• No results found

House of Cards. IaaS without storage performance testing. Howard Marks, Deep Storage Len Rosenthal, Load DynamiX

N/A
N/A
Protected

Academic year: 2021

Share "House of Cards. IaaS without storage performance testing. Howard Marks, Deep Storage Len Rosenthal, Load DynamiX"

Copied!
21
0
0

Loading.... (view fulltext now)

Full text

(1)

House of Cards

IaaS without storage

performance testing

(2)

Your NOT So Humble Speaker

25+ years of consulting and

writing for trade press

Columnist/blogger at

NetworkComputing.com

Chief Scientist DeepStorage, LLC

Independent test lab and analysts

@DeepStorageNet on Twitter

(3)

Today’s Agenda

What do we mean by IaaS?

Why IaaS is for the corporate

datacenter?

The three types of IaaS storage

How IaaS makes life difficult for

storage

The I/O blender and other hypervisor effects

Users choose their storage level

Noisy neighbors have a lease

(4)

Infrastructure As A Service?

Self service meets virtualization

Users create VM from service

catalog through web interface

CPU, memory, storage capacity, storage class

Can be delivered several ways

Public (eg: AWS, Microsoft Azure)

Private (eg: VMware vCAC, Openstack, Cloudstack)

(5)

IaaS Drivers

Agility

Workload deployment shrinks from days to minutes

Workloads billed by the hour

Ideal for seasonal, temporary workloads

Encourages users to release unneeded resources

Can lead to higher utilization

(6)

Storage = IaaS Magic

Scaling the compute side of IaaS

is science

Just manage the vCPU to core ratio

Storage for IaaS is still an art form

Storage costs vary over 10X for same capacity

Therefore, a large fraction of total costs

Storage guy metrics like IOPS alien to users

Poor visibility into how applications affect storage

(7)

IaaS Storage Types

Non-persistent

Block storage that’s released when VM shutdown

Persistent block storage

Eg: Amazon EBS, Openstack Cinder

Dynamically provisioned, traditional storage

Object Storage

The cheap stuff eg: Amazon S3

(8)

Block Storage in Classes

Defined in orchestration layer

Media

SSD, hybrid, spinning disk, Etc.

Data protection

RAID level, replication, snapshot schedule

Hopefully storage QoS

Minimum IOPS, IOPS quota

Priority

(9)

The Cloud’s Dark Side

Cloud providers have no visibility

Into how storage will be used

Into the application

So APM tools limited value

Into future requirements

The user/customer buys 200GB

of Silver

That may be a WORN archive or real time trading database

They may make it 2TB overnight

(10)

Virtualization is Hard on Storage

Disk drives like sequential

workloads

When multiple VMs do sequential

I/O they’re multiplexed together

That throws I/O into a blender…

all I/O is now random I/O!

(11)

Enter the Noisy Neighbor

Pre virtualization storage

performance managed by

spindles

More IOPS, more 15K RPM drives

Each app had dedicated spindles

In IaaS multiple VMs, from

different users, share a volume

When marketing re-indexes

(12)

Can’t I Just Throw Flash At It?

Flash based SSDs do provide

lots of IOPS

But even all flash arrays have their limits

and most of us can’t afford them

Hybrid solutions are application

sensitive

How much flash? How to deploy?

Storage performance validation

with load testing can reveal:

How much flash you need

How your system will react to new workloads, noisy neighbors

(13)

Storage Performance Validation

A key to successful IaaS

July 24, 2014

Len Rosenthal

(14)

Who is Load DynamiX?

Leader in Storage Performance Validation

Putting your storage technologies to the test

Provide insight into storage infrastructure behavior to assure performance & optimize cost for public and private clouds

Market entry March 2009; Privately-held Santa Clara, CA

Software: Load DynamiX Enterprise

Hardware: LDX Series load generation

(15)

Go Daddy Overview

 World’s largest web hosting provider: 5+ million accounts

 Over 50 million domain names under management

 3300 employees; 600+ IT developers, 5 storage admins

IT Challenge

 Determine most cost-effective storage systems to support

31PB growing 30% annually with 99.999% uptime across 5 data centers

Solution

 Multiple Load DynamiX 10G Series appliances

Go Daddy Benefits

 Projected storage cost savings of 65% over 2 years

 Full confidence that their storage systems can support

projected workloads “Load DynamiX is like my smart phone – I can’t imagine my life without it!” Julia Palmer Manager, Performance Engineering Go Daddy

Storage Performance Validation

(16)

Attaining IaaS Enlightenment

Validate with

If it ain’t broke don’t fix it!

Test in production and pray Validate with freeware tools (Iometer, etc.) Evolution

(17)

Storage Performance Validation Process

In 3 Simple Steps

Create

Model

Create storage workload emulations from production data or use our sample workloads.

Analyze

Results

Find performance limits. Compare

latency, throughput & IOPS of different

products and configs.

Run

Tests

Connect storage systems to Load DynamiX load generation appliance and run the tests.

1 2 3

Deploy In Production with

Iterate on different products & configurations

(18)

Go Daddy Production Workload

(19)

C

ost

per G

B

OS & Disk Change 56% drop Vendor B 23% drop Add Compression 35% drop Dedupe/Compression 62% drop Vendor C 56% drop Whitebox 67% drop

Every transition validated with Load DynamiX

Storage Cost Optimization

Driving cost optimization through technology validation

(20)

Benefits to Go Daddy by Using Load DynamiX

Faster Validation

Weeks to test in lab, not 12 months in

production

Storage Engineer

Superior Storage ROI

Pick most cost effective solution for workload

(21)

Q&A Session

Howard Marks DeepStorage.net

References

Related documents

Western Kentucky has already jumped out to an impressive 4-2 record with impressive wins at home over (ieorgia and Southern Illinois and an enormous victory on a neutral court

In neighbor- hoods with a scarce presence of immigrants the response rate is 30 percentage points lower for immigrants than for natives, while this differential decays towards zero

Table 2 gives hazard rates for total separation classified by initial job tenure and for years prior to and after the 1995 reform.. Most remarkably, these hazard rates exhibit a

We have developed a novel gas sensor based on an array of individual, suspended carbon nanowires decorated with PdNPs of various sizes for room temperature hydrogen gas sensing

Specifically we estimate the parameters of the following seven models: Standardized t-dis- tribution (ST), Twin t-distribution (TT) of Baker and Jackson ( 2014 ), Generalized

The STD400 is an electronic averaging transmitter that converts the average measured temperature into an electric current signal 4-20 mA.. The transmitter is used for

In this case, if we corrupt the clean recordings of a given speaker by different types and levels of distortion, we can artificially produce channel variability in recordings of

Provide support services for County departments, agencies, and employees such as miscellaneous necessary desktop/server applications; redundant backup services, virus protection