Platforms, Oracle RAC
DIMM Modules Installed on the Motherboard on the RAM Bus
• Extremely low latency


How do you size your RAID
groups or disk arrays?

By Capacity?
By Performance?
By Guess Work?

www.CambrIdgeComputer.com 8

Hard Drives = Units of Performance

Each hard drive (spindle) is like a little worker capable of
doing only so much work in a given day.
• Work is measured in IOPS – input/outputs per second.
• Faster hard drives deliver more IOPS than slower hard drives.
• Note: there is a diminishing return with increasing rotational speed.

Drive Type Typical Real-World IOPS


SSD Performance Comparison
Approximate – 4 KB Random I/O
Drive Type Typical Real-World IOPS


Cost Per Gigabyte versus
Cost Per Disk I/O
Drive Type Cost per GB

Drive Type Cost per IOP

Notes: 320GB SSD PCIe Card 128GB SSD Hard Drive 450GB 15k Hard Disk
IOPS Assumptions 100,000 20,000 180
Costs $7,500 $10,000 $1,000


Taking Advantage of Solid
State Storage


Applications for SSD

Any application or data set where the demand for random I/O
performance is high relative to the demand for capacity
• Databases: logs, temps, indexes, whole tables
– MS Exchange: logs and data
• Operating systems: boot images, page files
– Next Generation Virtual Desktop Solutions (VDI)
– Multi-host Platforms: VMware, Citrix
• Out-of-band file systems for metadata catalog
– Any file system where metadata can be isolated from file data
Modern advances enable cache and SSD for general purpose
computing
• Cache and SSD can be inserted anywhere along the storage I/O path
– Dynamic tiered storage, life cycle management, caching


How Do You Take Advantage of These
New Storage Tiers?
These new SSD tiers offer performance and cost
characteristics that are wildly different from
traditional storage.
New tiers drive the demand for automated tiered
storage.

Automated Tiered Storage
The ability to scale performance and
capacity independently of each other


The Storage I/O Path – Plenty of
Places to Insert Acceleration Logic

Host-based storage
management software
A p pl LAN-Based
O pe ic atio
ra tin n
g
S y s te
Appliance
Out-of-band File System m
Fi le S
y s te m
Out-of-band Volume Level V olu
me M
a nag
er

Server Virtualization Hypervisor

Switch-based Virtualization Fabric

SAN-based Volume
Virtualization Appliance

Disk Array Controller


Inserting Cache and SSD into the
Storage I/O Path

LAN/WAN


Solid State Storage with MS
Exchange


SSD in an Application: MS Exchange

Sizing disk I/O for MS Exchange
• ¼ to ½ I/O per active user
• Multiply by 4 for Blackberry/iPhone (any ActiveSync)
What do put on SSD
• The entire data set if it will fit
• EDBs holding mail for select power users
• Log files
– Big win for write-intensive mail systems
Exchange 2007 and 2010 offers HA clustering without the
need for a shared storage SAN
• CCR – Continuous Cluster Replication
• DAG – Database Availability Groups
• Leverage the low-cost of direct-attached SSDs


MS Exchange : Clustering with Direct-
Attached SSDs

LAN

SSD SSD
Exchange CCR Cluster

Replication and HA clustering without the need for
shared SAN storage
• CCR – Cluster Continuous Replication (2007)
• DAG – Database Availability Groups (2010)

MS Exchange: Direct-Attached SSDs
for Log Files Only
Exchange logs are stored on
internal SSD SSD
Exchange databases are
stored on traditional
enterprise SAN SAN
• Snapshots
• Replication
• Provisioning


Solid State Storage for
Databases


SSD for Databases

For small databases, put the entire database on SSD
Typical best practice is to determine which elements
of the database will benefit from SSD
• Usual suspects: Logs, Indexes, Temps, select tables
• Performance monitoring tools are available for all major
databases
– Example: Oracle StatPack

Use volume manager, intelligent SAN, or database
tools to migrate hotspots to SSD
• Oracle ASM allows seamless migration from one storage
device to another.

Adding SSD Device for Database
Hotspots

Database
Cluster
SAN Fabric

RAM
RAM-Based
SSD Flash-Based
Appliance Array

Disk Array with
Big Cache or
Flash SSD Drives


Database QoS with Operating System
Resource Partitioning
Quality of service can be enforced by manipulating
prioritization of operating system resources
Specific queries or patterns of queries are given
priorities or assigned performance threshold
• Prioritization policies can be set for disk I/O as well as CPU
utilization
Renegade queries are throttled
• As opposed to running wild or being killed
Allows you to size hardware around normal
conditions rather than worst scenario


Hyper-V, vSphere and
XenServer

Better Performance for Virtual
Environments


SSD with Clustered DAS Logical
Volume Manager and VMWare
Server Virtualization Hosts with Internal SSD

SSD SSD

“Clustered DAS” Software

SAN
SAN Add a redundant tier of
Conventional
SAN Storage
SSD to your VM
environment by installing
flash SSD cards into VM
hosts.


File Systems and NAS
Acceleration


File System Acceleration

If metadata can be isolated from data, put the metadata on
SSD drives
• Especially useful for SAN-enabled file systems that are heavily
constrained by metadata
Some file systems (notably ZFS) can cache recently accessed
files in SSD drives.
• This is especially useful when creating very large ZFS stripe groups
Some file systems allow the journal to be stored separately
Maybe your whole file system will fit on SSD drives!
• Use SSDs with hard drive form factor in place of a hard drive.
A caching appliance in front of your NAS
• Any place where NFS/CIFS improvement is needed


Clustered Scale-Out NAS: Living the
Dream!!!

RAM RAM RAM RAM

Clu stered NAS w ith Distrib u ted Co h eren t Cach e

Redundant SAN Fabric

RAM-Based Tier 0 Tier 1 Tier 2
SSD MLC Flash 15K Disk SATA Disk
for Metadata


A Clustered File System on IB with
Internal SSD Drives

Application Servers and/or Compute Nodes

Infiniband Fabric

Servers with Flash SSD cards
and lots of RAM


Clustered In-Line NAS Cache with SSD
and Tiered Storage

Local Area Network (1Gb, 10Gb, Infiniband, etc.)

Clustered Caching
Appliance with SSD
and Tiered Storage
File Server (NAS)

Use all SATA in your NAS

Retain snapshots,
replication, backup, etc.


SAN Acceleration


Hot-spotting with SSD or DRAM Using
SAN Volume Virtualization

SAN
Internal SSD
VIRTUALIZATION DEVICE

DISK CONTROLLERS
NOTE: Virtualization
layer is not drawn as
External SSD redundant.


Transparent In-band SAN Cache

SAN


SAN Topology View of Inline Cache

SAN


SSD in Dynamically Tiered Array

SSD can be leveraged across the
board with dynamic block-level
tiering.
Hottest blocks or fresh new
writes can stored on SSD and
migrate to lower tiers when not
in active use.
Makes it easy to deploy SSD
without having yet one more tier
to worry about.
• Secondary goal is to squeeze the
most I/O out of your disk pool.


SSD with Deduplication

Deduplication gets its greatest value when
• There is a lot of duplicate data
• The media is relatively expensive
New technologies allow block-level dedupe for
primary storage
Killer applications:
• Boot appliance for booting servers over the SAN
• Virtual desktop storage
• Source code libraries
• Oracle, Exchange


Conclusions

Solid state storage is the big game changer in the storage
industry right now
• The potential for increased random disk I/O performance
changes traditional assumptions about application performance
• The wide range of SSDs create more tiers and inspire demand
for automating tiered storage
New advances allow SSD to be inserted into the existing
storage infrastructure
• The challenge is to figure out the best way to insert SSD for any
given application
The end-goal is to purchase capacity and performance
independently of each other
• Grow capacity with larger, less expensive drives and grow
performance with various types of SSDs


Application acceleration from the data storage perspective

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

En vedette

En vedette (6)

Similaire à Application acceleration from the data storage perspective

Similaire à Application acceleration from the data storage perspective (20)

Plus de Interop

Plus de Interop (20)

Dernier

Dernier (20)

Application acceleration from the data storage perspective